Exploring Day-ahead Load Forecasting Based on Bi-LSTM Model

Abstract Power system planning and operation necessitate robust load forecasting, covering predictions across long, medium, and short timeframes. Of these, day-ahead forecasting emerges as a pivotal element for various smart grid functionalities, including distributed energy resource management systems (DERMS), demand response (DR) programs, and the effective functioning of electricity markets. Recent research has high-lighted the usefulness of Long Short-Term Memory (LSTM) models in superior predictive capabilities. This study contributes to the field by exploring a Bi-directional LSTM (Bi-LSTM) model incorporating two time-series features for enhanced day-ahead load forecasting. The proposed predictive model is empirically compared to multi-variable time series models, revealing the efficiencies of the Bi-LSTM. The paper proposes a model training approach predicated on data partitioning and the construction of sample datasets method. The model was extensively evaluated using real Southern Load datasets from Vietnam, demonstrating its effectiveness with low Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) values. Specifically, the Bi-LSTM model achieved a MAPE of 2.67% on regular days and a significantly lower 1.58% during the holiday period, thus outperforming the conventional LSTM baseline.