Enhancing LSTM prediction accuracy through hybrid feature engineering: An LSTM-based model for load forecasting of industry in Thailand

This study aims to improve the accuracy of electrical load forecasting in the industrial sector by enhancing the performance of the Long Short-Term Memory (LSTM) model and comparing three deep learning methods: the standard LSTM model, the hybrid CNN–LSTM model, and the Feature Engineering LSTM (FE-LSTM) model. This research uses historical industrial electricity consumption data, environmental temperature data, and data on working days and holidays. This data will be used to forecast electricity load for all three models and evaluate the models' performance using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and R² value. The research results show that the FE-LSTM model outperforms the other models, with the highest R² value of 0.9447, and can reduce the RMSE from 13,220.08 W (standard LSTM) to 8,231.62 W, indicating an improvement in electrical load forecasting by 37.7%. The study concludes that the integration of hybrid feature engineering techniques, such as lagged features, moving averages, and calendar variables, significantly enhances forecasting accuracy. The proposed FE-LSTM model is suitable for short-term industrial load forecasting and can support energy management planning, cost optimization, and operational decision-making in the industrial sector.