Data driven day-ahead electrical load forecasting through repeated wavelet transform assisted SVM model
Introduction
Electrical load requirement is an increasing function and is a reflection of a nation’s development index. Load forecasting is an essential part of the operation, maintenance, and expansion of the modern smart power network. Electrical load forecasting is an important study for the effective utilization of generation resources and their economic dispatch. The forecasted load data decide the unit commitment of generating resources. Besides these, electrical load forecasting finds its importance in electricity market transactions, optimal control of the attached energy storage, operation and maintenance scheduling of generating units, and in the expansion planning of the existing power network [1]. The horizon of electrical load forecasting depends on its end application. This leads to various categories of load forecasting horizons that differ in time-scale as [2]:
- •
Very short-term or ultra short-term (minutes to hours ahead)
- •
Short-term ( day to weeks ahead)
- •
Medium-term (months to year ahead)
- •
Long-term (years ahead)
Modeling is an intrinsic part of any system/phenomena identification, and estimation [3]. Different models for electrical load forecasting are proposed in the literature [1]. Some models are more accurate for specific horizons, while other models can be used for different horizons. The different load forecasting models available in the literature can be grouped under different heads based on specific criteria. For example, if the number of variables is considered while forecasting the electrical load, the models can be categorized under the univariant model and multivariant model [4]. In a univariant model, only the historical load data is necessary for electrical load forecasting. Time-series models come under the group of the univariant model. Load forecasting model that includes more than one variable such as temperature, humidity, and historical electrical load data belongs to the multivariant model. The artificial Neural Network (ANN) model belongs to the class of multivariant model if the inputs to ANN are more than one. Load forecasting models can also be differentiated based on the dependency on the initial conditions. If the initial conditions are varied and the electrical load forecasting model changes, such models are probabilistic models while the models, that are robust against initial conditions are deterministic [4]. The time-series model and regression model are deterministic models, while the ANN model is a probabilistic model. Depending upon the approach of electrical load forecasting model, the models can also be classified as statistical, or machine learning-based models [2]. Time-series and regression models are statistical models, while ANN and Support Vector Machine (SVM) based models are Machine Learning (ML) based models. Some load forecasting models cannot be categorized as they have a complimentary feature from different groups, and such models are hybrid models [2]. These models tend to condescend to others due to superior forecasting results, as they amalgamate advantages of the groups they are derived from.
Initially, time-series model-based approach [5], [6] was used for electrical load forecasting modeling. It uses the standard Box–Jenkin approach for the estimation of the load forecast model. Some of the hybrid models partially utilize time-series approaches [7], [8]. However, lately, these approaches are rarely used, and researchers are drifting towards machine learning-based models and their hybridization with other statistical or machine learning models [4].
In general, the machine learning-based model uses ANN [9], [10], SVM [11], and the associated hybrids. ANN models available in literature differ in the number of inputs to ANN and the associated training approaches. In [12], Particle Swarm Optimization (PSO) based training is used to develop an ANN-based model, where the input to ANN comprises information related to historical load data, temperature, humidity, holidays, and week-offs. BOOsted Neural Network (BooNN) based approach is used in [13]. BooNN has a set of ANN that are trained iteratively, where at each iteration, the error between the estimated output from previously trained ANN and the actual output is minimized. Feedforward neural network-based approach is used in [14] and a comparison with echo state network is made. The inputs to the ANN are the historical electrical load information with timestamps, temperature, humidity, and solar irradiance. The electrical load data is clustered using the K-means approach and clusters are based on load values and information related to weekdays. A generalized regression neural network for estimation of electrical load forecast is proposed in [15], where fruit fly optimization algorithm is used for training of the neural network. An SVM-based approach is presented in [16], where sub-sampling is used to reduce the computational complexity of SVM. Another variant of SVM modeling with ambient temperature and electrical load data as input is proposed in [17]. SVM with parameters tuned by PSO is proposed in [18]. Thus, these machine learning-based models differ in terms of the input parameters and training approaches, besides different machine models used.
One of the approaches of hybrid model estimation of electrical load forecast is to decompose the historical load data and use any of the statistical/ML-based models. Generally, decomposition of load data time-series helps better estimate the electrical load forecast model compared to the non-decomposed time-series load data. Apart from better estimation, decomposition of load data also help in denoising the historical load data. Decomposition of electrical load data is carried out using any available Multi-Resolution Analysis (MRA) techniques. The decomposed sub-series is then modeled through the statistical/ML method. The estimated sub-series model is used to forecast the electrical load for the corresponding sub-series. These forecasted sub-series are reconstructed using the inverse of the MRA technique. A hybrid model using Empirical Mode Decomposition (EMD) followed by SVM is proposed in [19], [20]. EMD is used to decompose and denoise the input data, and the decomposed sub-series is modeled through SVM. A similar approach is also taken in [21], however, the decomposed sub-series obtained after EMD is modeled through Least Square Support Vector Machine (LS-SVM). Variational Mode Decomposition (VMD), which is superior in denoising the input data as compared to EMD, is used in [22] followed by sub-series model estimation by Long Short Term Memory (LSTM) neural network. A combination of EMD and Deep Belief Network (DBM) is used in [23] for time-series modeling of electrical load data.
WT is another MRA technique used in decomposing the input data of any type and is widely used in various fields [24]. In [25], the input electrical load data is decomposed using Wavelet Transform (WT), and three different models are used to estimate the decomposed load data. A radial basis function-based neural network, time-series model, fuzzy inference neural network based models are used, and a comparison among them is shown in [25]. A second-order gray neural network-based model is used in [26] for the estimation of decomposed load data using WT. The decomposed load data time series are also estimated using SVM in [27], [28]. In [29], a hybrid model is proposed where WT is used to decompose the input data, and the sub-series is modeled using Triple Exponential Smoothing (TES) and Weighted Nearest Neighbor (WNN) models. The proposed hybrid model is superior to its building block models, which are TES and WNN. In [30], ANN is used to model the sub-series after input series decomposition through WT. However, the proposed model is not compared with any existing model. A similar study is also carried out in [31], where the parameter of ANN is optimized using Bat Algorithm (BA). Deep Neural Network (DNN) along with WT is used in [32] and its superiority over SVM is verified.
A general approach observed in all these decomposition-based hybrid models is the decomposition of input load data time-series using any of the available MRA techniques followed by estimating the decomposed sub-series using either statistical/ML models. Fig. 1 shows the generalized approach for these types of hybrid electrical load forecasting models. The first stage is the decomposition stage which can be carried out by MRA techniques such as WT/EMD/VMD. The second stage is the model estimation and forecasting of decomposed sub-series. The third stage is the reconstruction of the forecasted sub-series through the inverse of the decomposition. To the best of the author’s knowledge, none of the decomposition-based hybrid electrical load forecasting models have explored the contribution of forecasting error due to individual sub-series towards total forecasting error. In this work, WT-SVM hybrid model [27] is used as a reference model, and the effect of forecasting error due to individual sub-series is estimated. The sub-series with the highest forecasting error contribution is selected and a superior hybrid forecasting model is proposed.
The major contributions, in this paper, towards day-ahead electrical load forecasting are:
- 1.
A varying error contribution of decomposed sub-series towards total error in forecasted value is estimated and used.
- 2.
Utilizing the varying error contribution, a repeated wavelet-based SVM (RWT-SVM) model is proposed that gives a better day-ahead electrical load forecasting model as compared to wavelet transform-based SVM (WT-SVM) model [27].
The organization of the paper is as follows. A brief discussion of the basics of WT, SVM, and WT-SVM is done in Section 2. The proposed day-ahead electrical load forecasting model is given in Section 3. The proposed model is tested for real electrical load data for three different places, and the result with discussion is shown in Section 4. The conclusion of the paper is drawn in Section 5.
Section snippets
Wavelet transform [33]
Electrical load varies over a day, and thus its behavior is non-stationary over a time duration. This variation in electrical load can be effectively captured using wavelet transform. The wavelet transform of a signal gives information related to the time of occurrence of the various frequency components. The transform utilizes a set of basis function derived from a short duration wave that grows and decays within a limited time. Discrete WT (DWT) is preferred over continuous WT due to its
Proposed methodology
The electrical load is a time-varying signal and there appears a repetitive trend though there is no defined periodicity. This repetitive trend appearing in the electrical load can be attributed to the customer’s electrical power consumption behavior. WT can effectively capture the information contained in the time-varying signal. The output of WT of the electrical load signal is a set of sub-series that consists of information related to the time of occurrence of different frequencies in the
Electrical load data set and window size selection
The electrical load data set used to test the proposed algorithm is taken from [37]. The data set available are of various European countries from 2006–2015. Monthly electrical load data of different years for three different countries are taken for study. The data is having a time resolution of 1 h. The year is taken arbitrarily. The data set taken for the study is of Great Britain (2012), Germany (2013), and France (2014).
The first part of the work is to find a suitable window size. Window
Conclusions
Electrical load forecasting is the prerequisite to power system control, operation, planning, and extension. The trend in the electrical load data can be effectively captured using any MRA techniques. The decomposition of load data through WT leads to lower noise intrusion into the data. The modeling of the decomposed sub-series through SVM gives superior forecasting as compared to ordinary SVM. The individual sub-series forecasting through WT-SVM is estimated, and it is observed that different
CRediT authorship contribution statement
Aasim: Conceptualization, Methodology, Software, Writing – original draft. S.N. Singh: Conceptualization, Writing – review & editing. Abheejeet Mohapatra: Writing – review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (39)
- et al.
Electrical load forecasting models: A critical systematic review
Sustainable Cities Soc.
(2017) - et al.
Probabilistic electric load forecasting: A tutorial review
Int. J. Forecast.
(2016) - et al.
A hybrid dynamic and fuzzy time series model for mid-term power load forecasting
Int. J. Electr. Power Energy Syst.
(2015) - et al.
Analysis of an adaptive time-series autoregressive moving-average (ARMA) model for short-term load forecasting
Electr. Power Syst. Res.
(1995) - et al.
An intelligent hybrid short-term load forecasting model for smart power grids
Sustainable Cities Soc.
(2017) - et al.
Boosted neural networks for improved short-term electric load forecasting
Electr. Power Syst. Res.
(2017) - et al.
A short-term power load forecasting model based on the generalized regression neural network with decreasing step fruit fly optimization algorithm
Neurocomputing
(2017) - et al.
Sequential grid approach based support vector regression for short-term electric load forecasting
Appl. Energy
(2019) - et al.
Short-term electrical load forecasting using the Support Vector Regression (SVR) model to calculate the demand response baseline for office buildings
Appl. Energy
(2017) - et al.
An incremental electric load forecasting model based on support vector regression
Energy
(2016)
Empirical mode decomposition based denoising method with support vector regression for time series prediction: A case study for electricity load forecasting
Measurement
A hybrid application algorithm based on the support vector machine and artificial intelligence: An example of electric load forecasting
Appl. Math. Model.
A hybrid short-term load forecasting model based on variational mode decomposition and long short-term memory networks considering relevant factors with Bayesian optimization algorithm
Appl. Energy
Empirical mode decomposition based ensemble deep learning for load demand time series forecasting
Appl. Soft Comput.
Short term load forecasting using a hybrid intelligent method
Knowl.-Based Syst.
A novel decomposition-ensemble model for forecasting short-term load-time series with multiple seasonal patterns
Appl. Soft Comput.
Short term load forecasting using wavelet transform combined with Holt–Winters and weighted nearest neighbor models
Int. J. Electr. Power Energy Syst.
An ensemble method of full wavelet packet transform and neural network for short term electrical load forecasting
Electr. Power Syst. Res.
Optimization of neural network with wavelet transform and improved data selection using bat algorithm for short-term load forecasting
Neurocomputing
Cited by (39)
An instance based multi-source transfer learning strategy for building's short-term electricity loads prediction under sparse data scenarios
2024, Journal of Building EngineeringTFEformer: A new temporal frequency ensemble transformer for day-ahead photovoltaic power prediction
2024, Journal of Cleaner ProductionPower load combination forecasting system based on longitudinal data selection
2024, Engineering Applications of Artificial IntelligenceUse of weighted local constant method to short-term forecasting of electric load in cities at weekends
2023, Electric Power Systems ResearchSliding window and dual-channel CNN (SWDC-CNN): A novel method for synchronous prediction of coal and electricity consumption in cement calcination process
2022, Applied Soft ComputingCitation Excerpt :Aasim [16] combines a hybrid model of wavelet transformation (WT) and SVM features to estimate the regression model of power load prediction using historical time series information of the power load. Although SVM can deal with nonlinear regression problems, when the amount of training data is large, the performance of SVM is limited and cannot play a good role [17]. XGBoost algorithm based on tree model has many successful applications in time series prediction.
Forecasting power load: A hybrid forecasting method with intelligent data processing and optimized artificial intelligence
2022, Technological Forecasting and Social ChangeCitation Excerpt :In particular, SVM, capable of solving nonlinear problems and locating the global optimal solution, is widely used in power load forecasting (Ahmad and Chen, 2019). For example, Singh and Mohapatra (2021) used a hybrid forecasting model—based on wavelet transform (WT) and SVM—for power load forecasting which effectively improved the accuracy of load forecasting. Barman and Choudhury (2020) used the gray wolf optimizer (GWO) algorithm to optimize the parameters of SVM, which effectively improved the load forecasting accuracy, and the comparison results indicate that the prediction effects of SVM before and after optimization are better than that of ANN.