A multi-energy meta-model strategy for multi-step ahead energy load forecasting

Abstract Energy load forecasting (ELF) is necessary for optimal scheduling of electricity distribution to customers, reducing energy losses and for regulating power distribution. There are numerous approaches for ELF based on statistical, Machine Learning (ML) and deep learning (DL) methods. This paper performs a comparative multi-step ahead ELF utilising rolling (or recursive) versus direct forecasting techniques for three energy types (ETs), i.e. heating, cooling and electricity. It introduces a multi-energy meta-model strategy (MEMMS), a unified approach that combines the best outcomes of the comparison and utilises one model per time step ahead for all ETs. The findings were validated on three datasets, one per ET, employing metrics such as mean absolute error (MAE), root-mean-squared error (RMSE), $$R^{2}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup> <mml:mi>R</mml:mi> <mml:mn>2</mml:mn> </mml:msup> </mml:math> and execution time. The novel aspects of this work can be attributed to the possible identification of the best-performing ML and DL models across all ETs, as well as the multi-energy behaviour recognition of the multi-step ahead strategies and the employed MEMMS approach. Initial comparison results indicated that direct multi-step ahead voting ensemble with light gradient boosting machine (LGBM) and CatBoost (CB) was the best-performing baseline ensemble approach. MEMMS reduced training time by 50%, achieving up to 2 kW lower RMSE, 0.5 kW lower MAE and 5% higher $$R^{2}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup> <mml:mi>R</mml:mi> <mml:mn>2</mml:mn> </mml:msup> </mml:math> for cooling, and up to 2 kW lower RMSE, 4 kW higher MAE and 2% higher $$R^{2}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup> <mml:mi>R</mml:mi> <mml:mn>2</mml:mn> </mml:msup> </mml:math> for heating, showcasing versatility across ETs and forecasting horizons.