<p>The rapid digitalization of energy systems through smart meters provides vast data streams that, when contextualized with socioeconomic information, can reveal the underlying behavioral mechanisms of electricity use. However, standard machine learning pipelines often overlook the non-Euclidean relationships embedded within social structures and spatial proximity, leading to limited predictive reliability. This paper presents a Hypergraph-Driven Hybrid Intelligence Framework (HHIF) that synthesizes hypergraph attention, graph convolution, and temporal learning to create an adaptive model for residential load forecasting. The proposed system encodes complex socio-energy interactions using Hypergraph Attention Networks (HANs), integrates K-means clustering for structure-aware feature extraction, and fuses GCN and LSTM modules for joint spatial–temporal optimization. Experiments on multi-city smart meter datasets highlight substantial gains in accuracy, stability under noise, and interpretability, underscoring the framework’s applicability for grid management, peak load mitigation, and energy equity research.</p> Graphical abstract