A novel machine learning approach for estimation of electricity demand: An empirical evidence from Thailand

Highlights

• A hybrid approach is presented for the estimation of the electricity demand.

• The proposed method integrates the capabilities of GP and SA.

• The GSA model makes accurate predictions of the electricity demand.

Abstract

This study proposes an innovative hybrid approach for the estimation of the long-term electricity demand. A new prediction equation was developed for the electricity demand using an integrated search method of genetic programming and simulated annealing, called GSA. The annual electricity demand was formulated in terms of population, gross domestic product (GDP), stock index, and total revenue from exporting industrial products of the same year. A comprehensive database containing total electricity demand in Thailand from 1986 to 2009 was used to develop the model. The generalization of the model was verified using a separate testing data. A sensitivity analysis was conducted to investigate the contribution of the parameters affecting the electricity demand. The GSA model provides accurate predictions of the electricity demand. Furthermore, the proposed model outperforms a regression and artificial neural network-based models.

Introduction

The great influence of electric energy on economic development is well-understood. In general, electricity demand is defined as the amount of electricity generated and distributed in a specific region over a specific period. Accurate estimation of the electricity demand is a crucial factor for energy management. In this context, many technological developments and progresses have been achieved. Up to now, different methods are employed for the prediction of the electricity demand [1]. Generally, these methods can be categorized into three major groups [2]: (1) autoregressive integrated moving average (ARIMA), (2) multiple linear regression (MLR) and (3) machine learning techniques. ARIMA models are stochastic difference equations commonly used for the modeling of stochastic disturbances in time series analysis [3]. Cho et al. [4] conducted a comparative study between ARIMA and traditional regression approach for the customer short term load forecasting. Abdel-Aal and Al-Garni [5] used this technique to predict the domestic electric energy consumption in Saudi Arabia. Ediger and Akar [6] applied ARIMA to the estimation of the future energy demand in Turkey. Suhartono and Endharta [7] used double seasonal ARIMA model for short term electricity load demand forecasting in Indonesia. Mohamed et al. [8] investigated the double seasonal ARIMA model for forecasting the double seasonal (daily and weekly) Malaysia load demand time series. Taylor [9], [10] and Taylor et al. [11] employed various techniques such as double seasonal exponential smoothing; principal component analysis, and double seasonal ARIMA for electricity demand forecasting up to a day ahead for Rio de Janeiro and a series of half-hourly demand for England and Wales. MLR is another statistical approach which is widely used for the energy demand estimation [2]. A major drawback of the regression and other statistical analyses is that they model the nature of the corresponding problem by a pre-defined linear or nonlinear equation [12].

Over the last decade, machine learning has attracted much attention in both academic and empirical fields for tackling real world problems. The machine learning systems are powerful tools for design of computer programs. They automatically learn from experience and extract various discriminators [13]. Artificial neural networks (ANNs) are the most widely used branch of machine learning. ANNs have been widely applied to different problems in energy conversion and management (e.g. [14], [15], [16], [17], [18], [19]). Hsu and Chen [20] considered GDP, population and temperature for the ANN-based modeling of the regional peak load in Taiwan. Ringwood et al. [21] proposed prediction models for electricity demand on short, medium and long time scales in the Republic of Ireland using ANNs. Catalao al. [22] applied the ANN to predict the next-week prices in the electricity market of Spain. Azadeh et al. [23], [24] utilized ANNs for the energy demand estimation in Iran. Azadeh et al. [1] presented a novel integrated fuzzy system, data mining and time series framework to predict the electricity demand for seasonal and monthly changes in electricity consumption especially in developing countries such as China and Iran with non-stationary data. Azadeh et al. [25] used ANN and time series for forecasting electrical consumption. The advantage of ANN methodology through analysis of variance (ANOVA) was shown by Azadeh et al. [25]. Support vector model (SVM) is another machine learning technique that has been successfully applied to predict the electric consumption [26]. In this context, Fan and Chen [27] developed a load forecasting method for short-term load forecasting, based on an adaptive two-stage hybrid network with self-organized map (SOM) and SVM. Using the actual data taken from the New York ISO and the Western Farmers Electric Cooperative in Oklahoma, Fan et al. [28] proposed a novel forecasting model for day ahead electricity load forecasting based on bayesian clustering by dynamics (BCD) and support vector regression (SVR). The BCD–SVR model adopted an integrated architecture to handle the non-stationarity of time series. Despite having a good performance, ANNs and SVM are considered black-box models. That is, they do not provide the knowledge of process they obtain a solution. Thus, they are not capable of generating practical prediction equations. Another major limitation of the ANNs is that the structure of the network needs to be identified in advance [7], [29].

A robust machine learning approach for predicting electricity demand is genetic programming (GP) [30]. GP is a robust branch of evolutionary algorithms (EAs) that creates computer programs to solve a problem using the principle of Darwinian natural selection. GP has several advantages over the conventional and ANN techniques. A notable feature of GP is that it can produce prediction equations without a need to define the form of the existing relationship [29], [31], [32]. Lee et al. [33] showed that GP is a powerful tool for long-term forecasting of electric power demand in Korea. Later, Bhattacharya et al. [34] employed GP for modeling electricity demand prediction in Australia.

Simulated annealing (SA) is a well-known optimization method. The Metropolis algorithm [35] is known as the foundation of SA to simulate the annealing process [29]. Some researchers coupled GP and SA to improve the efficiency of GP [36], [37]. There are very limited applications of the hybrid GP and SA technique to solve real world problems [31], [32], [38], [39]. Unlike ANNs, SVM and classical GP, applications of this hybrid method in the field energy conversion and management are conspicuous by their absence [29].

This study investigates the potential of the coupled technique of GP and SA (GSA), to predict the total electricity demand in Thailand from 1986 to 2009. The results made by the developed GSA models were further compared with those obtained by a conventional statistical method. The paper is organized as follows: Section 2 presents brief descriptions of the standard GP and GSA techniques. Section 3 outlines the model development using GSA and reviews the results. The detailed performance analysis of the proposed model is discussed in Section 4. The results of the sensitivity analysis are given in Section 5. Finally, concluding remarks are outlined in Section 6.