A Residual Error Correction Hybrid CNN–LightGBM Model for Accurate Short-Term Electric Load Forecasting
Year: 2025
Type: article
Abstract: Short-term load forecasting (STLF) is crucial in maintaining power system reliability and supporting efficient operational planning. This research introduces a residual error correction hybrid model that combines Convolutional Neural Networks (CNN) and Light Gradient Boosting Machines (LightGBM) to enhance forecasting accuracy while assessing computational efficiency. The main contribution of this study is the development and evaluation of a CNN–LightGBM framework that leverages residual learning to reduce prediction errors compared with individual models. The proposed approach was validated using a real-world dataset of 72,096 hourly electricity load observations from the Ho Chi Minh City Power Corporation (HCMPC), covering eight consecutive years, with an average system demand of approximately 2,311.7 MW. Experimental results show that while CNN outperformed LightGBM in capturing temporal patterns, the hybrid CNN–LightGBM achieved the lowest forecasting errors: MAE of 35 MW, MSE of 2,421 MW², RMSE of 49 MW, and MAPE of 1.30%, significantly improving prediction accuracy over the standalone models. However, this gain comes with a higher computational cost, as the hybrid model required about 390 seconds to train, compared to ~4× faster CNN and ~18× faster LightGBM. These findings highlight the trade-off between accuracy and runtime, emphasizing the need to balance forecasting precision with computational efficiency depending on operational requirements. Future work will optimize training speed, explore model distillation, and improve scalability for real-time power system applications.
Authors: Tuan Anh Nguyen, Chi Kien Ha
Institution: Industrial University of Ho Chi Minh City
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7160926011
