Robust explainable forecasting via heterogeneous sensor fusion with applications in residential peak load management

High electricity consumption peaks in the residential sector can lead to expensive grid upgrades and higher consumer costs. Traditional forecasting methods often lack the contextual awareness to predict the behaviour-driven spikes. To address this problem, this paper proposes a simulated robust, closed-loop architecture for residential demand-side management that integrates multi-sensor fusion and machine learning. We perform a proof-of-concept evaluation of our proposed system on a publicly available household electricity consumption dataset. In our experimental evaluation, by leveraging internal and external environmental data as proxies for human activity, our system identifies high-demand peaks at the household level. We investigate the forecasting window from the perspective of the stability–sensitivity trade-off for peak and minimum consumption. Integrating heterogeneous sensors reduced peak forecasting errors by 29%. We perform a detailed analysis of the required sensors and their placement to improve peak detection, and evaluate sensor robustness to assess system performance under various hardware or network failure models. Simulation results in a limited-data scenario demonstrate a peak reduction of up to 36.36% without compromising essential household functions or total energy efficiency.