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

Area of Science:

• Building energy management
• Artificial intelligence in energy systems
• Predictive control

Background:

• Conventional predictive control prioritizes prediction accuracy over transparency.
• Building energy optimization requires interpretable models for effective decision support.

Purpose of the Study:

• To present EcoImpact, a novel interpretable predictive-control framework for building energy management.
• To integrate data-driven forecasting, iterative control optimization, and explainable artificial intelligence.
• To enhance control performance and provide transparent decision support.

Main Methods:

• Utilized Random Forest and XGBoost models within a custom predictive-control structure.
• Employed SHAP (SHapley Additive exPlanations) for model interpretability.
• Integrated machine learning-based control with explainable AI in a unified workflow.

Main Results:

• Demonstrated substantial error reduction across iterative runs for both Random Forest (final error: 109.19) and XGBoost (final error: 92.53).
• SHAP analysis identified key operational and environmental factors influencing energy predictions.
• Validated the effectiveness of the EcoImpact framework in improving predictive performance.

Conclusions:

• EcoImpact successfully combines predictive control and explainable machine learning for transparent building energy management.
• The framework enhances predictive performance while ensuring model behavior is interpretable.
• Supports more efficient, reliable, and explainable building energy-management strategies.