Short Term Electricity Load Forecasting with Alternative Cross Validation Approaches

Electricity load forecasting has served as the foundation for predictive and prescriptive analytics problems in the energy analytics domain. Accurate forecasts of the electricity demand provide an important advantage in estimating the hourly market clearing price for electricity since it can be seen as the main driver for its fluctuations. Such forecasts can be inputs to many optimization problems related to portfolio optimization for a power producer. In this study, short term electricity demand will be taken into consideration as a multivariate series forecasting problem. Hourly electricity consumption data starting from January 2016 up to January 2025 from Turkey has been included in the experiments. Several deep learning algorithms such as Temporal Fusion Transformer, N-Beats and NHits has been used alongside a relatively more conventional forecasting approach, LightGBM. A model selection technique that is developed for High-Frequency Trading domain, Combinatorial Purged K-Fold Cross Validation will be extended into a problem with a non-financial dataset.