Скачать или смотреть Forecasting household electricity demand using machine learning algorithms - A case of Thailand

Nattapong Puttanapong (1)
(1) Faculty of Economics, Thammasat University

Because electricity generation is generally a long-term operation requiring large capital investment, the accurate projection on future electricity demand is inevitably a crucial factor. To smoothly sustain the economic growth, the Thai government has recognized the significance of long-term planning enabling sufficient and efficient electricity generation. Therefore, the national Power Development Plan (hereafter referred to as "PDP") has been regularly formulated since 1992. Specifically, load forecasting and the stability of the power system are the main deliverables of PDP. In addition, household electricity consumption is one of the main components of electrical load. However, the load forecast error has been a serious concern. In November 2020, it caused the power generation reserve to reach 50% of total power generation capacity, substantially higher than the internationally recommended rate of 15-20%.

With the necessity of improving the accuracy of load forecasts, this study introduced the alternative forecasting technique for household electricity demand by applying the machine learning methods to the household-level data. Specifically, the methods of generalized least squares (GLS) regression, artificial neural network (ANN), random forest (RF) and support vector regression (SVR) were applied to the data of Socioeconomic Survey (SES), which is the nationwide household survey officially conducted by the National Statistical Office every two years. The datasets cover the period 2006--2015, and the number of samplings continuously increased from 41,814 in 2006 to 37,008 in 2015. Each sampling included the monthly average household’s electricity consumption (kWh), the monthly average income (Thai baht), the average temperature (degrees Celsius), and the numbers of possessed electrical appliances classified into 15 categories. Also, the data included the number of family members, owned vehicles, the number of rooms and the classification of building materials of the house. The 100-fold resampling technique was applied to the dataset, and the values of root-mean-square error (RMSE) obtained from four machine learning algorithms were compared. For all annual datasets, RF yielded the lowest RMSE, while those of SVR, ANN and GLS ranked second, third and fourth, respectively. In addition, RF incorporated the analyses of variable importance (VIMP) and minimal depth (MD), quantifying each factor's degree of influence on electricity consumption. VIMP and MD outcomes show that the ownership of air conditioners, the average temperature, the monthly average income and the number of family members are the most influential factors.

With unique characteristics of socioeconomic survey data composed of a diverse combination of discrete and continuous variables, all obtained results indicate that RF is the most appropriate technique. These outcomes also suggest the potential application of using RF as a bottom-up approach for residential load forecasting, which is the alternative to the top-down one conventionally used in PDP. In addition, this proposed bottom-up technique can indicate the electricity demand in the spatial dimension, supporting the planning of distribution networks with high accuracy.