Decoding Energy Usage Predictions: An Application of XAI Techniques for Enhanced Model Interpretability

Abstract

The growing complexity of machine learning models has heightened the need for interpretability, particularly in applications impacting resource management and sustainability. This study addresses the challenge of interpreting predictions from sophisticated machine learning models used for building energy consumption predictions. By leveraging Explainable AI (XAI) techniques, including Permutation Importance, SHapley Additive exPlanations (SHAP), and Local Interpretable Model-Agnostic Explanations (LIME), we have dissected the predictive features influencing building energy usage. Our research delves into a dataset consisting of various building characteristics and weather conditions, applying an XGBoost model to predict Site Energy Usage Intensity (Site EUI). The Permutation Importance method elucidated the global significance of features across the dataset, while SHAP provided a dual perspective, revealing both the global importance and local impact of features on individual predictions. Complementing these, LIME offered rapid, locally focused interpretations, showcasing its utility for instances where immediate insights are essential. The findings indicate that 'Energy Star Rating', 'Facility Type', and 'Floor Area' are among the top predictors of energy consumption, with environmental factors also contributing to the models' decisions. The application of XAI techniques yielded a nuanced understanding of the model's behavior, enhancing transparency and fostering trust in the predictions. This study contributes to the field of sustainable energy management by demonstrating the application of XAI for insightful model interpretation, reinforcing the significance of interpretable AI in the development of energy policies and efficiency strategies. Our approach exemplifies the balance between predictive accuracy and the necessity for model transparency, advocating for the continued integration of XAI in AI-driven decision-making processes.