Interpretable AI Hybrid Model for Electricity Demand Forecasting: Combining TFT and XGBoost in Smart Grid Data

Accurate electricity load forecasting is crucial for optimizing power distribution networks, especially in rapidly growing cities like Tabriz (annual consumption growth of 7.2%). This study presents a hybrid AI framework integrating the Temporal Fusion Transformer (TFT) and XGBoost for residual error correction. The model is trained and evaluated using actual consumption data from Tabriz’s distribution network (2021–2023). Compared to a baseline TFT model, the proposed framework demonstrates a 11.2% reduction in RMSE (from 0.1249 to 0.1109) and a 10.7% decrease in MAE (from 0.0998 to 0.0891). Attention mechanism analysis reveals temperature (importance coefficient = 0.32), weekly patterns (0.18), and industrial activity (0.21) as key factors influencing electricity consumption in Tabriz. Achieving a MAPE of 4.2%, the framework provides actionable insights into consumption drivers. This research demonstrates the effectiveness of the proposed model in managing load fluctuations characteristic of medium-sized cities and offers potential for adaptation to similar urban contexts. The dual capability for accurate prediction and interpretable feature analysis establishes a new benchmark for smart grid analytics in emerging smart city environments.