Content uploaded by Nadeem Javaid
Author content
All content in this area was uploaded by Nadeem Javaid on Jul 13, 2019
Content may be subject to copyright.
Data Analytics for Load and Price
Forecasting via Enhanced Support Vector
Regression
Tanzeela Sultana1, Zahoor Ali Khan2, Nadeem Javaid1(B
), Syeda Aimal1,
Aisha Fatima1, and Shaista Shabbir3
1COMSATS University Islamabad, Islamabad 44000, Pakistan
nadeemjavaidqau@gmail.com
2Computer Information Science, Higher Colleges of Technology, Fujairah 4114, UAE
3Virtual University of Pakistan, Kotli Campus, Azad Kashmir 1100, Pakistan
http://www.njavaid.com
Abstract. In this paper, month-ahead electricity load and price fore-
casting is done to achieve accuracy. The data of electricity load is taken
from the Smart Meter (SM) in London. Electricity load data of five
months is taken from one block SM along with the weather data. Data
Analytics (DA) techniques are used in the paper for month-ahead elec-
tricity load and price prediction. In this paper, forecasting is done in
multiple stages. At first stage, feature extraction and selection is per-
formed to make data suitable for efficient forecasting and to reduce com-
plexity of data. After that, regression techniques are used for prediction.
Singular Value Decomposition (SVD) is used for feature extraction after-
wards; feature selection is done in two-stages, by using Random Forest
(RF) and Recursive Feature Elimination (RFE). For electricity load and
price forecasting Logistic Regression (LR), Support Vector Regression
(SVR) is used. Moreover forecasting is done by the proposed technique
Enhanced Support Vector Regression (EnSVR), which is modified from
SVR. Simulation results show that the proposed system gives more accu-
racy in load and price prediction.
Keywords: Load forecasting ·Price forecasting ·Data Analytics ·
Logistic Regression ·Support Vector Regression ·
Enhanced Support Vector Regression
1 Introduction
With the increasing demand of electricity along with the growth of advanced
technologies, the energy markets are evolving consistently. The Traditional Grid
(TG) systems are failed to incorporate with new technology. TG systems are
centralized and often shows uncertainty. Only one-way transmission strategy is
used in TGs. There were several issues that were rising in TGs. Due to incre-
ment in power demand, unstable voltage became major issue. Quality of power
c
Springer Nature Switzerland AG 2019
L. Barolli et al. (Eds.): EIDWT 2019, LNDECT 29, pp. 259–270, 2019.
https://doi.org/10.1007/978-3-030-12839-5_24
260 T. Sultana et al.
also decreased due to voltage fluctuations. Resources were unpredictable, which
causes low feasibility. To deal with the energy issues with the TGs, the grid sys-
tems are attracting toward the smart grid systems. The traditional approaches
for power system are decreasing and the concept of Smart Meters (SMs) and
Smart Grids (SGs) is getting high. SG systems are becoming more popular
worldwide due to rise of technology. SG, can be defined as, an advanced develop-
ment that gathers the Information Communication and Technology (ICT) and
traditional energy systems to improve productivity and ability of the grid by
making a relation between consumer and utility systems. SG provides the dis-
tributed forms of energy generation to minimize uncertainty. SG also provides
efficient energy management, such as communication between utility store and
consumer, security, generation cost saving, processing and storing energy data
[1–4]. SG presents many new components like: SMs, Demand Response (DR)
management scheme, growth in IoT devices, online interactions and billing sys-
tems, and integration of renewable energy. SG works on Information Technology
(IT) systems, for performing several tasks such as processing and measurements,
experiments and simulations, management [5–7] (Table 1).
As the power system is moving towards SG, more electrical data is generated
from SMs. For efficient management of energy in utility markets, Data Analytics
(DA) is used. DA is the process of collecting data and performs several operations
on data such as, cleaning, transforming and modeling data, to provide solutions.
Different DA techniques are used to perform multiple operations on data. DA is
used almost in every field, for several purposes. However, in energy markets DA is
mostly used for forecasting purposes. Forecasting is mostly done to estimate the
electrical load and price of electricity in future and for efficient management of
electricity. Also to take account of the consumption rate by user and generation
of electricity. Several forecasting techniques are introduced till now, to provide
accurate forecasting. The best forecasting algorithm gives accurate forecasting
results that can minimizes the issues related to power [8] and it also helps to
improve the electricity management system. As electric data generated from the
smart meters is time series data, there are time series forecasting paradigms
for different intervals: short term, medium term and long term [9]. Electricity
grids are often affected by weather conditions. The PCR based method is used
for prediction of temperature on power grids in [10], the conductor temperature
is collected by the installation of a sensor over transmission lines of TaiPower
grid in Taiwan. The main objective of the paper is to make improvement in
transmission lines that are influenced by weather conditions. Many techniques
are used in literature for forecasting both load and price in different horizon.
However, in this paper, electricity load and price forecasting is done for month-
ahead in three stages by regression techniques LR, SVR and EnSVR by taking
the dataset of one block of smart meters in London. Load and weather data of
that block is also used for month-ahead load and price forecasting.
Load and Price Forecasting 261
Table 1. List of abbreviation
ABC Artificial Bee Colony
AGG Arian Golden Group
ARIMA Autoregressive Integrated Moving Average
CMI Conditional Mutual Information
DA Data Analytics
DR Demand Response
DTR Dynamic Thermal Rating
DTW Dynamic Time Warping
ELM Extreme Learning Machine
EMD Empirical Mode Decomposition
EnSVR Enhanced Support Vector Regression
ERCOT Electric Reliability Council of Texas
FWPT Flexible Wavelet Packet Transform
GS Greater Sydney
GTA Grid Traverse Algorithm
ICT Information Communication and Technology
LR Logistic Regression
MA Moving Average
MIMO Multi-Input Multi-Output
MLE Maximum Likelihood Estimation
MLR Multiple Linear Regression
NA North American
NLSSVM Non-linear Least Square Support Vector Machine
NN Neural Networks
NSW New South Wales
PCR Principle Component Regression
PLD Power Load Decomposition
PJM Pennsylvania-New Jersey-Maryland
PSO Particle Swarm Optimization
QLD QueensLand
RF Random Forest
RFE Recursive Feature Elimination
RVFL Random Vector Functional Link Network
SG Smart Grid
SM Smart Meter
SMA Simple Moving Average
SVC Support Vector classification
SVD Singular Value Decomposition
SVM Support Vector Machine
SVR Support Vector Regression
TG Traditional Grid
262 T. Sultana et al.
1.1 Motivation
A lot of work is done in literature for efficient management of electricity. Many
techniques are proposed to forecast electricity load and price in smart meters
and smart grids for short-term, medium-term and long-term forecasting. Many
papers proposed the hybrid model for prediction such as [2]proposesahybrid
of Grid Traverse Algorithm (GTA) and Particle Swarm Optimization (PSO) for
short-term load forecasting, [4] does a three-stage prediction for both load and
price through different methods at each stage, [6] mixes Non-Linear Regression
with SVM for month-ahead price forecast. [2,5–7] uses SVR for forecasting. By
taking all the studies into consideration, the proposed paper aims to forecast
both load and price for month-ahead to achieve accuracy. The main objective of
papers is to achieve more accuracy and performance in forecasting and to resolve
peak load and price variation in smart grids.
1.2 Problem Statement
In paper [5] short-term load forecast is predicted to calculate Demand Response
(DR) for an office building in eastern China. The model’s objective is to achieve
maximum forecasting accuracy under DR baseline using historical electric load
and weather data. The paper uses SVR for accurate predictions. [7] predicts very
short term half-hour ahead load forecasting, using on-line SVR where prediction
is done to overcome the issues of memory and storage using advanced on-line
methods. The paper used the dataset of Irish Commission for Energy Regulation
(CER) utility company and achieves accuracy and minimizes the storage issues
by using ensemble tree-based ensemble methods with advanced on-line methods.
However, model outperforms for storage and memory issues, besides this model
gives single output with only one-hour ahead prediction, a further methodology
is required for the multiple outputs. Paper [7] also lacks in performance to predict
separate day. Above papers forecast either load or price, these papers can only
predict for few hours. However, the proposed model forecast both load and price
by two existing schemes LR, SVR. Afterwards forecasting is done the scheme
proposed by the paper.
1.3 Contribution
In this paper load and price forecast is performed in a new way, by making a
combination of schemes that are never used before for forecasting. The major
purpose is to achieve accuracy in terms of electricity load and price prediction.
The main contributions of this paper are:
•Only load data of one block smart meter, along with weather data is given in
the dataset, the price is calculated from the given load data to predict price
along with load,
•Data preprocessing is done by both feature extraction and feature selection.
For extraction, SVD is used and for feature selection, first RF and after that
RFE is used,
Load and Price Forecasting 263
•Month-ahead forecasting for load and price is accomplished by executing mul-
tiple regressors such as LR, SVR. The EnSVR method is proposed to achieve
more accuracy and for better electricity forecasting.
2 Related Work
With the generation of electricity in smart grids, the problem of electrical load
and price is increasing. To tackle the electric load and price forecast issues,
different schemes are proposed in the literature. Paper [1] founds the limita-
tions in short-term forecasting for university buildings, because of diverse pat-
terns. Paper uses a 2-stage forecast model for the short-term electric load pre-
diction in the university campus. Five years data is taken from a university
campus in Korea and 2-stage prediction is done. At the first stage, by using
Moving Average (MA) method and secondly by using RF technique. Further-
more, the performance is evaluated by the time-series cross-validation process.
Short-term load forecasting is also done by paper [2] over distributed systems
using SVR and two-step hybrid optimization algorithm based on GTA and PSO.
The data is calculated from the Partner utility’s distribution feeder and the
prediction is done on seasoned base. This paper overcome the issue for total
load of small section of distributed feeder. However, this paper works effec-
tively by using external factors. Load prediction is also done by taking external
factors like temperature and weather data into consideration. Paper [3] deter-
mines the growth in household load forecast. The level of load in households
is increasing due to which error rate is also increasing. The paper proposed a
Table 2. Summary of related work
Technique used Features Dataset Region Limitations
RF, ANN [1]Academic year, load University
campus
Seoul, Korea -
SVR, PSO [2]Historical load Partner’s
utility
US Better perform only
with external factors
MLR [3]Calender effect, load,
weather data
Ausgrid NSW and GS Minimize accuracy
due to large dataset
3-Stage
methodology [4]
Price, demand NYISO and
PJM
US -
SVR [5]Temperature,
electricity load,
weather
Office building Eastern China Predict few hours of
one building only
Hybrid Non-Linear
Regression, SVR [6]
Price, reading rate,
average wages
ERCOT Tex a s, U S A Model works well only
for month-ahead
forecast
On-line SVR [7]Half-hour load, time,
weekdays, weather
data
Irish CER Ireland Model is not
applicable for long
term forecasting
EMD-ELM [8]Half-hour electric load NSW,
Victoria, QLD
Australia -
DWT, EM D, RVFL
[9]
Year, max, median,
mean, standard
deviation
AEMO Australia -
264 T. Sultana et al.
forecasting model for scheduling and optimum planning by using a historical
load with weather data and calendar effects of Australian grids of residential areas
of New South Wales (NSW) and Greater Sydney (GS). Calendar effects are used
for prediction in several periods of the calendar: days, months and years. Many
forecasting techniques such as Multiple Linear Regression (MLR), Neural Net-
works (NN), regression trees and Support Vector Machine (SVM) are used. Accu-
racy of the paper is minimized due to large data. Paper [4] captures the limita-
tion of forecasting algorithms, as they perform forecasting of load and price sepa-
rately. So paper proposed hybrid algorithms based on demand-side management
models to predict both price and load concurrently using data of New York Inde-
pendent System Operator (NYISO), NSW and Pennsylvania New-Jersey Mary-
land (PJM). Forecasting in this paper is performed in three stages; at first stage,
Flexible Wavelet Packet Transform (FWPT) and Conditional Mutual Informa-
tion (CMI) algorithms are used for selection, then Multiple-Input Multiple-Output
(MIMO) based Non-linear Least Square Support Vector Machine (NLSSVM) and
Autoregressive Integrated Moving Average (ARIMA) is applied to define a rela-
tionship between load and price, at third stage enhances ABC on the basis of Time-
Varying Coefficients and Stumble Artificial Bee Colony (TV-SABC) for parame-
ter optimization. Different algorithms are used, that are complex to be used by
the utility market. In paper [5] SVR model is employed reduce complexity and
to minimize the risk of a load during the peak hours. Data is taken from four
office buildings in China to examine DR. The model only predicts only few hours
for one building efficiently. Electricity price is forecast in paper [6] to examine
and overcome the abnormalities in the electric power industry. This paper also
founds inaccuracy in month-ahead forecasting. Month-ahead average price fore-
cast is done by hybrid Non-Linear Regression and SVM on data of Electric Reli-
ability Council of Texas (ERCOT) electricity market in Texas, United States of
America (USA). The limitation of this system is that it gives accurate forecasting
for month-ahead only. With the enhancement of electricity grid market, most of
the data is processed online. Paper [7] figures out the memory and storage prob-
lems in systems. Paper proposed a short-term load forecasting with online SVR to
overcome the biasness of the Irish CER grid market. Ten models are compared in
this paper, however, online SVR achieved more accuracy. The model is not applica-
ble for long-term forecasting. For short-term load forecasting, paper [8] employed
a novel method Empirical Mode Decomposition (EMD)-mixed-Extreme Learning
Machine (ELM), where EMD is used for the decomposition of complex load fea-
tures to smoothen the data. RBF and UKF kernels are used in EML method to han-
dle the limitations of using one kernel. All the features of data are not captured by
one kernel. In the paper load data of NSW, Victoria and Queensland (QLD) is used
for forecasting. Another hybrid technique is used for short-term load prediction,
which is comprised of EMD and Random Vector Functional Link Network (RVFL)
in paper [9]. For better forecasting, ensemble method RVFL based on Dynamic
Time Warping (DTW) and EMD is presented for accuracy and performance.
Load data is collected from AEMO along with NSW, Tasmania (TAS), QLD,
South Australia (SA) and Victoria. This paper targeted the insufficiency of the
incremental methods, is the unnoticed change in the target value (Table 2).
Load and Price Forecasting 265
3SystemModel
In this section, the proposed system model of the paper is presented in detail.
The flow of the system model is shown in Fig. 1.
Fig. 1. Framework of proposed system
3.1 Dataset Description
The dataset of smart meters in London is used. In the whole dataset, data of
household electric load of multiple blocks is given. Also, load data of half-hourly
and hourly basis is presented along with the daily temperature data and hourly
weather data of three years, from 2013 to 2015.
In this paper, for accurate forecasting of load and price, electrical load data
of only one block smart meter is considered along with weather data of that
block is used in this paper. Load and weather data of five months is used to
forecast one-month ahead load and price of that block. The main focus of the
paper is to forecast price and load. As the dataset does not have any price data
of smart meter, the paper calculated price from the given load data to predict
the future price for the smart meter of that block.
3.2 Feature Importance of Data
Feature importance of data is calculated for every feature in the dataset. It is
examined by observing the effect of every independent feature in the data for
the target feature. Figure 2, shows the feature importance for a load. For load
forecasting, price feature shows more influence on the target than other features.
However, the feature that does not affect the target is humidity.
Like load, feature importance for Price is also calculated. The Fig. 3,shows
the feature importance of every feature to the target. For price, pressure shows
more influence than other and humidity does not shows any effect also for price
forecasting.
266 T. Sultana et al.
Fig. 2. Feature importance for load
3.3 Proposed System Model
Figure 1demonstrates the flow of presented system model. In the proposed
model, first, the load data of one block smart meter is taken along with the
weather data of the block. Then the price is calculated for price forecasting and
data is preprocessed. Feature extraction is done by SVD. After that feature selec-
tion is executed by two techniques, RF and RFE for getting the best features
for accurate forecasting. In the proposed model, forecasting is done by LR, SVR
and the SVR is modified to EnSVR.
Fig. 3. Feature importance for price
4 Simulation Results and Reasoning
In this section, results of electricity load and price forecast are presented with
reasoning. The proposed schemes are implemented by using five months data of
Load and Price Forecasting 267
load and price. The data is divided into training and testing as, 70% training
data and 30% test data. Then feature and selection is done to extract mean-
ingful features from the data. Forecasting is performed by two existing and one
proposed scheme. Simulations are performed using a server of 2.50 GHz Intel
Core i5 CPU and 4 GB RAM. LR, SVR and EnSVR are implemented in Python
using spyder platform.
4.1 Load and Price Forecasting
In this paper, month-ahead load and price is predicted by taking five month
historical data of smart meter in one block. Load and price forecast are shown
in Figs. 4and 5.
Electricity load and price forecasting is done by using multiple techniques.
Feature extraction is done by SVD. Feature selection is performed by using two
techniques, RF and RFE. Regression techniques are used for load and price
forecasting. Existing techniques LR, SVR and proposed scheme EnSVR is used
for prediction. At first step, LR is used and at second stage, SVR is used. At the
final stage, EnSVR is used to achieve prediction accuracy.
Fig. 4. Load forecast
Logistic Regression (LR). Logistic Regression (LR) is basically a statisti-
cal model and used as a classification technique in machine learning. LR shows
nonlinear relationship of one dependent variable with multiple independent vari-
ables. Logistic or sigmoid function is used to show relationship by estimating
probabilities and to transform the output. Logistic function can be defined as:
g(z)=1/1+e−z(1)
Support Vector Regression (SVR). Support Vector Regression (SVR) is
the regression technique, based on Support Vector Machine (SVM) model. SVR
learning is based on ratio of training and testing data. Input parameters of SVR
are defined by all possible values.
268 T. Sultana et al.
Fig. 5. Price forecast
Enhanced Support Vector Regression (EnSVR). To improve the per-
formance of SVR, this paper proposed an Enhanced SVR technique named as
EnSVR. EnSVR is modified from original SVR technique. To achieve more fore-
casting accuracy, best parameters are enhanced. Only those parameters are con-
sidered that give better result. SVR is enhanced by making changes in parame-
ters such as cost, epsilon, gamma and tolerance. Best parameters are picked by
cost = 1.0, epsilon = 0.001, gamma = 10, tolerance = 0.1. Forecasting results
shows that EnSVR shows more accuracy than SVR for load and price as illus-
trated in accuracy table. Forecasting results are also shown in Figs. 4and 5.
4.2 Accuracy Measurement
The goal of this paper is to achieve more accuracy for load and price forecasting
on the data of smart meters. The accuracy is calculated for both load and price
by using LR, SVR and EnSVR. Where LR has achieved 89% SVR has 60%
accuracy and EnSVR got 94% accuracy. Accuracy is also examined for price
on the base of every technique used. LR has achieved 74% and SVR has 60%
accuracy for price. Whereas EnSVR achieves 91% accuracy that is more than
other techniques. Proposed technique EnSVR achieves more accuracy more than
other techniques used.
4.3 Performance Evaluation
To calculate the performance of forecasting techniques, performance metrics such
as Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE),
Mean Squared Error (MSE) and Root Mean Square Error (RMSE) are used.
These evaluators are mostly used to check the accuracy for forecasting. Table3
shows the performance percentage for LR, SVR and EnSVR for load. Table 4
shows the error values for price.
Load and Price Forecasting 269
Table 3. Performance evaluation for load
Method RMSE% MSE% MAE% MAPE%
LR 0.0024 0.0006 0.0021 inf
SVR 0.0005 2.704 0.0004 22.142
EnSVR 0.0001 1.065 8.3451 5.706
Table 4. Performance evaluation for price
Method RMSE% MSE% MAE% MAPE%
LR 0.0304 0.0927 0.0168 114.12
SVR 0.0071 0.0051 0.006 1893.3
EnSVR 0.002 0.0004 0.001 7.164
5 Conclusion
With the rapid growth of electricity it is becoming very difficult for utility market
to handle the issues relating to forecast for future load and price. In this paper,
short term load and price forecasting is done to achieve accuracy by taking
load data of smart meters in London with weather data of about five months.
There are many strategies proposed for forecasting. However in this paper, short-
term load and price forecasting is done by using multiple techniques and with
the proposed technique. For forecasting first LR and SVR is used. After that
the forecasting is done by the proposed scheme EnSVR. Results show that the
proposed EnSVR technique outperforms the existing techniques in both load
and price forecasting.
References
1. Zhang, D., Li, S., Sun, M., O’Neill, Z.: An optimal and learning-based demand
response and home energy management system. IEEE Trans. Smart Grid 7(4),
1790–1801 (2016)
2. Shafie-khah, M., Siano, P.: A stochastic home energy management system con-
sidering satisfaction cost and response fatigue. IEEE Trans. Industr. Inf. 14(2),
629–638 (2017)
3. Akhavan-Rezai, E., Shaaban, M.F., El-Saadany, E.F., Karray, F.: Online intelligent
demand management of plug-in electric vehicles in future smart parking lots. IEEE
Syst. J. 10(2), 483–494 (2016)
4. Mirjalili, S., Mirjalili, S.M., Lewis, A.: Grey wolf optimizer. Adv. Eng. Softw. 69,
46–61 (2014)
5. Jayabarathi, T., Raghunathan, T., Adarsh, B.R., Suganthan, P.N.: Economic dis-
patch using hybrid grey wolf optimizer. Energy 111, 630–641 (2016)
6. Geem, Z.W., Yoon, Y.: Harmony search optimization of renewable energy charging
with energy storage system. Int. J. Electr. Power Energy Syst. 86, 120–126 (2017)
270 T. Sultana et al.
7. Ouyang, H.B., Gao, L.Q., Kong, X.Y., Li, S., Zou, D.X.: Hybrid harmony search
particle swarm optimization with global dimension selection. Inf. Sci. 346, 318–337
(2016)
8. Ambia, M.N., Hasanien, H.M., Al-Durra, A., Muyeen, S.M.: Harmony search
algorithm-based controller parameters optimization for a distributed-generation
system. IEEE Trans. Power Delivery 30(1), 246–255 (2015)
9. Manzoor, A., Javaid, N., Ullah, I., Abdul, W., Almogren, A., Alamri, A.: An intel-
ligent hybrid heuristic scheme for smart metering based demand side management
in smart homes. Energies 10(9), 1258 (2017)
10. Javaid, N., Javaid, S., Abdul, W., Ahmed, I., Almogren, A., Alamri, A., Niaz,
I.A.: A hybrid genetic wind driven heuristic optimization algorithm for demand
side management in smart grid. Energies 10(3), 319 (2017)
11. Mahmood, D., Javaid, N., Ahmed, I., Alrajeh, N., Niaz, I.A., Khan, Z.A.: Multi-
agent-based sharing power economy for a smart community. Int. J. Energy Res.
41(14), 2074–2090 (2017)
12. Zhao, Z., Lee, W.C., Shin, Y., Song, K.B.: An optimal power scheduling method for
demand response in home energy management system. IEEE Trans. Smart Grid
4(3), 1391–1400 (2013)
13. Logenthiran, T., Srinivasan, D., Shun, T.Z.: Demand side management in smart
grid using heuristic optimization. IEEE Trans. Smart Grid 3(3), 1244–1252 (2012)
14. Rajalingam, S., Malathi, V.: HEM algorithm based smart controller for home power
management system. Energy Buildings 131, 184–192 (2016)
15. Ahmed, M.S., Mohamed, A., Khatib, T., Shareef, H., Homod, R.Z., Ali, J.A.: Real
time optimal schedule controller for home energy management system using new
binary backtracking search algorithm. Energy Buildings 138, 215–227 (2017)
16. Zhang, Y., Li, C., Li, L.: Wavelet transform and Kernel-based extreme learning
machine for electricity price forecasting. Energy Syst. 9(1), 113–134 (2018)
17. F¨orderer, K., Ahrens, M., Bao, K., Mauser, I., Schmeck, H.: Towards the modeling
of flexibility using artificial neural networks in energy management and smart grids:
note. In: Proceedings of the Ninth International Conference on Future Energy
Systems, pp. 85–90. ACM (2018)
18. Gao, W., Darvishan, A., Toghani, M., Mohammadi, M., Abedinia, O., Ghadimi, N.:
Different states of multi-block based forecast engine for price and load prediction.
Int. J. Electr. Power Energy Syst. 104, 423–435 (2019)
19. Nowotarski, J., Weron, R.: Recent advances in electricity price forecasting: a review
of probabilistic forecasting. Renew. Sustain. Energy Rev. (2017)
20. Bramer, L.M., Rounds, J., Burleyson, C.D., Fortin, D., Hathaway, J., Rice, J.,
Kraucunas, I.: Evaluating penalized logistic regression models to predict Heat-
Related Electric grid stress days. Appl. Energy 205, 1408–1418 (2017)
