Robust Day-Ahead Short-Term Energy Forecasting Using Cyclical Encoding and Attention-Driven Recurrent Networks

The growing demand for modern power systems necessitates robust short-term load forecasting to enhance energy efficiency and maintain grid reliability. However, forecasting electricity consumption presents unique challenges due to significant fluctuating usage patterns and external factors such as electricity prices and renewable energy generation. Traditional forecasting methods often fail to capture intricate temporal dependencies and non-linear load variations, leading to suboptimal predictions. To address these challenges, this paper proposes a novel hybrid forecasting model that integrates long short-term memory (LSTM) and recurrent neural network (RNN) networks with attention mechanisms (AM). Further, this applies cyclic encoding to enhance feature representation and effectively model periodic trends. Additionally, seasonal and trend decomposition using loess (STL) with empirical mode decomposition (EMD) is employed to extract meaningful seasonal and non-seasonal patterns, while improving prediction accuracy. The model uses key input features, including historical load, electricity prices, solar and wind generation data. The experimental results demonstrate significant improvements in forecasting accuracy, with the proposed approach achieving reductions of 17.95%, 13.8%, and 14.1% in MSE, RMSE, and MAE, respectively, compared to conventional methods. These findings highlight the effectiveness of the proposed model in advancing load forecasting and enhancing power system management. The proposed system also provides a strong foundation for real-time applications in smart grids, with implications for energy storage control and power converter operation.