CausalHES: A causal source separation framework for weather-invariant household energy behavior segmentation and personalized demand management

Effective energy efficiency policies and personalized demand management require a clear picture of household consumption behavior. Observed load profiles, however, are heavily confounded by meteorological conditions, which obscures intrinsic behavioral patterns and hinders targeted interventions. This paper presents CausalHES, a deep learning framework whose objective is to learn weather-invariant behavioral representations for robust household segmentation by reframing the problem as a causal source separation task. A Causal Source Separation Autoencoder (CSSAE) disentangles raw load profiles into weather-independent base consumption and weather-dependent effects, enforcing explicit statistical independence between the latent representations of base load and weather through a composite loss that combines mutual information minimization, adversarial training, and distance correlation. Deep Embedded Clustering (DEC) then groups the purified base load embeddings, yielding robust, weather-independent segments. On data from 1000 Irish smart-meter households, CausalHES attains 88.0% accuracy in recovering socio-economic household categories, exceeding the best traditional clustering baseline (56.0%), the best weather normalization baseline (60.4%), and the best VAE-based disentanglement baseline (66.2%) by 32.0, 27.6, and 21.8 percentage points, respectively. Ablation studies confirm the necessity of the causal independence enforcement and the architectural separation, and semantic consistency analysis (r=0.78) validates the interpretability of the separated weather component, providing actionable insights for energy analytics and policy making.