Forecasting short-term electric energy consumption using validated ensemble learning

Short-term load forecasting (STLF) for a country is the key to planning the day-to-day operations of its power system to match the demand and the supply without interruption. Thailand is an industrially emerging country where the demand tends to rise in large numbers in the coming years (about 78% by 2036). The daily demand is also subjected to fluctuation due to industrial operations, tropical climate, and the considerable amount of holidays per year. Practitioners have utilized classical and machine learning (ML) models to forecast such a demand. However, an emerging trend is to build ensemble learning (EL) models by effectively combining them. Therefore, this study primarily builds an EL model to forecast Thailand’s short-term electric energy demand.The proposed EL model uses the voting regression (VR) with weighted averaging. It was selected after an extensive comparison between the prediction accuracy of itself and five other individual models used to build it. Three of them are multiple linear regression (MLR)-based parametric models, which use linear regression (LR), ordinary least squares (OLS) regression, and generalized least squares auto-regression (GLSAR) as their estimators. The other ML models use nonlinear nonparametric regressive estimators: decision tree (DT) and random forest (RF).In contrast to many other time series forecasting (TSF) models, which omit the cross-validation (CV) in the literature, this research identifies an effective CV scheme called Blocked-CV. It was selected by comparing it with two other validation schemes: Random CV and expanding window forward validation (EWFV). In the beginning, the available dataset was divided into four groups, and several features, including historical loads, temperature, deterministic variables, and some other interaction terms, were realized. These features were then selectively used to build models in each group, and the validation and test errors were compared. Results suggest that the Blocked-CV outperforms the other two validation schemes producing a minimum difference between the validation error and the test error 75% times. Further, the VR model results in a minimum validation error for this selected Blocked-CV scheme in each group, proving that it also results in a minimum test error in each group.