Energy Load Modeling with Behaviorally Interpretable Clustering and LightGBM

A hybrid method for residential energy consumption modeling has been presented, in which unsupervised clustering has been integrated with gradient boosting regression. Daily load behavior has been characterized using engineered features, including time-of-day consumption averages, temperaturenormalized usage, and inter-day variability. Gaussian Mixture Models and Spectral Clustering have been applied to segment the data into behaviorally homogeneous groups, for which individual LightGBM models have been trained. Evaluation on real-world data has shown that cluster-specific models have outperformed a global model, particularly under conditions of partial data loss. Clustering has revealed behaviorally distinct groups with stable temporal profiles. Among the engineered features, total daily consumption, temperature-adjusted usage, and evening load consistently exhibited the strongest discriminative power across clusters. These findings demonstrate that the combination of Gaussian Mixture Model-based clustering and LightGBM regression yields competitive performance compared to recent approaches in the literature. The use of probabilistic soft clustering has enabled more flexible segmentation of consumption behavior, while LightGBM has been effective in modeling nonlinear dependencies within each group. The resulting approach has offered a scalable and interpretable alternative for improving prediction accuracy, even in data-sparse environments.