Research Information System
ENERGY AND BUILDINGS, vol.352, pp.1-12, 2026 (SCI-Expanded, Scopus)
The forecasting of electrical load is a very important step for the reliable and sustainable operation of electrical power systems. Bidirectional long short-term memory network (BILSTM) models have been widely applied to short-term load forecasting. However, their performance when combined with specific input features remains largely unexplored. To address this gap, this study propose a BILSTM model using a set of input features that has not been previously investigated. The model is evaluated across six scenarios, both with and without data decomposition. This analysis provides insights into how input configuration and preprocessing strategies affect forecasting accuracy for the Trabzon dataset. The most successful scenario is applied to another dataset, the Checz dataset, which has been utilised in the existing literature. Compared to existing studies, the most successful scenario proves capable of producing more accurate predictions while requiring fewer sensor measurements. It outperforms other studies that utilized the Czech dataset, reducing the forecasting error by 34.52% and 23.75% according to the mean absolute percentage error (MAPE) criterion.