Forecasting week-ahead hourly electricity prices in Belgium with statistical and machine learning methods

Electricity price forecasting is a challenging task as it involves predicting series that are influenced by numerous variables, such as weather conditions, electricity consumption, and seasonal factors. Yet, accurate forecasts are necessary for supporting operational management and short- to mid-term planning of energy companies. Over the years, various forecasting approaches have been proposed in the literature to perform this task, including statistical and machine learning (ML) methods. However, the studies comparing their performance have been inconclusive with regards to the superiority of the one type of technique over the other in the task of forecasting for energy markets. Moreover, the value-added of explanatory variables in extrapolation tasks must be separately assessed for statistical and ML methods. This chapter provides an overview of both approaches. It compares the forecasting performance of two popular ML methods, namely neural networks and random forest, to that of traditional, statistical ones by considering the electricity market in Belgium. Our results indicate that ML methods provide better forecasts, both in terms of accuracy and bias. While external variables improve the performance of both statistical and ML approaches, the improvements in the latter group are relatively more substantial. Moreover, we show the beneficial effects of combining forecasts within but also across families. Finally, we discuss the limitations of each approach and suggest avenues for future research.