An Efficient Artificial Intelligence Energy Management System for Urban Building Integrating Photovoltaic and Storage

Abstract

The emerging leading role of green energy in our society pushes the investigation of new economic and technological solutions. Green energies and smart communities increase efficiency with the use of digital solutions for the benefits of inhabitants and companies. The paper focuses on the development of a methodology for the energy management, combining photovoltaics and storage systems, considering as the main case study a multi-story building characterized by a high density of households, used to generate data which allow feasibility foresights. The physical model of the algorithm is composed by two main elements: the photovoltaics modules and the battery energy storage system. In addition, to gain information about the real-time consumption a machine learning module is included in our approach to generate predictions about the near future demand. The benefits provided by the method are evaluated with an economic analysis, which computes the return of the investment using the real consumptions of a Boarding School, located in Turin (Italy). The case study analyzed in this article showed an increase in purchased energy at the minimum price from 25% to 91% and a 55% reduction in the electricity bill compared to most solutions on the market, with no additional costs and a stabilizing effect on the grid. Finally, the economic analysis shows that the proposed method is a profitable investment, with a breakeven point of thirteen years, due to the very simple implementation and the zero additional cost requested.

Full text

Received 19 December 2022, accepted 14 February 2023, date of publication 22 February 2023, date of current version 27 February 2023.
Digital Object Identifier 10.1109/ACCESS.2023.3247636
An Efficient Artificial Intelligence Energy
Management System for Urban Building
Integrating Photovoltaic and Storage
ENRICO GIGLIO 1,2,3, GABRIELE LUZZANI 1, VITO TERRANOVA4,
GABRIELE TRIVIGNO5, (Graduate Student Member, IEEE), ALESSANDRO NICCOLAI 4, (Member, IEEE),
AND FRANCESCO GRIMACCIA 4, (Senior Member, IEEE)
1Dipartimento di Ingegneria Meccanica e Aerospaziale, Politecnico di Torino, 10129 Turin, Italy
2Marine Offshore Renewable Energy Laboratory, Politecnico di Torino, 10129 Turin, Italy
3Energy Center Laboratory, Politecnico di Torino, 10129 Turin, Italy
4Dipartimento di Energia, Politecnico di Milano, 20156 Milan, Italy
5Dipartimento di Automatica e Informatica, Politecnico di Torino, 10129 Turin, Italy
Corresponding author: Enrico Giglio (enrico.giglio@polito.it)
This work was supported by the Politecnico di Torino through the CRUI CARE Agreement.
ABSTRACT The emerging leading role of green energy in our society pushes the investigation of new
economic and technological solutions. Green energies and smart communities increase efficiency with the
use of digital solutions for the benefits of inhabitants and companies. The paper focuses on the development
of a methodology for the energy management, combining photovoltaics and storage systems, considering
as the main case study a multi-story building characterized by a high density of households, used to
generate data which allow feasibility foresights. The physical model of the algorithm is composed by
two main elements: the photovoltaics modules and the battery energy storage system. In addition, to gain
information about the real-time consumption a machine learning module is included in our approach to
generate predictions about the near future demand. The benefits provided by the method are evaluated with an
economic analysis, which computes the return of the investment using the real consumptions of a Boarding
School, located in Turin (Italy). The case study analyzed in this article showed an increase in purchased
energy at the minimum price from 25% to 91% and a 55% reduction in the electricity bill compared to most
solutions on the market, with no additional costs and a stabilizing effect on the grid. Finally, the economic
analysis shows that the proposed method is a profitable investment, with a breakeven point of thirteen years,
due to the very simple implementation and the zero additional cost requested.
INDEX TERMS Deep learning, energy management systems, energy storage, environmental economics,
renewable energy sources.
I. INTRODUCTION
The emerging leading role of green energy in our society
pushes the investigation of new economic and technological
solutions. However, at higher-level RESs penetration, the sys-
tems become more complex to be controlled and balanced [1].
The actual political situation and environmental awareness
are pushing toward the energy transition. In this context, for
example, the European Commission (EC) has published the
The associate editor coordinating the review of this manuscript and
approving it for publication was Lei Wang.
‘‘EU Solar Energy Strategy’’ [2], trough which it aims to
bring online over 320 GW of solar photovoltaic by 2025
(more than doubling compared to 2020) and almost 600 GW
by 2030. Traditional photovoltaic (PV) plants, on the other
hand, can only generate electricity during daylight hours.
Because of the lack of nighttime performance, expensive
batteries and grid connections to alternative energy sources,
most notably fossil fuels, are required [3].
In this scenario, the possible installation of Renewable
Energy Sources (RES) in urban context can have a key role in
a feasible energy transition. Due to this reason, there has been
VOLUME 11, 2023 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 18673

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
a surge in interest in Building-Integrated Photovoltaic (BIPV)
applications in urban structures during the last 10 years.
Reference [4] emphasizes that just a few studies look at the
integration of PVs with smart grids, particularly regarding
BIPV systems. These are integrated into the structure and
replace specific structural parts (roof, façade, etc.), whereas
building-added PV systems are added to a structure. BI solar
systems provide visually beautiful structures as well as the
replacement of some building components. More in general,
cities are undergoing a profound shift in terms of energy and
sustainability, with increased density and a scattering of urban
distribution: this phenomenon can represent an opportunity
for energy management and distribution [5]. Even if the just
past global Sars-CoV-2 pandemic has slowed this apparent
urbanization tendency, we can expect a comeback as soon
as the health crisis is over, without jeopardizing this human
aggregation paradigm [6]. The development of modern
residential units, whose main dimensional characteristic
is based on their exponential height, was prompted by
the increasing congestion of cities. The notion of energy
communities was born out of the urban concentration of
several homes in a single building, allowing resource and land
use management to follow a low-emissions strategy [7].
Due to their intrinsic stochastic nature, RES can introduce
instability in the electric network. Thus, a key element for
obtaining a feasible energy transition is the capability to
achieve a good forecasting of RES production and electric
load [8]. Many works in the literature address the problem of
either short-term [10] or long-term forecasting [9].
Around the world, energy communities are forming to
raise awareness and reduce waste, based on creating a smart
grid. This system allows users to trade energy while reducing
prices and waste. The notion of an energy community is
vital in urban government since it has direct implications for
various environmental issues [31].
As concluded in [4], one of the key challenges for the
evolution of the smart grids, integrated with PV systems,
is in the opportunity to deal with the intermittent power
generation through the smart data management. In this sense,
prediction on the demand and the production could permit
to attenuate the highlighted problems if this information
would be properly use. Indeed, this paper investigate the
opportunity coming from the store of electricity when the
electrical demand to the grid is lower (typically the night),
and to use it during the day, with the one produced by the
PV system. In other words, the system would forecast the
amount of energy used by the building’s residents as well
as the amount of energy produced by the PV system on the
same roof. The difference represents the energy that should
be purchased throughout the day: the program would acquire
it at night, when it is normally cheaper, and store it in the
building storage system for use the next day.
The aim of this work is to introduce a hybrid control
strategy based on physical models of the system components
and machine learning methods for the prediction of electrical
load and RES production. This control strategy has been
specifically designed for the management of loads in an urban
context; in particular, the system should optimize the cost of
electricity in condominiums where structural constraints limit
the installable capacity of photovoltaic and battery energy
storage (BESS) systems.
This control strategy will exploit both photovoltaic produc-
tion and the time-of-use energy tariff in order to maximize the
profits that can be obtained through the installation of energy
storage systems.
This study brings several contributions to the field. Mainly,
this research proposed a power flow algorithm that combines
the exploitation of RES and time-of-use tariff to improve the
economic advantage of using BESS. Indeed, the proposed
algorithm can lead to a reduction of the final energy price
perceived by the private users. Furthermore, the hybrid use of
physical models and of machine learning, with the Temporal
Convolution Neural Network, represents another significant
novelty of this work.
Aside from the reduction in energy price, there are two
other major advantages of the use of the proposed algorithm:
it maximizes the use of RES and it leads to a higher
installation of BESS that can be used to increase the electric
network stability.
The remaining of the paper is structured as follows.
Section II describes the State-of-the-Art of the research and of
the industrial applications of integrated BESS system. Then,
in Section III the proposed method is described, detailing
both the physical and the data-driven models implemented.
In Section IV the case study used to test the proposed method
is described and in Section Vthe obtained results are shown.
Finally, in Section VI some conclusions are drawn.
II. STATE OF THE ART
A. DEMAND FORECASTING
The problem of modeling and forecasting the electric load
can be categorized as a time series prediction problem.
It is characterized by a series of challenges that are due to
the specific nature of the measures of interest: fluctuations
due to random behavior of the users operating in the
system of interest, and seasonal and weekly trends. The
way in which these issues affect the overall load can vary
at different scales. In the literature there are mainly two
categories of approaches: traditional, statistical methods, and
Machine Learning based ones. They are hereby discussed,
and a summarization of them is present as well in Tab. 1
exemplyfiyng for each paper the used algorithms and metrics.
Statistical models like SARIMA model were used in
[11]. Other traditional approaches based on multi-predictor
regression were applied in [12] and [13] to study both hourly
and daily energy consumption profile. Researchers also tried
to boost the performance of simple ARIMA models by
Bootstrap aggregating them in [14]. The main drawback of
purely statistical method is their inability to well capture
the highly non-linear behaviors that arise in consumption
profiles. This effect is amplified at smaller scale; in fact with
18674 VOLUME 11, 2023

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
higher volumes considered, the random fluctuations tend to
even out.
Moving on towards Machine Learning based methods, the
first approaches that showed good results were based on
Support Vector Regressors [15], [16] or ensemble learning
techniques and Adaboost to forecast energy consumption
[17]. Despite the improvement they provided, these method
still failed to reliably complete the forecasting task, mainly
due to overfitting, caused by evolving correlation over time
among the data. The more modern approaches come from
Deep Learning, as they try to exploit the elevated number
of parameters to capture complex temporal dependencies.
Simple ANNs are adopted in [26], whereas Cascaded ANNs
are proposed by [29] for resource forecasting. Fuzzy systems
(ANFIS) are also employed in [21] and [29].
Other works were based on more complex architectures
such as Recurrent Neural Networks (RNN) like [18]. The
main issue with RNN is that they struggle in keeping
‘‘memory’’ i.e. detect pattern on long temporal scales, like
seasonal trends that are fundamental in consumption profiles.
An improvement in this sense comes from Long-Short Term
Memory networks (LSTMs), as they were applied in [19].
For many years the LSTMs have gained large popularity,
as witnessed by the large body of research expanding on
them. Indeed, they are also exploited in [25], where firstly
a CNN is used to extract the embedding for the LSTM
recurrent cells. Moreover, in [27] they embed the LSTM
cell in a Non-Linear Auto-Regressive framework, in which
the feedback loop of the regressor contains time-step delays.
A complete comparison of traditional ML methods like
Linear Regression and Support Vector Machines (SVMs)
versus Multi Layer Perceptrons (MLPs) and LSTMs can be
found in [26], applied to day-forward predictions in a context
of microgrid clusters. Despite their success, LSTM-based
architecture struggle when complex and multi-scale temporal
dependencies are present, as shown in [28]. Subsequently,
another kind of architecture that improves the capability of
detecting long-range temporal dependencies was proposed,
the Temporal Convolutional Network. This architecture was
applied to the energy demand task in [20], but only at a
national scale. An alternative approach is found in [24],
where they employ a model similar to TCNs, but they
tune the hyper-parameters via an evolutionary algorithm.
Moreover, [21] analyzed the potential and issues of increas-
ingly deep models specifically applied to the residential
case.
More recently some more elaborate models have been
proposed, like [22], that introduced an attention mechanism
in order to detect upcoming and unforeseen surges of demand.
Another modern architecture, that has recently been gaining
popularity across different fields is the Transformer architec-
ture, characterized by the mechanism of self-attention. The
work of [23] shows the potential of these kind of models.
Machine learning algorithms have been also adopted in order
to forecast changes in energy consumer pattern, due to event
as the past pandemic one, and to support the energy suppliers,
grid operators, and traders to better calculate the required
operational flexibility [30].
Overall, the analysis of the literature indicates that Deep
Learning-based methods are getting increasingly accurate
and robust, thanks to their capability of modeling complex
and non-linear functions, with dependencies evolving over-
time.
B. EXISTING SOLUTIONS FOR BUILDING ENERGY
MANAGEMENT
Nowadays, the generation of electricity from renewable
energy sources (RES) is becoming increasingly important
to achieving the ambitious targets of reducing greenhouse
emissions set in the 2015 Paris Agreement [32]. Thus,
renewable energy generation such as solar or wind power
generation is spreading worldwide, leading to the problems
of their intrinsic variable nature, whose outputs could vary
temporally on many scales, and their integration into the
energy system [33]. In this context, governments and energy
multiutilities are questioning themselves which could be
the most fruitful model to adopt in order to solve these
challenges. According to [34], it is evident how a viable
solution is represented by the integration of photovoltaic
panels and energy storage systems (ESS) into the buildings
in a perspective of community-scale micro grids within a
city. In fact, considering a bidirectional energy flow, where
people may both buy and sell energy, smart grids ensure
a high penetration of stochastic and intermittent renewable
energy. Due to their variable nature, these grid systems
are experiencing high voltage fluctuations leading to power
quality problems that need to be controlled and stabilised
in order to guarantee an ever-high quality service [35].
Therefore, in recent years there has been an increasing
amount of work seeking solutions to stabilise these processes.
In particular, [36] proposed and validated a new control
scheme applied to a solar and wind hybrid power generation
system (HPGS) linked to a power grid, introducing a
supercapacitor to smooth out the ripple on the distribution
side in the power grid. Instead, [37] adopted a deep
learning model to predict the stability of a simulated smart
grid. However, proper implementation of these technologies
in a building environment can be obtained only with an
appropriate energy management system, which controls the
flow of energy between energy production (through PVs),
energy consumption, and energy storage, optimizing costs
through a link with the energy grid network [38], [39].
Thus, analysing the state of the art of the energy systems
for PV and storage energy management, it can be noted that
most of the software already present on the market are private
tools, developed by multiutilities. As highlighted by [40],
it is rare to find literature that analyses a local energy market
in which prosumers and consumers can actively participate
in the power supply process by exploiting ESS. Only a few
examples of building energy management software can be
found in the literature.
VOLUME 11, 2023 18675

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
TABLE 1. State of the art papers comparison.
In particular, the most of research studies applied in this
field are related to peer-to-peer energy trading in the context
of future energy communities. According to [41] and [42],
energy communities have developed as new entities that not
only provide end-users with unique platforms for investing
in low-carbon assets but also as operational market entities
capable of exchanging energy excess (deficit) among their
peers. In fact [43] investigates the benefits of a three-layered
architecture comprising cloud, fog, and consumer layers in
a smart grid environment. It highlights how the application
of this architecture allows an optimization of the resources,
minimizing the system’s response time and processing
time. Whereas [44] and [45] analyse the advantages and
disadvantages of an optimization model to schedule peer-
to-peer transactions via the local electricity market, grid
transactions in the retail market, and battery management.
They consider the photovoltaic production of households in
local energy communities, showing that combining the use
of peer-to-peer transactions and energy storage systems can
potentially provide consistent energy savings in the future.
Similar results are obtained by [46] through its study on
the problem of a smart community, composed of a group of
grid-connected prosumers with controllable loads, renewable
generations, and energy storage systems. It demonstrates
that the energy exchange in the proposed scalable energy
trading system results in considerable increases in energy cost
reductions and renewable energy usage efficiency. However,
all this research relates to the future development of energy
communities, while there is a clear lack of solutions that can
be implemented in a city context today. The analysis provided
in [61] show some preliminary results on a day-basis time
frame, where the BESS is used for both RES integration and
exploitation of different tariffs. These results provided a first
economic analysis, but it is important to increase the time
resolution of the forecasting and optimization models to be
able to consider the power flow, and, possibly, new features
such as the possibility to provide grid services.
Nevertheless, one possible solution is provided by [47] that
developed an algorithm to manage a building energy storage
system (BESS) to reduce the electricity price and the peak
18676 VOLUME 11, 2023

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
TABLE 2. State of the art software comparison.
load acting on the charge and discharge phase of the BESS,
by predicting the monthly load data of the building. Their
results highlighted how the electricity price was minimized
by peak load and electricity usage reduction. However,
this system was based only on the building energy data
without considering the forecasting of possible renewable
energy production. This lack of forecasting capability is
a common property of lots of the commercial PV and
storage management software, developed to manage these
technologies in the facility environments. These software can
be identified in: Enel X’s DER Optimization Software [48],
SMA’s Power Plant Manager [49], Homer [50], Energy
Toolbase [51]. Hence, the forecasting capability is present
only in Tesla’s solutions like Opticaster or Autobidder [52],
which are however not directly feasible in a city building
context. In this environment, in fact, there are currently only
photovoltaic panel and storage system solutions such as those
of Sonnen, Tesla Energy, LG Chem and CellCube, which
are not equipped with an energy management, prediction
and optimization software, but only with the hardware and
a storage control unit.
In order to summarize and compare all these considerations
and assess whether or not there is a gap in the current energy
market, it was decided to make a comparison using the
following four drivers:
1) integration into buildings: the capability of a solutionto
be easily inserted into a smart building context; perform
prediction analysis to make appropriate choices in
energy management between storage systems, PV and
the grid;
2) predictive capability: perform prediction analysis to
make appropriate choices in energy management
between storage systems, PV and the grid;
3) independent energy management: the ability of a
solution to manage energy flow autonomously;
4) speak to users: the capacity of a solution to interact
with users and make them aware of the actions and
functioning of the solution itself.
Taking into account these drivers, the comparison analysis
shown in Tab. 2was obtained.
Hence it can be seen in Tab. 2, that there are no solutions
at the state of the art that contains all the strong points that
we identified as drivers of a successful solution, capable
of covering the aforementioned challenges of the energy
market. This analysis, indeed, was the starting point for
the development of our whole project, which has in these
characteristics its strength and its novelty: we aim to develop
a software that offers a solution for city buildings to predict
and optimize the energy flow among PV production, building
consumption and energy bought from the grid.
III. METHODS
A. PROPOSED METHOD
The proposed method plugs itself in a scenario that essentially
comprehends a building equipped with a PV and a storage
system. The overall system is depicted in Fig.1, which is
split in 2: above, the training scheme for the AI module
that predicts energy demand (discussed below). In the
bottom part of the figure, all the inputs to our system are
exemplified, as well as the relationship between modules
in our overall architecture, and how they interact with the
building. Individual modules are described in the remaining
parts of this Section.
The system has 2 main sources of inputs: one coming from
the building, where the software has to monitor the state of
the storage, and receive live-data about the energy demand.
Another important input comes from the weather forecast,
that is going to be applied to the physical model of the PV
to extrapolate the forecasted power output. The information
about the real-time consumption is used by the AI module
to generate predictions about the near future demand. The
solar irradiance forecast gathered from the web is instead
processed by the physical model component to compute the
power output of the PV.
Internally, these pieces of information are combined, and
then fed to the optimizer module that provides the strategy
to follow. The strategy in detail consists in deciding at what
time and what amount of energy has to be purchased from
the grid, with the goal of fully exploiting the storage asset
and maximize final energy self-consumption while reducing
the overall cost for the user. The strategy can also be updated
after each new weather forecast or overridden by the building
manager to deal with sudden severe weather events.
B. PHOTOVOLTAIC MODULES PHYSICAL MODEL
In order to buy the proper amount of energy during the
night, it is of paramount importance that the algorithm can
accurately predict the power produced by PVs. During the last
years many physical models have been developed in order to
forecast the power production of photovoltaic modules based
on weather forecasts [53]. The most common and accurate
model of PV cell is an equivalent electric circuit based on
three, five or seven parameters. Results in [53] show that there
is no clear advantage in using the five or seven parameters
VOLUME 11, 2023 18677

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
FIGURE 1. Architectural block diagram of the proposed solution. In the above picture, the training procedure
of the AI-based consumption predictor block is schematized. Below, the flow of information through the
whole pipeline is clarified: the physical models provide the Optimization block with information about the
SOC and forecasted energy production, whereas the consumption predictor outputs demand forecast. These
pieces of information are processed by the Optimizer which in turn acts on the building to maximize the
energy quota bought at minimum price.
FIGURE 2. Equivalent circuit representing the three-parameter model
adopted.
instead of three one. Hence, we decided to model the PV cell
as the three parameter circuit shown in Figure 2that includes
a current generator and a diode connected in parallel that is
described by the following equation:
I=IPV −I0eV
nsVt−1(1)
where IPV is the photocurrent, I0is the dark saturation
current, nsis the number of series connected cells in the
module and Vtis the module thermal voltage.
The obtained model allows to predict the power produced
having as input:
FIGURE 3. Graphic representation of the PV model.
•the PV datasheet,
•the weather forecast: the diffuse irradiance (GDIFF ), the
direct normal one (GDNI ) and the ambient temperature
(Tamb),
•the date and time (to evaluate the relative position
between the PVs and the sun).
A graphical representation of the model is shown in Figure 3.
The performance of a PV cell is strictly related to its
temperature (TC). A good approximation is given by the
Nominal Operating Cell Temperature (NOCT) equation:
TC=Tamb +NOCT −Tamb@NOCT
GNOCT
·GTOT (2)
By using this formula, the cell temperature is obtained
assuming that the difference between ambient temperature
and the PV one is proportional to the total irradiance (GTOT )
18678 VOLUME 11, 2023

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
FIGURE 4. Graphic representation of the BESS model.
starting from a reference condition: NOCT =50◦C,
Tamb@NOCT =20◦Cand GNOCT =800 W/m2.
Finally, the model was implemented using the Solcast™
environment tools [54] with irradiance and temperature data
as input for plant power forecasting.
C. BATTERY ENERGY STORAGE SYSTEM PHYSICAL MODEL
We modeled also the battery energy storage system (BESS)
using [55] as main reference.
In particular, we wrote the efficiency of the BESS for its
cycle of charge and discharge as a function of its state of
charge (SOC), its number of life cycles [56] and the power
used as input (Figure 4).
Note that data provided by the two papers want to be just
a support for the model-building, but they cannot be applied
at any storage. Indeed, the characteristic values of the actual
installed storage must be provided by datasheets or measured.
Considering a flux of power Passociated to the storage
(positive if it charges the BESS, negative otherwise), the
relative energy value can be calculated by integrating P over
time:
En=Zdis/char
P(t)·dt (3)
and, during this process, the SOC changes as





SOC(t=0) =SOCinit
SOC(t>0) =SOCold +En
Emax
(4)
where SOCinit represents the initial state of the BESS,
SOCold is the SOC before the new operation and Emax is the
maximum capacity.
The model proposed, before changing the state of charge,
verify that:
–In case of positive power, that the battery capacity can
admit the charging load. The storage admit energy up to
saturation.
–In case of negative power, that the SOC is not null.
–In any case (discharge and charge), that the power flux
does not overcome the maximum power, whose value is
provided by data sheet.
The efficiency of the BESS is taken into account as follow:
Pactual,discharge =Pdischarge
η(5)
where ηis the efficiency of the BESS considering both
the effect of the SOC and the number of cycles. In par-
ticular, to avoid excessive underestimation of the system
FIGURE 5. High level representation of a TCN.
performance, this correction is done only in the discharge
phase. Because the data of this model are extremely problem
dependent, the final algorithm will start using these data and it
will be trained continuously with real data measured, in order
to correct itself automatically.
D. AI MODULE: ENERGY DEMAND PREDICTION
This section presents the proposed framework with an AI
module trained on an energy consumption dataset. Within
such framework, at inference time (i.e. when the trained
model is being used by the system with live data) the output
of the model is used jointly with the physical models of PV
and storage and the irradiance forecast from the Solcast API
to optimize the self-consumption of the final user, minimizing
the overall energy purchase while maximizing, among the
necessary purchases, the energy quota bought at the lower
price range.
The considered problem of energy forecasting has many
factors of variability that have to be captured by a model
that has to explain the data. Besides the amount of energy
that can be considered as baseload, that is constituted by
always-on appliances and recurrent habits (e.g. refrigerators,
night lights), there are both seasonal,weekly and daily trends.
Moreover, demand peaks are also frequent due to random use
of energy hungry appliances. Therefore the model of choice
will require a high capacity, in order to be able to quickly
respond to the peaks and learn the different time-scaled
trends. This means that for example a simple or multiple
regression will not be sufficient; and a more powerful model
is required.
In this work we propose the use of Temporal Convolutional
Networks, as discussed in Sec. I, also adopted in [20]. The
basic block of a TCN is depicted in Fig. 5, showing the
causal dilated convolution and how the depth allows upper-
level layers to have a broad receptive field. The formulation
for a dilated causal convolution over sequential layer is as
follows:
xt
l=g(
K−1
X
k=0
ωk
l·xt−(kxd)
l−1+bl) (6)
In this formalism xt
lis the output of the neuron at position
tin the l−th layer; K represents the convolutional kernel
size; ωk
lstands for the weight at position k, whereas dis the
VOLUME 11, 2023 18679

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
FIGURE 6. Representation of the operative way of using the model.
dilation factor and bl. The function gis the activation function
of choice which in modern DL in convolutional layers is
traditionally the Rectified Linear Unit.
1) TRAINING PROCEDURE
This paragraph describes the training procedure of the
consumption predictor block in our system. It is tasked with
predicting the demand, based on the inputs of the recent
past consumption, day and month. We describe operatively
the devised framework for training and inference with
the developed model. Mainly this means discussing the
preprocessing of the dataset, choice of a training loss, and
inference scheme.
The adopted dataset is a private collection of data from
1047 houses monitored for 2 years with data points every
15 minutes. The data was normalized with a min-max
normalization scheme.
Another point to discuss is how many houses to group
together to predict their total demand. This is mainly a trade-
of between the cost of the storage, which is of course higher
as you have higher demand, and the availability of data;
since grouping together a high number of samples drastically
reduces the final number of samples actually available for
training. The final choice turned out to be 20.
As for the inference scheme (i.e. prediction horizon), this
was mainly an issue of performances; the further you try to
predict in the future, the less performant it will be; whereas
the further you look back in the past the better, but having
huge sequences yields also a heavier model and more difficult
training. After some experiments the final model uses data
about the previous 2 days, outputting predictions for 8 hours
in the future. The model is used in a sliding window fashion,
as depicted in Fig. 6.
Regarding the choice of the optimizer, in the field of
Deep Learning SGD-based algorithms (Stochastic Gradi-
ent Descent) have become the de-facto standard. In the
energy community they are fairly popular, adopted by
[20], [21], [22], and [28]; however there are works that
exploit other algorithms such as Levenberg–Marquardt in
[26] and [29]. We chose to rely on SGD given its efficient
implementations available in the PyTorch framework, and
among its existing variants (SGD, Adam, AdamW, LARS..)
we selected SGD by experimental validation.
TABLE 3. Monthly averages over the years 2015-2019 for each price
bands in Italy.
The network has been trained for 30 epochs with a
Stochastic Gradient Descent optimizer and a Learning Rate of
10−3. The loss used was the Mean Squared Error (MSE),
suited to penalize errors when a quantity is the objective of
prediction. Therefore, from a statistical standpoint the model
estimates a Gaussian distribution conditioned on the input
sequence, which is a very common setting for Regression
tasks in which a real quantity is the objective.
As an additional form of regularization, the weight decay
technique has been used. It consists in adding a term to the
loss that penalizes the L2-norm of the network’s weights.
While this may seem odd, it is a very simple yet powerful
technique whose effect is the one of a statistical regularizer;
meaning it limits the capacity of the model in order to obtain
an estimator with a lower variance. Put in practical terms,
it avoids overfitting by asking that the weights be all close to
zeros; this will encourage the network to increase the value
of a weight only if it encodes useful information that actually
minimizes the loss of interest (MSE in our case).
Therefore the final objective of the training optimization is
reported below:
L(y,ˆy)=KMSE
N
N
X
i=0
(y− ˆyi)2+KWD ∗
wM
X
wi=0
||Wwi||2(7)
E. OPTIMIZING SELF-CONSUMPTION
This paragraph describes the devised algorithm that combines
the outputs of the demand forecasting model and the physical
models in order to predict demand, consumption, and act on
the system in order to optimize the use of energy. Regarding
the energy price, we referred to the Italian system of energy
pricing, which shares many similarities with countries in
Europe and across the world. It divides the day into 3 price
bands, each with its price. The price, as it is common, is lower
in off-peak time bands, therefore during the night. In our
experiments, we used as price values the monthly average of
prices for each slot in the last 5 years. In Tab. 3details on the
prices used can be found.
Regarding the choice of the timing of when to load the
storage, it is mainly a trade-off between the consumption
prediction horizon of the AI model, and the time required
to physically charge the storage, since we want to be able to
18680 VOLUME 11, 2023

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
buy energy before the more costly price band kicks in. Due
to the fact that our demand prediction window is of 8 hours,
the choice has been for the algorithm to operate on the basis
of 8-hours slots. Therefore, considering the nominal power of
the storage in our test-case (detailed in Sec. V-B), in order to
have enough time to load the storage during the lowest price
band, the first decision for the day is taken starting at 5 am,
relying on the prediction for the subsequent 8 hours and so
until 1 pm.
Considering that we want to account for as much of the
consumption of the day as possible, we add to the prediction
of the model a simple heuristic to account for the rest of the
day after 1 pm. The heuristic is the following: we take the
history about the past month, compute the sum of energy
deltas for the time slot of interest (1 am - 11 pm), and take
the daily average over them. This simple yet powerful ideas
has 2 advantages: has a bound with the period of the year at
hand (accounting for seasonal trends), and is able to smooth
out the effect of day-level outliers. The decision output is
how much energy to be bought and loaded into the storage,
according to the following formulation:
GAPi,j=
j
X
i=0
Ei
cons −Ei
prod (8)
Ei,j
buy =GAPi,j+EST _GAP1am−11 pm
+ −Estorage +MIN _THRESH (9)
where the index irunning in 0..jrepresents the current time
slot. GAPi,jis the sum of the energy deltas that occur in
the mentioned slot, computed subtracting from the predicted
demand Econs the forecasted energy produced Eprod.
EST _GAP1am−11 pm is computed with the heuristic
described in the previous paragraph, and its purpose is to
account for the energy deltas that occur in the rest of the day
that is beyond the predictive model horizon.
The term Estorage is needed to take into account the current
state of the storage; so, if the stored energy already covers
the predicted delta for the morning plus the estimated one
for the afternoon, no energy will be bought. The last term,
MIN _THRESH, is present as a kind of safety measure to
avoid wearing out the storage; in fact it is recommended to
increase the expected life at the rated efficiency of the storage,
to not discharge it completely. It can be seen as a safety net
to avoid discharging completely the battery and account for
unforeseen demand peaks. In our experiments it has been set
to 10% of the storage capacity.
Below is reported the pseudo-code that encodes the
principles explained; it does not take into account the time
required to load the storage (as the original code does) nor
does it track events, and the main purpose of this code is to
exemplify in an algorithmic fashion the concepts enounced.
The complete script that makes use of our software package
to run the optimizer routine can be found in our Github
repository.
FIGURE 7. Pseudo-code of our optimization algorithm.
TABLE 4. Case study specifications.
Note that the decision on how much to load the storage
is only considered during the F3 price band (the cheapest,
nightly), and never, say, during F2. So during the night
an estimate of the consumption of the following day is
computed, and a purchase is made accordingly. The reason
for this is the following: the F2 slot is valid from 7 am to
8 am and from 7 pm - 11pm. Given this hourly division, it is
ineffective to consider doing any computation to buy in this
slot. Indeed, if after 7 pm the storage happens to be empty
because consumption exceeded foreseen demand, energy will
be bought at this price anyway. Regarding the mere 1 hour
between 7-8 am, at that time the prediction has just been
computed and there are no sufficient further information.
IV. CASE STUDY
A key aspect of the project is represented by the understand-
ing of the economic advantage that our solution can provide
in the building energy management market. Therefore, it was
decided to consider a specific case study, in which evaluating
if an energy optimization management software could offer
an effective benefit in terms of cost savings on the bill.
A. BOARDING SCHOOL ENERGY CONSUMPTION DATA
The Italian Boarding School ‘‘Collegio Universitario Renato
Einaudi Torino 1935,’’ which main characteristics are
reported in Tab. 4[57], was chosen as the reference building
to perform our economic analysis.
The starting point of this analysis was to obtain the energy
consumption data for our case study, reported in Tab. 5, from
[61]. In particular, these energy consumption data are divided
into three price ranges: F1, F2, and F3. In Italy, indeed, the
energy price fluctuates, depending on the market, leading to
VOLUME 11, 2023 18681

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
TABLE 5. Case study monthly consumption data - 2019.
daily variations, depending on the request. According to the
Gestore dei Servizi s.p.a. report [58], the energy price ranges
in 2019 were divided as reported in the following:
•F1 range constitutes the most expensive one, covering
most of the daytime from 8.00 to 19.00 from Monday to
Friday;
•F2 is middle range, from 7.00 to 8.00 and from 19.00 to
22.00 from Monday to Friday and also from 7.00 to
22.00 on Saturday;
•F3 determines the hours in which the price decreases,
from 23.00 to 7.00 from Monday to Saturday and from
0.00 to 24.00 on Sunday.
The Boarding School energy consumption data provides a
monthly trend, while hourly subdivided values are required
in order to assess the profitability of the solution. Indeed,
comparing the hourly energy consumption data of the
Collegio with the values of solar panel production, it is
possible to evaluate how the delta of energy between
consumption and production changes during the year. This
delta of energy, indeed, is equivalent to the energy that
needs to be bought at the minimum price day by day during
the year. This kind of analysis will also make possible to
achieve a first storage sizing, which will be necessary for
the investment assessment that will be explained in the next
section. Therefore, it was decided to scale these monthly
values on an hourly energy consumption profile found in the
literature.
B. DATA ANALYSIS
Due to privacy matters determine a lack of information
in Italy regarding energy consumption, it was decided
to use an hourly consumption profile, obtained from the
London Datastore [59], coming from research undergone by
the Acorn Energy Group using SmartMeter technologies.
Acorn Energy [60], that is a conglomerate investing in
electricity generation and security, observed a sample of
5,567 London households between November 2011 and
February 2014 which have been monitored in their energy
consumptions, taking readings every half hour. From this
dataset it was possible to access data for the period from 16th
October 2012 at 00:30 to 16th October 2013 at 00:00. Since
FIGURE 8. ‘‘Collegio Einaudi’’ annual energy consumption profile.
energy consumption data are available for each half-hour, this
allows an amount of 17520 values.
C. ‘‘Collegio ENAUDI’’ HOURLY ENERGY CONSUMPTION
EVALUATION
Once the literature energy consumption profile was validated,
the Collegio Einaudi’s data were manipulated and scaled on
the basis of the London values. Thus, an hourly distribution
of energy consumption for the Boarding School was obtained
from an overlay of both data:
1) the London households profile data was reordered to
have a trend starting on 1 January at 00:30 and ending
on 31 December at 00:00;
2) for each i-th consumption value (with i=1:
17520), its percentage weight on the total monthly
consumption (% Monthly cons(i)) was calculated
through the following expression:
%Monthly cons(i)=Hourly cons(i)
Nday
P
j=1
Hourly cons(j)
(10)
where Hourly cons(i) is the i-th value of energy
consumption given by the reference dataset and Nday
is the number of days per month;
3) these monthly percentage values were multiplied by
the corresponding monthly consumption data of the
Boarding School, already shown in Tab. 5. During
this passage, a special attention was given to multiply
the correct punctual percentage value to the respective
monthly Boarding School one, taking into account the
hourly energy price range division (F1, F2 and F3)
previously described;
4) the next step consisted in omitting some out-of-scale
values, replacing them with the average between the
previous and the following one;
5) then, the week in which the UK household went on
holiday (from 24 June 2019 at 00:30 to 1 July 2019 at
00:00), therefore not matching the Boarding School
trend, was replaced with the consumption profile of the
previous week;
The results are shown in Fig. 8. It is possible to observe
a considerable discontinuity of data in August caused,
obviously, by the very low consumption of the college linked
to the students’ summer vacations.
18682 VOLUME 11, 2023

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
The next step was to assess a first estimation of the
storage system sizing, exploiting the previously introduced
PV model. In fact, following the procedure described in [61],
it was possible to estimate the PV energy production profile
of the Boarding School in 2019. Therefore, knowing the delta
of energy between consumption and PV production for each
day of the 2019, it was possible to assess a first storage sizing
of ≃1500 kWh.
V. RESULTS
A. INVESTMENT ANALYSIS
To evaluate whether the investment related to our solution
could be viable or not, we decided to perform an economic
analysis based on one of the well-known and most effective
indicators in the PV and BESS energy system market: the
Net Present Value (NPV). This metric indicates the difference
between the present value of cash inflows and the current
value of cash outflows under a period of time [62], [63],
[64]. Moreover, further consideration of our analysis is based
on [65], showing that tax deductions reduced the payback
period for the investment and increased the annual savings
from the energy storage systems.
As reported in the article [61], the break even time for
the case study considered in the section IV is 13 years.
This number was obtained by discounting the cash flows
using the NPV considering the initial investment for the
purchase of photovoltaic modules and BESS, the utility bill
savings possible from the proposed solution, and the annual
maintenance costs of the PVs. The main reason why the
investment is paid back after so many years is the high
initial cost associated with PVs and BESS. However, both
the prices of PVs and BESS will decrease in the next years.
Indeed, PV have exhibited the most rapid cost decrease
among energy technologies, economy of scale being one of
the main factor to comport this decrease and their economic
potential lies in the further reduction of price expected in
the next decades. US investment bank Lazard’s edition of
its annual Levelized Cost of Energy Report 2020 showed a
7% decrease on a year basis, expecting to reach values lower
than for any other power source [66]. Considering the BESS,
according to Bloomberg New Energy Finance New Energy
Outlook (BNEF 2018) [67], over 1,200 GW of additional
Li-ion battery capacity (a choice motivated by their high
specific energy) is expected to be deployed by the year 2050.
Investments over the next few years are expected to be located
mainly in Asia and Europe reaching a combined total cost of
$544 billion (BNEF 2018). Between 2010 and 2017, battery
prices have fallen by 80 percent, getting to an average of
$200/kWh, projections estimating that the price will reach
approximately [68].
The combined use of photovoltaics and energy storage
systems set the opportunity to generate a profitable invest-
ment characterized by increasingly growing profit margins.
Hence, considering both PVs and BESSs prices reduction,
we decided to analyze different scenarios in which such
FIGURE 9. Cash flow analysis - investment comparison.
an investment made in different years would be evaluated.
Furthermore, as can be seen from the following analysis,
the implementation of our solution was not considered in
the initial investment. This is because it is a zero-cost
solution which can be perfectly and easily integrated into
any hardware system that a company has already developed.
The main components figuring in the initial investment costs
are [61]:
•The PV plant: 129,600 e
•Li-ion batteries: 570,000 e
We projected the PV price by using the 7% decrease on a
year basis [66] and the BESS price by applying the current
mark-up to the projected price of the raw materials expected
by Bloomberg (BNEF 2018). We considered also an average
market cost of 35 e/kW of PV maintenance annual cost.
Hence, we evaluated different scenarios which differ for
the considered moment of installation: the cost projections
have been proposed varying the starting year from 2021 to
2030 and each investment scenario is evaluated for the
following 30 years period (chosing, according to literature,
a discount rate of 5% [69] for the NPV). We evaluated
different scenarios which differ for the considered moment of
installation: the cost projections have been proposed varying
the starting year from 2021 to 2030 and each investment
scenario is evaluated for the following 30 years period.
Results are shown in Figure 9.
Final considerations show how NPV of the investment only
becomes positive during the last year, the necessity to repay
the expenses for the storage every fifteen years playing a
relevant role on the profit margin. Meanwhile the cash flow
of the project as seen in Fig.9underlines how the profitability
margin is steadily increasing over the years. Considering the
batteries and PV trends in the future is acceptable to say
the project investment is going to increase its profitability
exponentially in a horizon of thirty years.
B. OPTIMIZED SOLUTION RESULTS
In this section are discussed the results obtained by applying
the full proposed framework on the case study described in
Section IV and comparing the results of a PV +storage
system that implements the proposed solution, and a standard
one that does not.
VOLUME 11, 2023 18683

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
FIGURE 10. Results applying the ULISSE sofware: (a) behavior of the
storage in a ULISSE system in response to consumption/production
curves (b) behavior of the storage in a standard system in response to
consumption/production curves.
The analysis begins by looking at the curves reported in
Figure 10, regarding the behavior of the system across 5 days
in the month of May having some days below consumption,
same days slightly above.
Specifically, Figure 10b shows what would be the state
of the storage in the depicted days. Even though including
a storage system is definitely a choice that will be more
and more convenient in the future and it avoids to waste
the excess of production, this curves show clearly how
it can happen, quite frequently, that production cannot
keep up with the demand, causing the storage to be
unused.
Considering Figure 10a, which shows the behavior of the
system during the same days, on a system that includes
the proposed solution. These curves are quite explicative
of the goal of the proposed system and also of the quality
of results that it is able to provide. In particular, it can be
noted how, for the days of May the 2nd and 3rd respectively,
in which production rates were quite low, how our optimizer
successfully applied the predictive model and the heuristics
to decide to buy an extra quota of energy in the night shift,
were all the charging spikes in the SOC are present. It can be
seen how the system was able to reach the next charging slot
without running out of energy and thus without having to pay
for energy in the more costly bands.
FIGURE 11. Results applying the ULISSE software in a situation of
over-production in the month of August.
FIGURE 12. Results applying the ULISSE software in a situation of
under-production in the month of January.
For the second part of analysis results we will consider days
in which production is consistently higher than demand, and
viceversa.
Figure 11 depicts the situation, in the month of August,
where the production can consistently, for more consecutive
days, overcome demands. To take into consideration this
setting is important to assert the ability of our software to
carefully administrate the balance between prediction of the
energy deltas for the morning, and estimate of the gap for the
rest of the day.
Figure 11 shows how in the first 2 days, where production
is comparable to demand, it is able to understand that it is
necessary to fully load the storage during the night in order to
survive until the next day; whereas in the last 3 days depicted
where production outweighs demand it is remarkable how the
system understands that it needs to leave some headroom in
the storage capacity to account for the surplus of production,
that makes the storage load again. This behavior is clear from
the curve of the SOC.
Instead, Figure 12 reports the behavior of the system
during 5 particularly low-irradiated days in the month of
January. In this situation the hard-limit on the savings that
our optimization can provide is directly proportional to the
18684 VOLUME 11, 2023

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
TABLE 6. Self-consumption support schemes in each country reported in [71].
FIGURE 13. Results applying the ULISSE software: Difference in a full year
of expenses using ULISSE vs a standard system.
size of the storage, as it can be clearly understood from the
chart.
Indeed, the model is correctly able to estimate that
production will not be able to keep up with demand and
therefore the storage gets fully charged every night; however
it is not enough to supply energy for all the remaining day.
Therefore if one wants to further reduce the cost of the bill
the only solution is to increase storage capacity, provided that
that is compatible with the desired time to break-even.
Moving to the analysis of saving on the bills, the curves in
Fig.13 are devoted to show practically, at the end of the day
what kind of advantages our software can provide. The value
of capacity that has been utilized to generate all the charts
reported in this last Section, as already stated, corresponds to
the use of 14 unit of the Sonnen standard module, accounting
for a total capacity of 14 ∗14 kWh =196 kWh. The saving
that can be obtained using this capacity, with respect to a
standard installment amounts to 1398 eon the total cost of
energy for the building on a yearly basis, so approximately
70eper household. Reading from the graph it also emerges
what was reasoned upon looking at Fig.12; that is how
having consistent under-production over several days creates
a bottleneck on the savings that depends on the capacity of
the storage. This last chart shows how the value of capacity
at which the obtainable saving plateaus is beyond 400 kWh,
where a saving of over e2200 is registered, so an increase in
savings of 60 % when doubling capacity.
The chart also shows how for a standard system, the ability
to exploit storage capacity is really limited to the situations
in which productions outweighs demand, and therefore once
the storage is capable enough to store the highest daily
delta in production, increasing it will not provide any more
advantages.
Contrarily, the solution proposed by this paper will provide
a consistent saving on the total bill. It is important to underline
how the implementation of the proposed solution comes with
no additional cost, and it is therefore very easily pluggable-
in to whatever remote control system the storage providing
company has already in place.
C. SELF-CONSUMPTION REMUNERATION MECHANISM
Furthermore, European governments have activated remu-
neration policies to incentive green-energy self-consumption,
which can be variable or fixed, working with different
mechanisms [70]. Note how the economic saving computed
in the previous section does not take into account these
policies, which vary nation by nation. For example, [71]
explains how the remuneration mechanism works in Spain,
France, Germany Italy, United Kingdom, and Finland.
It highlights how, in the different countries, the granted feed-
in-tariff (FIT) can strongly vary country by country, as also
the plant size for which the remuneration can be applied.
The different self-consumption support schemes described
by [71] have been reported in table 6.
This mechanisms of remuneration have not been taken
into account in the economical analysis previously proposed
because strongly depends on the nation considered, but also
on the policy in effect at the time when the project is installed.
This kind of variable aspect cannot be modelled as done for
the expected reduction cost, weakening the validity of the
results. For this reason, this further economic instruments
have not been considered, but they have been proposed for the
economic push that they bring to the economical feasibility of
the solution proposed.
The proposed methodology can also be integrated with
electricity demand management algorithms. In this sense,
EVs represent an opportunity on both the demand and supply
side, as they could positively influence the effectiveness of
the storage system, and brings to further economic benefits in
terms of remunerations as well. Reference [73] shows how
VOLUME 11, 2023 18685

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
PV systems with BESS can operate in vehicle-to-grid (V2G)
and grid-to-vehicle (G2V) modes, increasing the efficiency
of the whole system. The integration of such modes in the
proposed case study could mean higher economic savings,
thanks to the revenues coming from the sale of energy to grid
(and then to the EVs).
Lastly, the Ancillary Service Market (MSD) represents
another opportunity for accessing to remuneration mecha-
nisms for energy system with storage, that can give flexibility
in power fed into the grid and the load demanded from the grid
[74]. The TSO usually pays for such flexibility (on demand)
because it helps to balance the local grid frequency. In order
to consider the mechanism of the MSD into the optimization,
several trade-off phenomena have to be considered, such as:
•the additional costs necessary to meet the technical
requirements needed to enter market such as the MSD;
•the model of the possible revenues from this type of
electricity market;
•an estimation of when the TSO might require a
demand/production shift.
In summary, this paper proposes an optimisation strategy
that takes into account the economic benefits that could be
derived from the electricity bill based on the hourly rate.
Moreover, the proposed methodology could be extended to
take into account other phenomena that allow access to other
remuneration mechanisms, thus increasing the economic
profitability of the adopted optimisation strategy.
VI. CONCLUSION
In the context of energy transition in a densely populated
urban environment, this work aimed to introduce a hybrid
control strategy based on physical models of system compo-
nents and machine learning methods for predicting electrical
load and RES production. In order to optimize the profits
that can be obtained by installing energy storage, this control
strategy will exploit both the photovoltaic production and
the energy price at the time of use. The application of the
proposed methodology has a number of advantages. Firstly,
it maximizes the use of RES; secondly, it can reduce the final
energy price perceived by residential customers; finally, it can
encourage the installation of BESS, which can be used to
increase the stability of the electricity grid.
The economic analysis shows that our proposed method is
a profitable investment today, even taking into account the
cost of the storage system and the PV installation. The break-
even point will indeed be reached in thirteen years, with the
costs of the storage systems, which represent a large part of
this investment, playing a key role.
The presented work evaluated the economic and energetic
feasibility of an Computational Intelligence-based system
for an efficient management of a Battery Energy Storage.
Additional studies can investigate the possibility of exploiting
this system also for providing grid services for improve its
stability: this additional analysis can highlight new possible
business that can further enhance the economic feasibility of
the proposed system.
Given Bloomberg’s forecast for the cost of the storage
systems, we obtained that in the following years, especially
from 2026, a more convenient and fruitful investment could
be reached leading to break-even in less than six years. This
process could be expanded exponentially over the next few
years as storage costs continue to decrease. However, this
trend can be accelerated with solutions for the optimization of
energy management, and, in this sense that the point comes to
the impact of our approach. Once the prediction system was
implemented and trained, the economic results confirmed the
efficiency of the proposed method even under completely
realistic conditions. In summary, it is able to increase the
percentage of energy purchased at the minimum price from
25% to 91% and reduce the electricity bill by 55% compared
to the most current solution on the market. And all this
without additional costs and while stabilising the grid.
ACKNOWLEDGMENT
The authors acknowledge the Boarding School ‘‘Einaudi,’’
Turin, Italy, for having shared the data necessary to test the
proposed methodology.
REFERENCES
[1] M. S. Alam, F. S. Al-Ismail, A. Salem, and M. A. Abido, ‘‘High-level
penetration of renewable energy sources into grid utility: Challenges
and solutions,’’ IEEE Access, vol. 8, pp. 190277–190299, 2020, doi:
10.1109/ACCESS.2020.3031481.
[2] European Commission (EC) Communication COM/2022/221, EU Solar
Energy Strategy, Brussels, Belgium, May 2022.
[3] T. Deppe and J. N. Munday, ‘‘Nighttime photovoltaic cells: Electrical
power generation by optically coupling with deep space,’’ in ACS
Photonics, vol. 7, 1st ed. Washington, DC, USA: American Chemical
Society, 2020, pp. 1–9.
[4] C. Lamnatou, D. Chemisana, and C. Cristofari, ‘‘Smart grids and smart
technologies in relation to photovoltaics, storage systems, buildings and
the environment,’’ Renew. Energy, vol. 185, pp. 1376–1391, Feb. 2022.
[5] C. Connolly, S. H. Ali, and R. Keil, ‘‘Onthe relationships between COVID-
19 and extended urbanization,’’ Dialogues Hum. Geography, vol. 10, no. 2,
pp. 213–216, Jul. 2020, doi: 10.1177/2043820620934209.
[6] J. Parasonis and A. Keizikas, ‘‘Possibilities to reduce the energy demand
for multistory residential buildings,’’ in Proc. 10th Int. Conf. Modern
Building Mater., 2010, pp. 989–993.
[7] R. Roversi, A. Boeri, S. Pagliula, and G. Turci, ‘‘Energy community
in action—Energy citizenship contract as tool for climate neutrality,’’
Smart Cities, vol. 5, no. 1, pp. 294–317, Mar. 2022, doi: 10.3390/smartci-
ties5010018.
[8] A. Niccolai, A. Dolara, and E. Ogliari, ‘‘Hybrid PV power forecasting
methods: A comparison of different approaches,’’ Energies, vol. 14, no. 2,
p. 451, Jan. 2021.
[9] K. Zheng, B. Qian, S. Li, Y. Xiao, W. Zhuang, and Q. Ma, ‘‘Long-short
term echo state network for time series prediction,’’ IEEE Access, vol. 8,
pp. 91961–91974, 2020, doi: 10.1109/ACCESS.2020.2994773.
[10] Y. Wang, Y. Kong, X. Tang, X. Chen, Y. Xu, J. Chen, S. Sun, Y. Guo,
and Y. Chen, ‘‘Short-term industrial load forecasting based on ensemble
hidden Markov model,’’ IEEE Access, vol. 8, pp. 160858–160870, 2020,
doi: 10.1109/ACCESS.2020.3020799.
[11] V. G. Tran, V. Debusschere, and S. Bacha, ‘‘One week hourly electricity
load forecasting using neuro-fuzzy and seasonal ARIMA models,’’ IFAC
Proc. Volumes, vol. 45, no. 21, pp. 97–102, 2012.
[12] N. Fumo and M. A. R. Biswas, ‘‘Regression analysis for prediction of
residential energy consumption,’’ Renew. Sustain. Energy Rev., vol. 47,
pp. 332–343, Jul. 2015.
[13] K. P. Amber, M. W. Aslam, and S. K. Hussain, ‘‘Electricity consumption
forecasting models for administration buildings of the U.K. higher
education sector,’’ Energy Buildings, vol. 90, pp. 127–136, Mar. 2015.
18686 VOLUME 11, 2023

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
[14] E. M. de Oliveira and F. L. C. Oliveira,‘‘Forecasting mid-long term electric
energy consumption through bagging ARIMA and exponential smoothing
methods,’’ Energy, vol. 144, pp. 776–788, Feb. 2018.
[15] R. K. Jain, K. M. Smith, P. J. Culligan, and J. E. Taylor, ‘‘Forecasting
energy consumption of multi-family residential buildings using support
vector regression: Investigating the impact of temporal and spatial
monitoring granularity on performance accuracy,’’ Appl. Energy, vol. 123,
pp. 168–178, Jun. 2014.
[16] Y. Liu, H. Chen, L. Zhang, X. Wu, and X.-J. Wang, ‘‘Energy consumption
prediction and diagnosis of public buildings based on support vector
machine learning: A case study in China,’’ J. Cleaner Prod., vol. 272,
Nov. 2020, Art. no. 122542.
[17] T. Pinto, I. Praça, Z. Vale, and J. Silva, ‘‘Ensemble learning for electricity
consumption forecasting in office buildings,’’ Neurocomputing, vol. 423,
pp. 747–755, Jan. 2021.
[18] A. Rahman, V. Srikumar, and A. D. Smith, ‘‘Predicting electricity
consumption for commercial and residential buildings using deep recurrent
neural networks,’’ Appl. Energy, vol. 212, pp. 372–385, Feb. 2018.
[19] J. Q. Wang, Y. Du, and J. Wang, ‘‘LSTM based long-term energy
consumption prediction with periodicity,’’ Energy, vol. 197, Apr. 2020,
Art. no. 117197.
[20] P. Lara-Benítez, M. Carranza-García, J. M. Luna-Romera, and
J. C. Riquelme, ‘‘Temporal convolutional networks applied to energy-
related time series forecasting,’’ Appl. Sci., vol. 10, no. 7, p. 2322,
Mar. 2020, doi: 10.3390/app10072322.
[21] M. Fayaz and D. Kim, ‘‘A prediction methodology of energy consumption
based on deep extreme learning machine and comparative analysis in
residential buildings,’’ Electronics, vol. 7, no. 10, p. 222, Sep. 2018.
[22] Y. Gao, Y. Ruan, C. Fang, and S. Yin, ‘‘Deep learning and transfer learning
models of energy consumption forecasting for a building with poor
information data,’’ Energy Buildings, vol. 223, Sep. 2020, Art. no. 110156.
[23] L. S. Saoud, H. Al-Marzouqi, and R. Hussein, ‘‘Household energy
consumption prediction using the stationary wavelet transform and
transformers,’’ IEEE Access, vol. 10, pp. 5171–5183, 2022, doi:
10.1109/ACCESS.2022.3140818.
[24] S. M. J. Jalali, S. Ahmadian, A. Khosravi, M. Shafie-Khah, S. Nahavandi,
and J. P. S. Catalao, ‘‘A novel evolutionary-based deep convolutional
neural network model for intelligent load forecasting,’’ IEEE Trans.
Ind. Informat., vol. 17, no. 12, pp. 8243–8253, Dec. 2021, doi:
10.1109/TII.2021.3065718.
[25] S. H. Rafi, S. R. Deeba and E. Hossain, ‘‘A short-term load forecasting
method using integrated CNN and LSTM network,’’ IEEE Access, vol. 9,
pp. 32436–32448, 2021, doi: 10.1109/ACCESS.2021.3060654.
[26] S. N. V. B. Rao, V. P. K. Yellapragada, K. Padma, D. J. Pradeep,
C. P. Reddy, M. Amir, and S. S. Refaat, ‘‘Day-ahead load demand
forecasting in urban community cluster microgrids using machine
learning methods,’’ Energies, vol. 15, no. 17, p. 6124, Aug. 2022, doi:
10.3390/en15176124.
[27] M. Massaoudi, I. Chihi, L. Sidhom, M. Trabelsi, S. S. Refaat, H. Abu-Rub,
and F. S. Oueslati, ‘‘An effective hybrid NARX-LSTM model for point and
interval PV power forecasting,’’ IEEE Access, vol. 9, pp. 36571–36588,
2021, doi: 10.1109/ACCESS.2021.3062776.
[28] S. Bai, J. Zico Kolter, and V. Koltun, ‘‘An empirical evaluation of generic
convolutional and recurrent networks for sequence modeling,’’ 2018,
arXiv:1803.01271.
[29] M. Amir and A. Zaheeruddin, ‘‘Intelligent based hybrid renewable energy
resources forecasting and real time power demand management system
for resilient energy systems,’’ Science Progress, vol. 105, no. 4, 2022,
Art. no. 00368504221132144, doi: 10.1177/00368504221132144.
[30] S. V. Oprea, A. Bara, V. Diaconita, C. Ceaparu, and A. A. Ducman,
‘‘Big data processing for commercial buildings and assessing flexibility
in the context of citizen energy communities,’’ IEEE Access, vol. 9,
pp. 168715–168730, 2021, doi: 10.1109/ACCESS.2021.3137352.
[31] IEA. (2020). Renewables 2020 Data Explorer, IEA, Paris, France, 2020.
[Online]. Available: www.iea.org/articles/renewables-2020-data-explorer,
last seen: 06/07/2021.
[32] United Nations/Framework Convention on Climate Change, Adoption of
the Paris Agreement, 21st Conference of the Parties, United Nations, Paris,
France, 2015
[33] H.-K. Ringkjøb, P. M. Haugan, and I. M. Solbrekke, ‘‘A review of
modelling tools for energy and electricity systems with large shares of
variable renewables,’’ Renew. Sustain. Energy Rev., vol. 96, pp. 440–459,
Nov. 2018, doi: 10.1016/j.rser.2018.08.002.
[34] C. Kennedy, I. D. Stewart, A. Facchini, and R. Mele, ‘‘The role of utilities
in developing low carbon, electric megacities,’’ Energy Policy, vol. 106,
pp. 122–128, Jul. 2017, doi: 10.1016/j.enpol.2017.02.047.
[35] Z. Shi, W. Yao, Z. Li, L. Zeng, Y. Zhao, R. Zhang, Y. Tang, and
J. Wen, ‘‘Artificial intelligence techniques for stability analysis and
control in smart grids: Methodologies, applications, challenges and future
directions,’’ Appl. Energy, vol. 278, Nov. 2020, Art. no. 115733, doi:
10.1016/j.apenergy.2020.115733.
[36] M. Amir, A. K. Prajapati, and S. S. Refaat, ‘‘Dynamic performance evalu-
ation of grid-connected hybrid renewable energy-based power generation
for stability and power quality enhancement in smart grid,’’ Frontiers
Energy Res., vol. 10, p. 222, Mar. 2022, doi: 10.3389/fenrg.2022.861282.
[37] A. Salam and A. El Hibaoui, ‘‘Applying deep learning model to predict
smart grid stability,’’ in Proc. 9th Int. Renew. Sustain. Energy Conf.
(IRSEC), Nov. 2021, pp. 1–9, doi: 10.1109/IRSEC53969.2021.9741136.
[38] G. Chaudhary, J. J. Lamb, O. S. Burheim, and B. Austbø, ‘‘Review
of energy storage and energy management system control strategies
in microgrids,’’ Energies, vol. 14, no. 16, p. 4929, Aug. 2021, doi:
10.3390/en14164929.
[39] K. Parvin, M. S. H. Lipu, M. A. Hannan, M. A. Abdullah, K. P. Jern,
R. A. Begum, M. Mansur, K. M. Muttaqi, T. M. I. Mahlia, and Z. Y. Dong,
‘‘Intelligent controllers and optimization algorithms for building energy
management towards achieving sustainable development: Challenges
and prospects,’’ IEEE Access, vol. 9, pp. 41577–41602, 2021, doi:
10.1109/ACCESS.2021.3065087.
[40] E. Mengelkamp, J. Garttner, and C. Weinhardt, ‘‘The role of energy storage
in local energy markets,’’ in Proc. 14th Int. Conf. Eur. Energy Market
(EEM), Jun. 2017, pp. 1–6, doi: 10.1109/EEM.2017.7981906.
[41] M. Grzanić, J. M. Morales, S. Pineda, and T. Capuder, ‘‘Elec-
tricity cost-sharing in energy communities under dynamic pricing
and uncertainty,’’ IEEE Access, vol. 9, pp. 30225–30241, 2021, doi:
10.1109/ACCESS.2021.3059476.
[42] E. Oh and S.-Y. Son, ‘‘Peer-to-peer energy transaction mechanisms
considering fairness in smart energy communities,’’ IEEE Access, vol. 8,
pp. 216055–216068, 2020, doi: 10.1109/ACCESS.2020.3041838.
[43] Z. A. Khan, A. A. Butt, T. A. Alghamdi, A. Fatima, M. Akbar,
M. Ramzan, and N. Javaid, ‘‘Energy management in smart sectors using
fog based environment and meta-heuristic algorithms,’’ IEEE Access,
vol. 7, pp. 157254–157267, 2019, doi: 10.1109/ACCESS.2019.2949863.
[44] A. Shrestha, R. Bishwokarma, A. Chapagain, S. Banjara, S. Aryal, B. Mali,
R. Thapa, D. Bista, B. P. Hayes, A. Papadakis, and P. Korba, ‘‘Peer-to-peer
energy trading in micro/mini-grids for local energy communities: A review
and case study of Nepal,’’ IEEE Access, vol. 7, pp. 131911–131928, 2019,
doi: 10.1109/ACCESS.2019.2940751.
[45] R. Faia, J. Soares, T. Pinto, F. Lezama, Z. Vale, and J. M. Corchado,
‘‘Optimal model for local energy community scheduling considering peer
to peer electricity transactions,’’ IEEE Access, vol. 9, pp. 12420–12430,
2021, doi: 10.1109/ACCESS.2021.3051004.
[46] H. Zhu, K. Ouahada, and A. M. Abu-Mahfouz, ‘‘Peer-to-peer energy
trading in smart energy communities: A Lyapunov-based energy control
and trading system,’’ IEEE Access, vol. 10, pp. 42916–42932, 2022, doi:
10.1109/ACCESS.2022.3167828.
[47] N.-K. Kim, M.-H. Shim, and D. Won, ‘‘Building energy management
strategy using an HVAC system and energy storage system,’’ Energies,
vol. 11, no. 10, p. 2690, Oct. 2018, doi: 10.3390/en11102690.
[48] (2022). How Enel X’s DER.OS Software Maximizes the Value of
Battery Energy Storage. Accessed: Jun. 9, 2022. [Online]. Available:
https://www.enelx.com/n-a/en/resources/videos/enel-x-der-optimization-
software-maximizes-value
[49] (2022). Manage Energy and Digitalize Power Plants. Accessed:
Jun. 9, 2022. [Online]. Available:https://www.sma.de/en/products/
monitoring-control/power-plant-manager.html#c463285
[50] (2022). Optimize the Value of Your Hybrid Power System—From Utility-
Scale and Distributed Generation to Standalone Microgrids. Accessed:
Jun. 9, 2022. [Online]. Available: https://www.homerenergy.com/
[51] (2022). Solar + Storage Deployed Simply. Accessed: Jun. 9, 2022.
[Online]. Available: https://www.energytoolbase.com/
[52] B. Khosla, ‘‘Tesla, Inc. A case study,’’ Int. J. Sci. Res., vol. 10,
no. 12, pp. 45–69, Dec. 2021. [Online]. Available: https://www.ijsr.net/
get_abstract.php?paper_id=SR211128172051
[53] A. Dolara, S. Leva, and G. Manzolini, ‘‘Comparison of different physical
models for PV power output prediction,’’ Sol. Energy, vol. 119, pp. 83–99,
Sep. 2015.
VOLUME 11, 2023 18687

E. Giglio et al.: Efficient Artificial Intelligence Energy Management System
[54] Solcast. Accessed: Dec. 10, 2022. [Online]. Available: www.solcast.com
[55] G. Rancilio, A. Lucas, E. Kotsakis, G. Fulli, M. Merlo, M. Delfanti, and
hboxM. Masera, ‘‘Modeling a large-scale battery energy storage system
for power grid application analysis,’’ Energies, vol. 12, no. 17, p. 3312,
Aug. 2019.
[56] A. Carnovale and X. Li, ‘‘A modeling and experimental study of
capacity fade for lithium-ion batteries,’’ Energy AI, vol. 2, Nov. 2020,
Art. no. 100032.
[57] (2020). Fondazione Collegio Universitario Einaudi, Guida Ai Servizi:
Sezione Crocetta. [Online]. Available: https://www.collegioeinaudi.it/
wpcontent/uploads/2021/09/guida-ai-servizi-collegio-
einaudisezionecrocetta.pdf
[58] (2019). Prezzi Medi Mensili Per Fascia Oraria e Zona di
Mercato, Gestore dei Servizi Energetici S.P.A. [Online]. Available:
https://www.mercatoelettrico.org/it/Statistiche/ME/DatiSintesi.aspx
[59] G. Piggot. (2015). Electricity Consumption in a Sample of London House-
holds. London Datastore. [Online]. Available: https://data.london.gov.
uk/blog/electricity-consumption-in-a-sample-of-london-households/
[60] Acorn Energy Group. Accessed: Dec. 10, 2022. [Online]. Available:
http://acornenergy.com/
[61] E. Borretta, E. Giglio, G. Luzzani, V. Terranova, G. Trivigno, F.Grimaccia,
and A. Niccolai, ‘‘Software-based solutions for the optimization of
a building electric bill using integrated PV and storage systems: A
case study,’’ in Proc. IEEE Int. Conf. Environ. Elect. Eng. IEEE Ind.
Commercial Power Syst. Eur. (EEEIC/I&CPS Eur.), Sep. 2021, pp. 1–6.
[62] F. Cucchiella, I. D’Adamo, and M. Gastaldi, ‘‘Economic analysis of a
photovoltaic system: A resource for residential households,’’ Energies,
vol. 10, no. 6, p. 814, Jun. 2017, doi: 10.3390/en10060814.
[63] M. Chandel, G. D. Agrawal, S. Mathur, and A. Mathur, ‘‘Techno-
economic analysis of solar photovoltaic power plant for garment zone of
Jaipur city,’’ Case Stud. Thermal Eng., vol. 2, pp. 1–7, Mar. 2014, doi:
10.1016/j.csite.2013.10.002.
[64] A. Vargiu, R. Novo, C. Moscoloni, E. Giglio, G. Giorgi, and G. Mattiazzo,
‘‘An energy cost assessment of future energy scenarios: A case study on
san pietro island,’’ Energies, vol. 15, no. 13, p. 4535, Jun. 2022.
[65] I. D’Adamo, M. Gastaldi, and P. Morone, ‘‘The impact of a subsidized
tax deduction on residential solar photovoltaic-battery energy storage
systems,’’ Utilities Policy, vol. 75, Apr. 2022, Art. no. 101358, doi:
10.1016/j.jup.2022.101358.
[66] SolarPower Europe. (2020). EU Market Outlook for Solar Power
2020–2024. [Online]. Available: https://www.solarpowereurope.org/wp-
content/uploads/2020/12/3520-SPE-EMO-2020-report-11-
mr.pdf?cf_id=23124
[67] L. P. Bloomberg. (2018). BloombergNEF’s New Energy Outlook 2018.
[Online]. Available: https://about.bnef.com/new-energy-outlook/
[68] Energy Storage Technology and Cost Characterization Report,
HydroWIRES U.S Department of Energy, Washington, DC, USA,
2019.
[69] E. Drury, P. Denholm, and A. Margolis, The Impact of Different Economic
Performance Metrics on the Perceived Value of Solar Photovoltaics,
document NREL/TP-6A20-52197, NREL: Nat. Reewable Energy Lab.,
2011.
[70] S. Bracco, E. Bianchi, G. Bianco, A. Giacchino, A. Ramaglia, and
F. Delfino, ‘‘On the participation of small-scale high performance
combined heat and power plants to the Italian ancillary services market
within virtually aggregated mixed units,’’ Energy, vol. 239, Jan. 2022,
Art. no. 122275, doi: 10.1016/j.energy.2021.122275.
[71] P. Escobar, E. Martínez, J. C. Saenz-Díez, E. Jiménez, and J. Blanco,
‘‘Profitability of self-consumption solar PV system in Spanish households:
A perspective based on European regulations, Renew. Energy, vol. 160,
pp. 746–755, Nov. 2020, doi: 10.1016/j.renene.2020.06.135.
[72] M. L. D. Silvestre, M. G. Ippolito, E. R. Sanseverino, G. Sciumè, A. Vasile,
‘‘Energy self-consumers and renewable energy communities in Italy:
New actors of the electric power systems,’’ Renew. Sustain. Energy Rev.,
vol. 151, Nov. 2021, Art. no. 111565, doi: 10.1016/j.rser.2021.111565.
[73] A. Sangswang and M. Konghirun, ‘‘Optimal strategies in home energy
management system integrating solar power, energy storage, and vehicle-
to-grid for grid support and energy efficiency,’’ IEEE Trans. Ind. Appl.,
vol. 56, no. 5, pp. 5716–5728, Sep. 2020, doi: 10.1109/TIA.2020.2991652.
[74] F. Moscetti, S. Paoletti, and A. Vicino, ‘‘Analysis and models of
electricity prices in the Italian ancillary services market,’’ in Proc.
IEEE PES Innov. Smart Grid Technol., Eur., Oct. 2014, pp. 1–6, doi:
10.1109/ISGTEurope.2014.7028755.
ENRICO GIGLIO received the B.S. and M.S.
degrees in mechanical engineering from the
Politecnico di Torino, in 2019 and 2021, respec-
tively, where he is currently pursuing the Ph.D.
degree in mechanical engineering.
His research interest includes short-term energy
scenario models, with a focus on islanded systems
with high-RES penetration.
GABRIELE LUZZANI received the B.S. and
M.S. degrees in aerospace engineering from the
Politecnico di Torino with dual degree pathway
‘‘Alta Scuola Politecnica,’’ where he is currently
pursuing the Ph.D. degree in aerospace engineer-
ing in partnership with Leonardo S.p.A.
His research interest includes physiological
response to stress and cognitive workload in a
safety-critical environment, focusing on the civil
aviation context.
VITO TERRANOVA received the B.S. degree in
aerospace engineering and the M.S. degree in
aeronautical engineering from the Politecnico di
Milano University, in 2019 and 2022, respectively.
His master’s thesis focused on combining numer-
ical methods and machine learning techniques to
solve Burgers’ equation.
He is specialized in aerodynamics and compu-
tational fluid dynamics (CFD).
GABRIELE TRIVIGNO (Graduate Student Mem-
ber, IEEE) received the B.S. degree in computer
engineering and the M.S. degree in data science
and engineering from the Politecnico di Torino, in
2019 and 2021, respectively, where he is currently
pursuing the Ph.D. degree in AI for computer
vision.
His main research interest includes visual
place recognition, with a focus on developing
fully learnable scalable and cross-domain robust
methods.
ALESSANDRO NICCOLAI (Member, IEEE)
received the bachelor’s and master’s degrees
(summa cum laude) in mechanical engineering and
the Ph.D. degree (cum laude) in electrical engi-
neering from the Politecnico di Milano, in 2014,
2016, and 2019, respectively. He is currently a
Researcher with the Politecnico di Milano. His
main research interest includes computational
intelligence, focused on evolutionary optimization
algorithms and neural networks.
FRANCESCO GRIMACCIA (Senior Member,
IEEE) received the M.S. and Ph.D. degrees
(cum laude) in electrical engineering from the
Politecnico di Milano, Milan, Italy, in 2003 and
2007, respectively. He is currently an Associate
Professor with the Energy Department, Politecnico
di Milano. His main research interest includes soft-
computing technique development and applica-
tions in different electric engineering fields, such
as unmanned systems, photovoltaics, and other
energy-harvesting devices. He is a member of the IEEE Computational
Intelligence Society and the Vice-President of the Associazione Italiana di
Elettrotecnica, Elettronica and the Automazione, Informatica e Telecomuni-
cazioni, Milan Section.
Open Access funding provided by ‘Politecnico di Torino’ within the CRUI CARE Agreement
18688 VOLUME 11, 2023