Improving Load Forecasting with Feature Selection via XAI and Expert-Guided Prompting

Accurate electricity load forecasting plays a vital role in optimizing energy production and management within smart grids. While deep learning models such as LSTM, GRU, and Transformer architectures have achieved strong predictive performance, they remain susceptible to overfitting and inefficiency when trained with redundant or irrelevant features. Traditional feature selection methods mitigate these issues but often ignore domain knowledge and rely solely on statistical or model-derived importance. This study proposes a novel Explainability- and Knowledge-driven Feature Selection framework that integrates explainable AI (XAI) techniques with structured expert reasoning via large language models. By combining SHAP- and LIME-based feature importance scores with domain-informed prompting, the proposed approach enables more reliable and interpretable feature selection. The framework was validated on both short-term and long-term load forecasting tasks using two benchmark datasets (EPC and UMass Smart*), and across diverse architectures including CNN-LSTM and Transformer models. Experimental results demonstrate that combining XAI-derived feature importance with LLM-based expert knowledge leads to more accurate and consistent feature selection. The retrained models using these selected features achieved notable improvements in forecasting performance and efficiency across different datasets and architectures. These findings confirm that integrating XAI and expert-guided prompting provides a practical and effective mechanism for data-driven feature optimization in electricity load forecasting.