Drift-Aware Adaptive Fusion of Machine Learning Models for Interpretable Residential Energy Forecasting

Main Article Content

Ali Layth Alzaidy
Zainab Zaydon Alnuaimy

Abstract

Well-crafted residential energy forecasting forms a fundamental part of Smart Home Energy Management Systems (SHEMS); however, the traditional static models face challenges in learning the highly volatile consumption patterns. To meet this challenge, this work introduces an innovative technique of a drift-aware adaptive fusion framework that combines and adapts machine learning algorithms in real-time via Particle Swarm Optimization (PSO). PSO adapts ensemble weights every time concept drift is found (i.e., whenever the prediction error exceeds 25% of the baseline RMSE), avoiding full model retraining, which can be computation-costly. When used with high-resolution smart home data, the adaptive ensemble offered superior performance across all evaluation metrics (overall RMSE = 0.686 kW, MAE = 0.336 kW, R² = 0.580) and exhibited higher prediction stability regardless of the characteristics/complexity level of individual models. Secondly, by employing Shapley Additive Explanations (SHAP) on integrated ensemble outputs, we eliminated the “black box” approach of the ensemble, demonstrating ambient temperature was the main predictor. This framework balances predictability, computational efficiency, and explicability to provide a practical and interpretable decision-support tool for accelerated energy optimization with high credibility in the field.

Downloads

Download data is not yet available.

Article Details

How to Cite
Layth Alzaidy, A., & Zaydon Alnuaimy, Z. (2026). Drift-Aware Adaptive Fusion of Machine Learning Models for Interpretable Residential Energy Forecasting. Al-Ma’moon College Journal , 46(2), 24–34. Retrieved from https://journal.almamonuc.edu.iq/index.php/jac/article/view/280
Section
Articles