A hybrid temporal fusion transformer and LightGBM framework for accurate multi-horizon industrial electricity load forecasting
TL;DR
A novel interpretable hybrid forecasting framework that integrates the Temporal Fusion Transformer (TFT) with LightGBM that offers a balanced combination of high predictive accuracy, computational efficiency, and explainability, making it particularly suitable for smart grid applications in industrial contexts facing similar demand dynamics and data constraints.
Abstract
Accurate short-to medium-term electricity load forecasting is essential for efficient operation and planning of smart grids, particularly in medium-sized urban-industrial cities experiencing rapid demand growth and heterogeneous consumption patterns. This study proposes a novel interpretable hybrid forecasting framework that integrates the Temporal Fusion Transformer (TFT) with LightGBM. The TFT component captures complex multivariate temporal dependencies and provides variable importance through its attention and variable selection mechanisms, while LightGBM performs efficient residual correction and nonlinear pattern modeling. The model was trained and evaluated on real hourly consumption data recorded from 2023 to 2024 across active industrial units in Tabriz, incorporating meteorological, calendar, and industrial activity variables. Compared to the standalone TFT baseline, the proposed TFT-LightGBM hybrid reduced root mean square error (RMSE) from 0.128 to 0.113 (11.7% improvement), mean absolute error (MAE) from 0.107 to 0.0921 (13.9% improvement), and mean square error (MSE) from 0.0164 to 0.0128. It also outperforming six established benchmarks (LSTM, GRU, Informer, SVR and Patch-TST) across all metrics. The proposed framework offers a balanced combination of high predictive accuracy, computational efficiency, and explainability, making it particularly suitable for smart grid applications in industrial contexts facing similar demand dynamics and data constraints.