Interpretable Deep Learning for Heat Demand Forecasting in Buildings : Explaining BiLSTM Predictions with SHAP

Buildings account for a substantial share of global energy consumption. Enhancing energy efficiency in buildings is crucial for both environmental sustainability and economic efficiency. Accurate and interpretable forecasting of energy demand is particularly critical in cold-climate regions to ensure efficient energy management, reduce wastage, and provide stakeholders with transparent insights into decision-making. This thesis presents the application of interpretable deep learning technique to forecast daily heat consumption in a school located in Northern Sweden. It specifically focuses on enhancing prediction accuracy while maintaining model transparency. A Bidirectional Long Short-Term Memory (BiLSTM) neural network is designed to capture temporal dependencies in multivariate time series data, comprising weather variables, calendar features, and historical consumption. To enable model interpretability, SHAP(SHapley Additive exPlanations) is applied to analyze global feature contributions and provide local explanations for individual predictions.