Econometric and machine learning techniques for electricity load forecasting

Abstract

Load Forecasting is fundamental for the energy planning sector since a time-ahead power market requires demand-scheduling for power generation, transmission, distribution etc. Forecasting can be performed with different methods, the selection of each method relies on several factors including the quality and the relevance of the available historical data. Also, the methodology used is strictly correlated with the forecast horizon and the level of accuracy of the available data. The time horizon is being adopted taking into account the specific applications in power system planning. Id est, distribution and transmission planning need a short-term horizon while financial or power supply planning require a long-term horizon. In addition, short-term horizon is vital for the hour and day ahead market. Finally, energy forecasting is extremely important for the end consumers as they can get informed about their expected energy consumption and avoid unexpected bill shocks at the end of the year. Having stated the importance of the load forecasting, the methodology that is followed in order to forecast energy consumption is explained. After having preprocessed the data and having chosen the aggregation level and the time-horizon, the adequate model is being used in order to perform the analysis. Examples of these models are: Regression models, ANNs, SARIMAX (ARIMA family models), ANNs, time series analysis etc. For short-term and high-resolution forecasting, time series analysis and ANNs are preferred while long-term and low-resolution forecasting is mostly done with regression models. The high-resolution forecasting of a single household is extremely challenging due to the stochastic nature of the appliances’ usage. On an aggregated basis since the consumption of several households is added up extreme values and/or abnormalities are reduced to a minimum, making it easier to forecast the consumption on a higher resolution with a smaller error.