EcoImpact: energy conservation using data-driven model predictive control and interpretable machine learning in the...

This paper presents EcoImpact, a novel interpretable predictive-control framework for building energy management that integrates data-driven forecasting, iterative control optimization, and explainable artificial intelligence in a unified workflow. Unlike conventional predictive-control approaches that focus mainly on prediction accuracy, EcoImpact explicitly combines machine-learning-based control with SHAP-based interpretability to identify the operational and environmental factors driving energy-related predictions. This enables not only improved control performance but also transparent decision support for building energy optimization. The proposed framework uses Random Forest and XGBoost models within a custom predictive-control structure to iteratively reduce prediction error and improve energy-management decisions. Experimental results demonstrate that both models achieve substantial error reduction across iterative runs, with final errors of 109.19 for Random Forest and 92.53 for XGBoost, indicating the effectiveness of the proposed framework. In addition, SHAP analysis reveals the most influential features affecting model predictions, providing insight into how the control model prioritizes input variables during decision-making. The main innovation of this study lies in combining predictive control and explainable machine learning into a transparent optimization framework for building energy management. The results show that EcoImpact can improve predictive performance while making model behavior interpretable, supporting more efficient, reliable, and explainable building energy-management strategies.