Conference PaperPDF Available

An Enhanced Convolutional Neural Network model based on weather parameters for short-term electricity supply and demand

Conference Paper

An Enhanced Convolutional Neural Network model based on weather parameters for short-term electricity supply and demand

Abstract and Figures

Short-term electricity supply and demand forecasting using weather parameters including: temperature, wind speed, and solar radiations improve the operational efficiency and accuracy of power systems. There are many weather parameters which have influential affect on the supply and demand of electricity, but temperature, solar radiations, and wind speed are the most important parameters. Our proposed time series model is based on preprocessing, feature extraction, data preparation, and Enhanced Convolutional Neural Network referred as ECNN module for short-term weather parameters forecasting up to 6-hours ahead. The proposed ECNN time series model is applied on 61 locations of United States, collected from National Solar Radiation Database (NSRDB). Model trained on 15-years data and validated on additional two-years out of sample data. Simulation result shows that our proposed model performs better than traditional benchmark models in terms of Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Relative Root Mean Square Error (RMSE%) performance metrics. Result shows that the proposed model is effective for short-term forecasting of temperature, solar radiations, and wind speed. Moreover, proposed model improves the accuracy and operational efficiency of power systems.
Content may be subject to copyright.
An Enhanced Convolutional Neural Network
model based on weather parameters for
short-term electricity supply and demand
Zeeshan Aslam, Nadeem Javaid, Muhammad Adil, Muhammad Tariq Ijaz, Atta ur
Rahman, and Mohsin Ahmed
Abstract Short-term electricity supply and demand forecasting using weather pa-
rameters including: temperature, wind speed, and solar radiations improve the op-
erational efficiency and accuracy of power systems. There are many weather pa-
rameters which have influential affect on the supply and demand of electricity, but
temperature, solar radiations, and wind speed are the most important parameters.
Our proposed time series model is based on preprocessing, feature extraction, data
preparation, and Enhanced Convolutional Neural Network referred as ECNN mod-
ule for short-term weather parameters forecasting up to 6-hours ahead. The proposed
ECNN time series model is applied on 61 locations of United States, collected from
National Solar Radiation Database (NSRDB). Model trained on 15-years data and
validated on additional two-years out of sample data. Simulation result shows that
our proposed model performs better than traditional benchmark models in terms of
Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Relative Root
Mean Square Error (RMSE%) performance metrics. Result shows that the proposed
model is effective for short-term forecasting of temperature, solar radiations, and
wind speed. Moreover, proposed model improves the accuracy and operational effi-
ciency of power systems.
1 Introduction
Electric power industry plays very important role in the well being of a country, it’s
efficient performance helps in the economic development of a country. One of the
main mode of electric power generation is thermal, which is costly and emits car-
bon in large amount . Rapidly growing interests in renewable energy sources such
as wind and solar power and decreasing cost of power generation, renewable en-
Zeeshan Aslam, Nadeem Javaid (Corresponding Author), Muhammad Adil, Muhammad Tariq Ijaz,
Atta ur Rahman, and Mohsin Ahmed
COMSATS University Islamabad, Pakistan, email: nadeemjavaidqau@gmail.com
1
2 Authors Suppressed Due to Excessive Length
ergy generation gains lot of importance because of less expensive and no carbon
emission. Globally, wind and solar power generation grows rapidly from 80 to 790
GW from 2006 to 2016. Accurate forecasting is important because 1% reduction in
Mean Absolute Percentage Error (MAPE) means 0.1% to 0.3% reduction in gener-
ation cost of electricity, which is approximately $1 million annually on large scale
smart grid.
The supply and demand of electricity is highly dependent on weather parame-
ters, therefore, market participants need a useful and reliable technique to increase
their profit ratio by accurately forecasting weather parameters. Weather parame-
ters have great affect on both supply and demand side of electricity. From large
number of weather parameters, temperature, wind speed, and solar radiations are
the most influencing factors in supply and demand of electricity. Temperature has
great effects on the individual side of electricity, whereas, wind speed and solar
radiations effected the on supply side of electricity. Accurate short-term forecast-
ing of these weather parameters is important for several reasons including: efficient
supply management of electricity, to reduce amount of electricity consumption, to
improve energy efficiency level of stations, to prepare effective production plans, to
improve operational efficiency, and to adjust and control power stations [1-2]. How-
ever, inaccurate forecasting of these weather parameters is one of the most important
challenge in supply and demand of electricity.
Traditionally, short-term weather parameters forecasting techniques are based
on statistical models such as Vector Autoregression (VAR) model, regression tech-
niques, and artificial intelligence models such as Support Vector Machine (SVM),
Artificial Neural Networks (ANNs), and Deep learning models. These models re-
quire extensive memory and computation time, slow convergence, less accurate,
and require weights adjustment and heavy preprocessing in order to refine input
which make them less accurate, and lead to overfitting problem. Moreover, as the
size of data increases, these models become complex and requires more training
time [2]. Based on above mentioned limitations of traditional models, we propose
ECNN weather model for short-term forecasting of three weather parameters in-
cluding: temperature, wind speed, and solar radiations using 61 locations data of
United States.
In the light of above mentioned weather parameters forecasting techniques, this
paper have following contributions. First, we identify and remove outliers from
weather parameters time series. Second, we apply Fast Independent Component
Analysis (FastICA) to reduce dimensionality of features. Third, we apply Grange
Causality (GCA) and Augmented Dickey-Fuller (ADF) tests to check the useful-
ness and stationarity of time series. After that, we find and adjust trend and seasonal
patterns in time series. Fourth, we employ ECNN time series model for short-term
forecasting of three weather parameters to overcome the limitations of above men-
tioned traditional models. Finally, we evaluate the performance of our proposed
model with traditional benchmark models according to three accuracy metrics in-
cluding: RMSE, RMSE%, and MAE.
Rest of this paper is organized as follows. In section II, we investigate and sum-
marize the related work of short-term weather parameters forecasting. In section III,
Title Suppressed Due to Excessive Length 3
we present our proposed methodology. In section IV, we examine the forecasting
performance using simulation results. Finally, in section V, we draw conclusion and
future work of this paper.
2 Related Work
Related work based on weather parameters have been extensively studied and cat-
egorized into four sections: solar radiations forecasting, wind speed forecasting,
temperature or load forecasting, and multivariate weather parameters forecasting. In
paper [26, 27, 28, 31, 32], authors propose solutions for short-term to medium-term
electricity price and load forecasting. In [29,30], authors perform wind and photo-
voltaic power forecasting. Table 1 shows the summary of further related work based
on short-term weather parameters forecasting.
3 Methodology
Modeling three weather parameters for short-term forecasting first requires prepro-
cessing of input features. Afterward, it requires appropriate extraction and prepa-
ration of input features to reduce dimensionality and keep necessary information.
Then, define a suitable structure of ECNN to perform short-term forecasting of input
weather parameters. Following sub-sections describe each module of our proposed
system model, before describing these steps in detail, an overview of methodology
is presented.
3.1 Overview of methodological approach
In this paper, we propose an approach for short-term forecasting of multivariate time
series based on historical input features known as: wind speed, solar radiations, and
temperature. In our proposed approach, first apply preprocessing on input features,
then identify and remove outliers using Z score technique, and lastly normalize in-
put features using min-max normalization (see Section 3.3). After preprocessing
of input features, an appropriate feature extraction technique FastICA is applied to
reduce the dimensionality of features (see Section 3.4). After preprocessing and ex-
traction of relevant data, identify and adjust its specific characteristics and seasonal
patterns (see Section 3.5) in order to prepare data for forecasting. After the prepara-
tion of input features, the ECNN model is develop and pass input data to ECNN. In
order to define the structure of ECNN, it is necessary to define elements of ECNN
(see Section 3.6). Overview of proposed approach is presented in Fig. 1.
4 Authors Suppressed Due to Excessive Length
Table 1: Summary of short-term weather parameters forecasting
Proposed
Method
Description Limitations
VAR, 2018 [1] Authors employ VAR model for weather parame-
ters.
Linear in nature, need more memory, high compu-
tation time and preprocessing.
WKNNRW,
2018 [3]
Authors propose a hybrid model for multivariate
time series forecasting.
Overfitting, extensive computation and memory
requirements, and random weights selection.
DNN, 2018 [4] Authors apply deep learning to forecast solar radi-
ations.
Overfitting, local optima, and slow convergence.
CNN-LSTM,
2018 [5]
Authors propose a hybrid model to forecast solar
power.
Overfitting and computationally expensive.
DNN-TPE,
2018 [6]
Authors propose a global DNN model for solar ir-
radiance forecasting.
Overfitting, poor accuracy, and TPE become ex-
pensive on large data set.
VMD-RMWK,
2018 [7]
Authors propose a hybrid model for solar irradi-
ance forecasting.
Computationally expensive, more training time,
more memory, and Less accurate on large data set.
IEMD-ARIMA-
WNN-FOA,
2018 [8]
Authors propose a hybrid model for short-term
load forecasting.
Computationally expensive, overfitting, more
training time, and less accurate on large data set.
PSR-BSK, 2018
[9]
Authors propose a hybrid model for short-term
load forecasting.
Require parameter tuning, overfitting, extensive
memory and computation requirements, and less
accurate.
SVR, 2017 [10] Authors propose SVR model for short-term fore-
casting of demand response.
Overfitting, slow convergence, requires more
memory, less accurate, and computationally ex-
pensive on large data set.
EWT-LSTM-
RELM-IEWT,
2018 [11]
Authors propose a hybrid model to forecast wind
speed.
Computationally expensive, overfitting, requires
more memory, complex and expensive on large
data set.
RWT-ARIMA,
2019 [12]
Authors propose a hybrid model for short-term
forecasting of wind speed.
Non-linear characteristics, poor accuracy, less ac-
curate and expensive on large data set.
CBA-ANN-
SVM, 2019
[2]
Authors propose a hybrid model to forecast load
and weather parameters.
Overfitting, local optima and slow convergence,
and requires more memory.
VMD-GSO-
ELM, 2018
[13]
Authors propose a hybrid model for wind speed
forecasting.
Overfitting, extensive memory and computation
requirements, and GSO become expensive on
large data set.
ICEEMDAN-
GWO, 2018
[14]
Authors propose a hybrid model to forecast wind
speed.
Overfitting, extensive memory and computation
requirements, and GWO become expensive on
large data set.
EEMD-AWNN,
2018 [15]
Authors propose a hybrid model for short-term
wind speed forecasting.
Overfitting, slow convergence, and extensive
memory and computation requirements.
3.2 Weather Data
We apply our proposed model on 61 locations of United States, collected from Na-
tional Solar Radiation Database (NSRDB), which is prepared by National Renew-
able Energy Laboratory, National Climate Data center, and other partners [16]. We
use three input parameters from data set. Input parameters that have influential ef-
fect on both the supply and demand side of electricity such as temperature in Kelvin
Title Suppressed Due to Excessive Length 5
Fig. 1: Overview of proposed system model
(K), wind speed in meters per second (m/s), and global solar radiations in Watt hour
per square meter (Wh/m2). We use 15-years data from 1991-2006 to train model and
additional 2-years data for validation.
3.3 Data Preprocessing
In preprocessing module, we perform three operations: data cleansing and analyz-
ing, outliers detection and removal, and data normalization. In data cleansing and
analyzing, we first analyze the data by identifying the type of time series and con-
taining them null or incorrect values, then cleaning those values through filling them
with mean value of time series. In outliers detection and removal, we identify values
using Z score which reside outside the distribution and do not have any major influ-
ence on final output. An outlier has serious impact on mean and standard deviation,
and causes to skew the data. In paper [17], authors define the formula of Z score to
identify and remove outliers, which is defined as:
Zscore = (Observat ion Mean)/StandardDeviation.(1)
After first two operations in preprocessing module, we normalize the time series
using Min-Max normalization that make it easy for ECNN to handle data in same
scale and increases the training speed of ECNN. It scales the input features between
0 and 1. In paper [18], authors define the normalization formula which is defined as:
6 Authors Suppressed Due to Excessive Length
x0= (xmax xmin)(xixmin )
(xmax xmin)+xmin ,(2)
where (xmax xmin) = 0 when (xmax xmin ) = 0 for a feature which shows a constant
value for that input feature. Preprocessing module improves the accuracy of our
proposed system, reduces required memory, training time, complexity of model,
and overfitting problem.
3.4 Feature Extraction
In feature extraction module, we apply FastICA, which is computationally powerful
method for estimation of independent component analysis. It is 10-100 times faster
than traditional methods for independent component analysis task, which are based
on gradient descent approach. FastICA is used as feature extraction technique to re-
duce dimensionality of features, while retaining key information. It compresses the
features that take less memory to store these features [19]. By reducing dimension-
ality of features, it reduces: the amount of memory required to store features, com-
plexity of model, training time of model, and improves visualization of data because
in high dimensions it is very difficult to understand and visualize data. FastICA is
similar to Principle Component Analysis (PCA) technique that maps collection of
features to uncorrelated features, whereas, FastICA do more by maximizing the sta-
tistical independence (or minimize mutual information) rather than developing just
uncorrelated features. FastICA performs better than PCA and easy to use [19]. It
reduces number of variables in time series.
3.5 Data Preparation
In data preparation module, we first apply Granger Causality (GCA) test, which
is statistical hypothesis test to check weather a time series is useful in forecasting
other time series. GCA is used for time series analysis. GCA result shows that if the
probability outcome is less than any αlevel, then the hypothesis would be rejected at
that level [20]. In [20], authors describe the equation for GCA test which is defined
as:
X(t) =
p
j=1
AjX(tj) + E1(t),(3)
where pis the lagged observations, the matrix Acontains the coefficients, E1 repre-
sents residual (prediction errors) for each time series. After that we find trend and
seasonal patterns in time series and adjust them. Time series data are mainly com-
posed of trend and seasonal patterns. It is possible to decompose the time series
data into major sub-components such as trend and seasonal components to check
their affects on time series data [21]. There are two different decomposition models:
Title Suppressed Due to Excessive Length 7
additive decomposition and multiplicative decomposition. In this paper, we use ad-
ditive decomposition model to get the trend and seasonal components of time series,
which is defined as:
Xt =t rend(T t) + Seasonal (St) + random.(4)
We use additive decomposition model because additive decomposition is useful to
find trend and seasonal components when time series change with respect to changes
in weather and do not vary much [21]. Additive model works more efficient in our
time series data than multiplicative model. After finding trend and seasonal com-
ponents in time series, we make the series stabilize by differencing method. Dif-
ferencing is the most popular method that make the series stable by reducing or
eliminating trend and seasonality [22], say,
xt=ytyt1,(5)
where ytis original time series and yt1is the lagged version of original time series.
Then apply Augmented Dickey-Fuller (ADF) test to ensure that weather the time
series is stationary or not. The ADF test result shows that time series is stationary
with p-values less than 0.01 for all locations. Results of differencing and ADF sug-
gest that stationarity is not an issue with our time series.
3.6 Model Structure
For the structure of ECNN, certain elements have to be defined as: number of hidden
layers to use in ECNN, activation function, type of optimizer, padding size, window
size, and adjustment of regularization terms to avoid overfitting and increase fore-
cast accuracy. As, for the choice of these decisions, there is no optimal choice given
in literature. All of these elements of proposed ECNN found by a trial and error pro-
cess. In ECNN, we use Leaky Relu activation function, which is most commonly
used activation function in deep neural networks. Leaky Relu improves the training
process and reduces vanishing gradient problem of ECNN. For ECNN learning, we
use Adaptive moment estimation (Adam) optimizer, which is extension of stochastic
gradient descent to optimize the model. Moreover, the choice of hidden layers and
activation function depend on the type of data and problem to solve. Fig. 2 shows
the structure of ECNN.
4 ECNN: Our Proposed Model
CNN is a deep neural network, which uses multiple layered neural network structure
to represent the information. CNN was first proposed [23] for automatic classifica-
8 Authors Suppressed Due to Excessive Length
tion of digit images. One of the most important property of CNN is that it continues
to improve as the size of data increases. It successfully improves the performance
with less memory requirements, because CNN is fully connected network and it
has parameter sharing property. We propose an enhanced version of CNN, named
as ECNN, in which we add some additional hidden layers and adjust parameters of
hidden layers to avoid overfitting problem and improve model performance. CNN
works in two parts; in first part, CNN learns the high level features from the given
input features with weight sharing property, and in second part, CNN flattens the
output of above layers and perform prediction. In ECNN, we add one convolution
layer with filter size two using Leaky Relu as activation function. Mathematical
equation of convolution layer to perform convolution operation is defined as in [4];
Oj,k=f(
c
l=0
c
m=0
wk,lij+l,k+m),(6)
where fis the activation function, wis the weight values of kernel, and iis the input
features. In most of deep neural networks, relu is used as activation function. Relu
has key advantages over other activation functions that it does not activate all the
neurons at the same time. However, the limitation of relu activation function is that
it saturates at the negative region, which means that gradient at negative region is
zero. When the gradient is zero, during back propagation all the weights will not be
updated. To overcome such limitation, we use Leaky Relu. In general, it solves the
dying relu problem. After convolution layer we add two dense layers, dense layer
represents a matrix vector multiplication, and the values in the matrix are trainable
parameters that get updated during back propagation. Dense layer is fully connected
layer whose neurons receive input from all the neurons of previous layers. After the
dense layer, dropout layer comes which is used to prevent overfitting problem. Dur-
ing training time, at each iteration, number of neurons with some certain probability
is temporarily dropped. The reason is that dropout prevents the network to be de-
pendent on a small number of neurons and force every neuron to be able to operate
independently, which increases the accuracy, shortens the training time and combats
Fig. 2: Proposed Enhanced Convolutional Neural Network
Title Suppressed Due to Excessive Length 9
overfitting. The purpose of dropout layer is to not rely on some or combination of
neurons, but to learn different representations to avoid overfitting.
In ECNN, we add two Maxpooling layers. Maxpooling layer reduces the amount
of parameters, model computation, dimensionality, and control overfitting by reduc-
ing the spatial size of the network. Key advantage of pooling operation is to generate
small feature maps, which summarize the large input feature maps. After that flatten
layer flattens the output of above layers in order to feed next fully connected layers
to perform prediction. Output of a jth hidden layer neuron can be calculated using
following equation:
oj=f(b+
p
l=1
wj,lol),(7)
where fis the activation function, w(j,l)is the weight between neurons ojand ol,
and pis the total number of neurons [4]. Fig. 2 shows the proposed ECNN model.
ECNN trains using Adam optimizer to update the weights of model iteratively.
Adam has key some key advantages over traditional stochastic gradient descent
optimization algorithms, which are defined as: it is straightforward to implement,
computationally efficient, requires less memory, best for problems that are large in
terms of data or parameters, and appropriate for noisy problems [24]. Following are
the key equations to update the model parameters:
mt=β1mt1+ (1β1).gt,(8)
vt=β2vt1+ (1β2).g2
t,(9)
where mand vare moving averages, gis gradient along time t, and betas are hyper-
parameters of the algorithm.
4.1 Performance metrics
In order to evaluate the model performance, there are many performance evaluation
metrics available in literature. In this paper, we consider three standard performance
metrics, which are defined by [30] as: MAE, RMSE, and RMSE%. Following are
the equations of these metrics; MAE is defined as:
MAE =1
N
N
i=0
|FiOi|,(10)
RMSE is defined as:
RMSE =s1
N
N
i=0
(FiOi)2,(11)
and RMSE% is defined as:
10 Authors Suppressed Due to Excessive Length
RMSE %=q1
NN
i=0(FiOi)2
1
NN
i=0Oi
,(12)
where Nis the total number of input samples, Fiis the actual value, and Oiis the
predicted value. In this paper, we compare our proposed model with existing bench-
mark models including: SVM, VAR, and ANN.
5 Simulation Results
In this section, we discuss the results of our proposed ECNN time series model with
existing benchmark models. We evaluate the performance of our proposed model
with existing benchmark models in terms of MAE, RMSE, and RMSE% perfor-
mance metrics which are describe in Sub-section 4.1. Fig. 3 shows that the proposed
ECNN has very less average (among 61 locations) MAE and RMSE as compared
to other existing models. In terms of accuracy, ECNN performs better than existing
benchmark models. Fig. 4 shows ECNN forecasting results of wind speed, solar ra-
diations, and temperature for California state.
Table II and III summarize the ECNN forecasting results of wind speed, solar radi-
ations, and temperature according to three performance metrics (i.e., MAE, RMSE,
and RMSE%). These tables show the average MAE, RMSE, and RMSE% for three
weather parameters. For weather parameters forecasting, our results clearly indicate
that the proposed ECNN model has very less error rate and more accurate than tra-
ditional models. Furthermore, the execution time of ECNN is less than the VAR
benchmark model. VAR execution time for one location is 6 minutes and 25 sec-
onds, while ECNN has 3 minutes and 50 seconds. Results concluded that our pro-
posed model have better performance than other benchmark models for supply and
demand forecasting of electricity.
6 Conclusion
In order to solve the problem of short-term weather parameters forecasting, we pro-
pose an ECNN weather model based on three weather parameters including: temper-
ature, wind speed, and solar radiations. The proposed model is based on four major
modules. In first module, we perform preprocessing of time series data including:
data analyzing and cleansing, outliers detection and removal, and data normaliza-
tion. In second module, we perform feature extraction using FastICA. FastICA is
used to reduce dimensionality, model complexity, and improve model training. In
third module, we prepare time series data by applying GCA test to check the use-
fulness of time series, find and adjust trend and seasonality in time series, and ADF
Title Suppressed Due to Excessive Length 11
Fig. 3: Average 6-hours MAE, RMSE, and RMSE% of ECNN and other models
Table 2: ECNN average MAE, RMSE, and RMSE% of wind and solar forecasting
Forecast
Horizon
MAE (m/s) RMSE
(m/s)
RMSE%
(%)
1-hour
ahead
0.0084 0.0103 23.15
2-hour
ahead
0.0095 0.0109 26.17
3-hour
ahead
0.0098 0.0115 28.21
4-hour
ahead
0.0105 0.0121 35.62
5-hour
ahead
0.0107 0.0124 38.67
6-hour
ahead
0.0110 0.0126 38.90
Forecast
Horizon
MAE
(Wh/m2)
RMSE
(Wh/m2)
RMSE%
(%)
1-hour
ahead
0.0114 0.0135 17.15
2-hour
ahead
0.0116 0.0138 20.17
3-hour
ahead
0.0116 0.0140 22.21
4-hour
ahead
0.0118 0.0142 25.12
5-hour
ahead
0.0124 0.0145 27.41
6-hour
ahead
0.0126 0.0146 30.90
test to check stationarity of time series. In fourth module, we perform short-term
forecasting by employing proposed ECNN model. Therefore, the proposed model
gets the advantages of deep neural networks. The simulation results based on two-
years real world time series weather data shows that the proposed model has more
12 Authors Suppressed Due to Excessive Length
Fig. 4: ECNN wind speed, solar radiations, and temperature forecasting results
Table 3: ECNN average MAE, RMSE, and RMSE% of temperature forecasting
Forecast Hori-
zon
MAE (K) RMSE (K) RMSE%
(%)
1-hour ahead 0.0988 0.9886 0.15
2-hour ahead 0.0988 0.9889 0.25
3-hour ahead 0.0989 0.9890 0.28
4-hour ahead 0.0989 0.9895 0.32
5-hour ahead 0.0989 0.9898 0.35
6-hour ahead 0.0990 0.9899 0.40
accurate and effective results than existing benchmark models. Altogether it can be
concluded that developing a deep model is complex process, especially when it is
applied to weather parameters. Hence, the deep model development requires exten-
sive effort to determine the optimal number of hidden layers and their parameters
such as hyper-parameters tuning, since there are no clear instructions available for
such process. In future, we enhance the proposed system by incorporating more
weather parameters and improves its performance. Furthermore, we will enhance
Title Suppressed Due to Excessive Length 13
the performance of proposed model through optimization techniques by tuning their
hyper-parameters.
References
1. L. Yixian, M. C. Roberts, and R. Sioshansi, ”A vector autoregression weather model for elec-
tricity supply and demand modeling,” Journal of Modern Power Systems and Clean Energy,
vol. 6, pp. 763-776, 2018.
2. M. Torabi, S. Hashemi, M. R. Saybani, S. Shamshirband, and A. Mosavi, ”A Hybrid clus-
tering and classification technique for forecasting short?term energy consumption,” Environ-
mental Progress & Sustainable Energy, vol. 38, pp. 66-76, 2019.
3. K. Lang, M. Zhang, Y. Yuan, and X. Yue, ”Short-term load forecasting based on multivariate
time series prediction and weighted neural network with random weights and kernels,” Cluster
Computing, pp. 1-9, 2018.
4. K. Kaba, M. Sarıg¨
ul, M. Avcı, and H. M. Kandırmaz, ”Estimation of daily global solar radi-
ation using deep learning model,” Energy, vol. 162, pp. 126-135, 2018.
5. W. Lee, K. Kim, J. Park, J. Kim, and Y. Kim, ”Forecasting Solar Power Using Long-Short
Term Memory and Convolutional Neural Networks,” IEEE Access, vol. 6, pp. 73068-73080,
2018.
6. J. Lago, K. De Brabandere, F. De Ridder, and B. De Schutter, ”Short-term forecasting of solar
irradiance without local telemetry: A generalized model using satellite data,” Solar Energy,
vol. 173, pp. 566-577, 2018.
7. I. Majumder, P. Dash, and R. Bisoi, ”Variational mode decomposition based low rank robust
kernel extreme learning machine for solar irradiation forecasting,” Energy conversion and
management, vol. 171, pp. 787-806, 2018.
8. J. Zhang, Y.-M. Wei, D. Li, Z. Tan, and J. Zhou, ”Short term electricity load forecasting using
a hybrid model,” Energy, vol. 158, pp. 774-781, 2018.
9. G.-F. Fan, L.-L. Peng, and W.-C. Hong, ”Short term load forecasting based on phase space
reconstruction algorithm and bi-square kernel regression model,” Applied energy, vol. 224,
pp. 13-33, 2018.
10. Y. Chen, P. Xu, Y. Chu, W. Li, Y. Wu, L. Ni, et al., ”Short-term electrical load forecasting
using the Support Vector Regression (SVR) model to calculate the demand response baseline
for office buildings,” Applied Energy, vol. 195, pp. 659-670, 2017.
11. Y. Li, H. Wu, and H. Liu, ”Multi-step wind speed forecasting using EWT decomposition,
LSTM principal computing, RELM subordinate computing and IEWT reconstruction,” En-
ergy Conversion and Management, vol. 167, pp. 203-219, 2018.
12. S. Singh and A. Mohapatra, ”Repeated wavelet transform based ARIMA model for very
short-term wind speed forecasting,” Renewable energy, vol. 136, pp. 758-768, 2019.
13. C. Li, Z. Xiao, X. Xia, W. Zou, and C. Zhang, ”A hybrid model based on synchronous op-
timization for multi-step short-term wind speed forecasting,” Applied energy, vol. 215, pp.
131-144, 2018.
14. J. Song, J. Wang, and H. Lu, ”A novel combined model based on advanced optimization
algorithm for short-term wind speed forecasting,” Applied energy, vol. 215, pp. 643-658,
2018.
15. M. Santhosh, C. Venkaiah, and D. V. Kumar, ”Ensemble empirical mode decomposition based
adaptive wavelet neural network method for wind speed prediction,” Energy conversion and
management, vol. 168, pp. 482-493, 2018.
16. Wilcox, S. M. (2012). National solar radiation database 1991-2010 update: User’s manual
(No. NREL/TP-5500-54824). National Renewable Energy Lab.(NREL), Golden, CO (United
States).
17. P. J. Rousseeuw and M. Hubert, ”Robust statistics for outlier detection,” Wiley Interdisci-
plinary Reviews: Data Mining and Knowledge Discovery, vol. 1, pp. 73-79, 2011.
14 Authors Suppressed Due to Excessive Length
18. T. Jayalakshmi and A. Santhakumaran, ”Statistical normalization and back propagation for
classification,” International Journal of Computer Theory and Engineering, vol. 3, pp. 1793-
8201, 2011.
19. D. Langlois, S. Chartier, and D. Gosselin, ”An introduction to independent component anal-
ysis: InfoMax and FastICA algorithms,” Tutorials in Quantitative Methods for Psychology,
vol. 6, pp. 31-38, 2010.
20. H. L¨
utkepohl, New introduction to multiple time series analysis: Springer Science & Business
Media, 2005.
21. V. Prema and K. U. Rao, ”Time series decomposition model for accurate wind speed forecast,
Renewables: Wind, Water, and Solar, vol. 2, p. 18, 2015.
22. D. C. Montgomery, C. L. Jennings, and M. Kulahci, Introduction to time series analysis and
forecasting: John Wiley & Sons, 2015.
23. Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hubbard, et al., ”Back-
propagation applied to handwritten zip code recognition,” Neural computation, vol. 1, pp.
541-551, 1989.
24. D. P. Kingma and J. Ba, ”Adam: A method for stochastic optimization,” arXiv preprint
arXiv:1412.6980, 2014.
25. R. Marquez and C. F. Coimbra, ”Proposed metric for evaluation of solar forecasting models,
Journal of solar energy engineering, vol. 135, p. 011016, 2013.
26. Samuel, O., Alzahrani, F.A., Hussen Khan, R.J.U., Farooq, H., Shafiq, M., Afzal, M.K. and
Javaid, N., 2020. Towards Modified Entropy Mutual Information Feature Selection to Fore-
cast Medium-Term Load Using a Deep Learning Model in Smart Homes. Entropy, 22(1),
2020.
27. Khalid, R., Javaid, N., Al-zahrani, F.A., Aurangzeb, K., Qazi, E.U.H. and Ashfaq, T., 2020.
Electricity Load and Price Forecasting Using Jaya-Long Short Term Memory (JLSTM) in
Smart Grids. Entropy, 22(1), 2020.
28. Mujeeb, S. and Javaid, N., 2019. ESAENARX and DE-RELM: Novel schemes for big data
predictive analytics of electricity load and price. Sustainable Cities and Society, 51.
29. Mujeeb, S., Alghamdi, T.A., Ullah, S., Fatima, A., Javaid, N. and Saba, T., 2019. Exploiting
Deep Learning for Wind Power Forecasting Based on Big Data Analytics. Applied Sciences,
9(20).
30. Naz, A., Javaid, N., Rasheed, M.B., Haseeb, A., Alhussein, M. and Aurangzeb, K., 2019.
Game Theoretical Energy Management with Storage Capacity Optimization and Photo-
Voltaic Cell Generated Power Forecasting in Micro Grid. Sustainability, 11(10).
31. Naz, A., Javed, M.U., Javaid, N., Saba, T., Alhussein, M. and Aurangzeb, K., 2019. Short-
term electric load and price forecasting using enhanced extreme learning machine optimiza-
tion in smart grids. Energies, 12(5).
32. Mujeeb, S., Javaid, N., Ilahi, M., Wadud, Z., Ishmanov, F. and Afzal, M.K., 2019. Deep long
short-term memory: A new price and load forecasting scheme for big data in smart cities.
Sustainability, 11(4), p.987.
... They incorporate the generation of electric power through cheap and efficient resources and further distributed to electricity users through the power transmission and distribution lines [25]. Moreover, smart grids allow an efficient integration of renewable energy sources [26]- [29]. Hence, many researchers provide solution for efficient energy management [30]- [35]. ...
Thesis
Electricity is widely used around 80% of the world, which depicts the significance of secure and efficient use of electricity. Nontechnical losses (NTLs) become one of the biggest issues for electric utilities around the world. The intentional malfunctioning with electric meters and false data injection cover the largest proportion of NTLs and have hazardous effects on the power systems. The digitalization of traditional grids gives a new revolution for an efficient exchange of information over short periods, which allows the electric utilities to devise innovative datadriven solutions for electricity theft detection (ETD). The existing data-driven approaches for ETD have limited ability to handle high-dimensional, noisy and imbalanced data. Moreover, these approaches have limited potential to derive features' associations during feature extraction. These limitations raise the misclassification rate, which makes the existing ETD approaches unacceptable for electric utilities. Therefore, in this thesis, we propose a new data-driven methodology, which consists of four new solutions to systematically detect the electricity fraudsters in the smart grid environment. Particularly, in first system model, we present a new class balancing mechanism based on the interquartile minority oversampling technique to handle the data imbalance issues. Then, a combined ETD model composed of long short-term memory (LSTM), UNet and adaptive boosting (Adaboost), termed as LSTM-UNet-Adaboost, is presented to detect electricity frauds. Afterwards, in second solution, we introduce a new mechanism that is based on two scenarios. In the first scenario, a new supervised learning based mechanism is presented, which is a combination of UNet and generative adversarial network (GAN), named as UNet-GAN. The GAN's structure is mainly comprised of two neural networks: generator and discriminator. Due to the excellent performance of UNet, we utilize it in both generator and discriminator parts. These two neural networks contest with each other in a game-theoretic manner to significantly boost the ETD performance. In the second scenario, we propose a new dynamic learning based semi-supervised solution, which consists of probabilistic guider (PG) and Ladder network. This solution is termed as PG-Ladder network. The PG dynamically guides the proposed PG-Ladder network to further improve its performance in terms of ETD. Furthermore, the conventional approaches require extensive experts' involvement and lose data relationships during feature extraction for effective theft detection. Therefore, in third system model, we solve these issues by presenting the new solution that is based on relational denoising autoencoder (RDAE) with the attention guided (AG) TripleGAN, named as RDAE-AG-TripleGAN. The limitations of conventional clustering mechanisms and scarcity of labeled electricity consumption (EC) data are solved by presenting the new two-fold end-to-end semi-supervised solution, referred as fourth solution. In the first fold, it groups the similar EC cases by employing the grey wolf optimization (GWO) based clustering mechanism, namely clustering by fast search and find of density peaks (CFSFDP), known as GC. In the second fold, we design a new relational stacked denoising autoencoder (RSDAE) enabled semi-supervised GAN, termed as RGAN, for ETD. Therefore, the combined solution is named as GC-RGAN. In the system, RSDAE performs as feature extractor and the generator model of proposed RGAN. The proposed semi-supervised solutions efficiently gain the potential benefits of both labeled and unlabeled representations. Furthermore, the proposed solutions are simulated and evaluated over the real-time smart meter dataset of state grid corporation of China using the most suitable performance indicators, e.g., area under the curve and Mathews correlation coefficient. The simulation outcomes validate that the proposed methodology surpasses other traditional methods, such as semi-supervised support vector machine and random forest, for ETD and become acceptable for real-time practices.
Article
Full-text available
Over the last decades, load forecasting is used by power companies to balance energy demand and supply. Among the several load forecasting methods, medium-term load forecasting is necessary for grid’s maintenance planning, settings of electricity prices, and harmonizing energy sharing arrangement. The forecasting of the month ahead electrical loads provides the information required for the interchange of energy among power companies. For accurate load forecasting, this paper proposes a model for medium-term load forecasting that uses hourly electrical load and temperature data to predict month ahead hourly electrical loads. For data preprocessing, modified entropy mutual information-based feature selection is used. It eliminates the redundancy and irrelevancy of features from the data. We employ the conditional restricted Boltzmann machine (CRBM) for the load forecasting. A meta-heuristic optimization algorithm Jaya is used to improve the CRBM’s accuracy rate and convergence. In addition, the consumers’ dynamic consumption behaviors are also investigated using a discrete-time Markov chain and an adaptive k-means is used to group their behaviors into clusters. We evaluated the proposed model using GEFCom2012 US utility dataset. Simulation results confirm that the proposed model achieves better accuracy, fast convergence, and low execution time as compared to other existing models in the literature.
Article
Full-text available
Recently, power systems are facing the challenges of growing power demand, depleting fossil fuel and aggravating environmental pollution (caused by carbon emission from fossil fuel based power generation). The incorporation of alternative low carbon energy generation, i.e., Renewable Energy Sources (RESs), become crucial for energy systems. Effective Demand Side Management (DSM) and RES incorporation enable power systems to maintain demand, supply balance and optimize energy in an environment friendly manner. The wind power is a popular energy source because of its environmental and economical benefits. However, the uncertainty of wind power, makes its incorporation in energy systems really difficult. To mitigate the risk of demand-supply imbalance, an accurate estimation of wind power is essential. Recognizing this challenging task, an efficient deep learning based prediction model is proposed for wind power forecasting. The proposed model has two stages. In the first stage, Wavelet Packet Transform (WPT) is used to decompose the past wind power signals. Other than decomposed signals and lagged wind power, multiple exogenous inputs (such as, calendar variable and Numerical Weather Prediction (NWP)) are also used as input to forecast wind power. In the second stage, a new prediction model Efficient Deep Convolution Neural Network (EDCNN) is employed to forecast wind power. A DSM scheme is formulated based on forecasted wind power, day-ahead demand and price. The proposed forecasting model's performance is evaluated on big data of Maine wind farm ISO NE, USA.
Article
Full-text available
In order to ensure optimal and secure functionality of Micro Grid (MG), energy management system plays vital role in managing multiple electrical load and distributed energy technologies. With the evolution of Smart Grids (SG), energy generation system that includes renewable resources is introduced in MG. This work focuses on coordinated energy management of traditional and renewable resources. Users and MG with storage capacity is taken into account to perform energy management efficiently. First of all, two stage Stackelberg game is formulated. Every player in game theory tries to increase its payoff and also ensures user comfort and system reliability. In the next step, two forecasting techniques are proposed in order to forecast Photo Voltaic Cell (PVC) generation for announcing optimal prices. Furthermore, existence and uniqueness of Nash Equilibrium (NE) of energy management algorithm are also proved. In simulation, results clearly show that proposed game theoretic approach along with storage capacity optimization and forecasting techniques give benefit to both players, i.e., users and MG. The proposed technique Gray wolf optimized Auto Regressive Integrated Moving Average (GARIMA) gives 40% better result and Cuckoo Search Auto Regressive Integrated Moving Average (CARIMA) gives 30% better results as compared to existing techniques.
Article
Full-text available
A Smart Grid (SG) is a modernized grid to provide efficient, reliable and economic energy to the consumers. Energy is the most important resource in the world. An efficient energy distribution is required as smart devices are increasing dramatically. The forecasting of electricity consumption is supposed to be a major constituent to enhance the performance of SG. Various learning algorithms have been proposed to solve the forecasting problem. The sole purpose of this work is to predict the price and load efficiently. The first technique is Enhanced Logistic Regression (ELR) and the second technique is Enhanced Recurrent Extreme Learning Machine (ERELM). ELR is an enhanced form of Logistic Regression (LR), whereas, ERELM optimizes weights and biases using a Grey Wolf Optimizer (GWO). Classification and Regression Tree (CART), Relief-F and Recursive Feature Elimination (RFE) are used for feature selection and extraction. On the basis of selected features, classification is performed using ELR. Cross validation is done for ERELM using Monte Carlo and K-Fold methods. The simulations are performed on two different datasets. The first dataset, i.e., UMass Electric Dataset is multi-variate while the second dataset, i.e., UCI Dataset is uni-variate. The first proposed model performed better with UMass Electric Dataset than UCI Dataset and the accuracy of second model is better with UCI than UMass. The prediction accuracy is analyzed on the basis of four different performance metrics: Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Mean Square Error (MSE) and Root Mean Square Error (RMSE). The proposed techniques are then compared with four benchmark schemes. The comparison is done to verify the adaptivity of the proposed techniques. The simulation results show that the proposed techniques outperformed benchmark schemes. The proposed techniques efficiently increased the prediction accuracy of load and price. However, the computational time is increased in both scenarios. ELR achieved almost 5% better results than Convolutional Neural Network (CNN) and almost 3% than LR. While, ERELM achieved almost 6% better results than ELM and almost 5% than RELM. However, the computational time is almost 20% increased with ELR and 50% with ERELM. Scalability is also addressed for the proposed techniques using half-yearly and yearly datasets. Simulation results show that ELR gives 5% better results while, ERELM gives 6% better results when used for yearly dataset.
Article
Full-text available
This paper focuses on analytics of an extremely large dataset of smart grid electricity price and load, which is difficult to process with conventional computational models. These data are known as energy big data. The analysis of big data divulges the deeper insights that help experts in the improvement of smart grid’s (SG) operations. Processing and extracting of meaningful information from data is a challenging task. Electricity load and price are the most influential factors in the electricity market. For improving reliability, control and management of electricity market operations, an exact estimate of the day ahead load is a substantial requirement. Energy market trade is based on price. Accurate price forecast enables energy market participants to make effective and most profitable bidding strategies. This paper proposes a deep learning-based model for the forecast of price and demand for big data using Deep Long Short-Term Memory (DLSTM). Due to the adaptive and automatic feature learning mechanism of Deep Neural Network (DNN), the processing of big data is easier with LSTM as compared to the purely data-driven methods. The proposed model was evaluated using well-known real electricity markets’ data. In this study, day and week ahead forecasting experiments were conducted for all months. Forecast performance was assessed using Mean Absolute Error (MAE) and Normalized Root Mean Square Error (NRMSE). The proposed Deep LSTM (DLSTM) method was compared to traditional Artificial Neural Network (ANN) time series forecasting methods, i.e., Nonlinear Autoregressive network with Exogenous variables (NARX) and Extreme Learning Machine (ELM). DLSTM outperformed the compared forecasting methods in terms of accuracy. Experimental results prove the efficiency of the proposed method for electricity price and load forecasting.
Article
Full-text available
With increased integration of wind energy systems, an accurate wind speed forecasting technique is a must for the reliable and secure operation of the power network. Statistical methods such as Auto-Regressive Integrated Moving Average (ARIMA) and hybrid methods such as Wavelet Transform (WT) based ARIMA (WT-ARIMA) model have been the popular techniques in recent times for short-term and very short-term forecasting of wind speed. However, the contribution of the forecasting error due to different decomposed time series on the resultant wind speed forecasting error has yet not been analyzed. This paper, thus explores this shortcoming of the ARIMA and WT-ARIMA models in forecasting of wind speed and proposes a new Repeated WT based ARIMA (RWT-ARIMA) model, which has improved accuracy for very short-term wind speed forecasting. A comparison of the proposed RWT-ARIMA model with the benchmark persistence model for very short-term wind speed forecasting, ARIMA model and WT-ARIMA model has been done for various time-scales of forecasting such as 1min, 3min, 5min, 7min and 10min. This comparison proves the superiority of the proposed RWT-ARIMA model over other models in very short-term wind speed forecasting.
Article
Full-text available
Due to the increasing integration of solar power into the electrical grid, forecasting short-term solar irradiance has become key for many applications, e.g. operational planning, power purchases, reserve activation, etc. In this context, as solar generators are geographically dispersed and ground measurements are not always easy to obtain, it is very important to have general models that can predict solar irradiance without the need of local data. In this paper, a model that can perform short-term forecasting of solar irradiance in any general location without the need of ground measurements is proposed. To do so, the model considers satellite-based measurements and weather-based forecasts, and employs a deep neural network structure that is able to generalize across locations; particularly, the network is trained only using a small subset of sites where ground data is available, and the model is able to generalize to a much larger number of locations where ground data does not exist. As a case study, 25 locations in The Netherlands are considered and the proposed model is compared against four local models that are individually trained for each location using ground measurements. Despite the general nature of the model, it is shown show that the proposed model is equal or better than the local models: when comparing the average performance across all the locations and prediction horizons, the proposed model obtains a 31.31% rRMSE (relative root mean square error) while the best local model achieves a 32.01% rRMSE.
Article
Full-text available
This paper presents a hybrid approach to predict the electric energy usage of weather‐sensitive loads. The presented method utilizes the clustering paradigm along with ANN and SVM approaches for accurate short‐term prediction of electric energy usage, using weather data. Since the methodology being invoked in this research is based on CRISP data mining, data preparation has received a great deal of attention in this research. Once data pre‐processing was done, the underlying pattern of electric energy consumption was extracted by the means of machine learning methods to precisely forecast short‐term energy consumption. The proposed approach (CBA‐ANN‐SVM) was applied to real load data and resulting higher accuracy comparing to the existing models.
Article
As solar photovoltaic (PV) generation becomes cost-effective, solar power comes into its own as the alternative energy with the potential to make up a larger share of growing energy needs. Consequently, operations and maintenance cost now have a large impact on the profit of managing power modules, and the energy market participants need to estimate the solar power in short or long terms of future. In this paper, we propose a solar power forecasting technique by utilizing convolutional neural networks (CNNs) and long-short term memory (LSTM) networks recently developed for analyzing time series data in the deep learning communities. Considering that weather information may not be always available for the location where PV modules are installed and sensors are often damaged, we empirically confirm that the proposed method predicts the solar power well with roughly estimated weather data obtained from national weather centers as well as it works robustly without sophisticatedly preprocessed input to remove outliers.
Article
Solar radiation (SR) is an important data for various applications such as climate, energy and engineering. Because of this, determination and estimation of temporal and spatial variability of SR has critical importance in order to make plans and organizations for the present and the future. In this study, a deep learning method is employed for estimating the SR over 30 stations located in Turkey. The astronomical factor, extraterrestrial radiation and climatic variables, sunshine duration, cloud cover, minimum temperature and maximum temperature were used as input attributes and the output was obtained as SR. The datasets of 34 stations, spanning the dates from 2001 to 2007, were used for training and testing the model, respectively, and simulated values were compared with ground-truth values. The overall coefficient of determination, root mean square error and mean absolute error were calculated as 0.980, 0.78 MJm-2day-1 and 0.61 MJm-2day-1, respectively. Consequently, DL model has yielded very precise and comparable results for estimating daily global SR. These results are generally better than or they are comparable to many previous studies reported in literature, so one can conclude that the method can be a good alternative and be successfully applied to similar regions.