Insights into Long-Term Electrical Load Forecasting: Explainable AI Approach on Multivariate LSTM

With machine learning (ML) and deep learning (DL) methods becoming evident, electrical load forecasting has become more accurate than traditional ways. Most applications have been focused on small and midterm forecasting, ranging from a few hours to a few days. One reason is the lack of proper data that encapsulates long-term load trends. Moreover, ML and DL techniques are claimed to be black box models, as the generated output has little to no explanation or insight into the reasoning. This study creates a dataset for long-term load forecasting by merging several economic, electrical load and pricing factors. Several ML and DL methods were tested on the dataset and compared. XGBoost, Random Forest, Boosted Linear Trees and Prophet model were selected from ML methods. Long Short-Term Memory (LSTM) was the DL method. We propose a novel n-point multivariate LSTM approach for robust electrical load forecasting. This approach can perform short-term, midterm and long-term load forecasting, capturing the demand trend for any period. Our model showed a greater R2 score and a much lower MAPE than existing studies. The extended research focused on the interpretability of the model’s output, by utilizing SHAP analysis. This allowed for an in-depth explanation of time and economic components affecting the load trend while making the decision-making robust. Our novel and comprehensive time series forecasting framework provides insights into electrical load forecasting by combining new data preparation approaches, model selection, performance assessments, and interpretability analyses. This will aid in capacity expansion, infrastructure development and maintenance scheduling of power plants for both generation and distribution.