Feature Modelling for Electricity Load Forecasting Using Hybrid EEMD-BiLSTM with Spatially Correlated Weather Data

Abstract

Accurate electricity load forecasting is crucial for efficient energy management, allowing utility companies to optimize resource allocation, reduce operational costs, and ensure a reliable power supply. Various factors, such as weather conditions, economic activities, and demographic changes, significantly influence electricity load. To address these challenges, we designed a deep learning Bidirectional Long Short-Term Memory (BiLTSM-based model) for electricity load forecasting, leveraging weather data as the primary feature and signal decomposition as an additional feature. To select the best weather parameters, we utilize the Principal Component Analysis (PCA) and Correlation Coefficient (CC) to choose the most correlated weather parameters as inputs for the model. Moreover, we introduce spatial correlation to select the best weather locations. Additionally, we perform a signal decomposition technique using the Ensemble Empirical Mode Decomposition (EEMD) to decompose the historical load signal into Intrinsic Mode Functions (IMFs). We also investigate procedures to choose the best IMF to give the best accuracy. As a case study, we investigate the electricity load data in the Jakarta-Banten region of Indonesia, known for its high consumer and industrial activity. The results demonstrate that combining the BiLSTM model and EEMD achieves high prediction accuracy, with a Correlation Coefficient (CC) of 0.956 and a Mean Absolute Percentage Error (MAPE) of 2.824%.