Data-driven multi-step energy consumption forecasting with complex seasonality patterns and exogenous variables: Model accuracy assessment in change point neighborhoods

Abstract

Energy forecasting tools became a significant scope of application for time series modeling due to the specific challenges in energy trading — the forecast of consumption for the whole next trading day based on the limited data availability at the forecast origin. The research article addresses the scope of high-frequency time series data, multiple seasonal patterns, exogenous variables, and nonstationary properties in a multi-step forecast horizon tasks. The contribution of the research is the introduction of a machine- and deep-learning-based data-driven approach for multi-output time series forecasting and mainly an introduction of the new evaluation metric called the Change Point Neighborhood Error (CPNE). The purpose of the metric is to provide a distinctive measure of forecasting accuracy of the proposed models in parts of the time series where a change point or a data drift emerges. The experimental findings indicate a notable improvement in accuracy achieved by machine and deep learning models, resulting in a substantial reduction of the mean absolute percentage error (MAPE) by approximately 45 % compared to the optimal statistical model across both datasets used, and also in terms of change point neighborhood error in comparison to statistical models due to their requirement for stationary data input. Deep learning models may be a viable alternative to machine learning approaches; however, deep learning models require long input sequences for accurate forecasting, whereas machine learning methods require shorter input sequences and can benefit more from feature engineering.