Edge-optimized and interpretable deep learning framework for high-volatility residential load prediction

The transition toward intelligent smart grids requires highly accurate, minute-resolution residential load forecasting. However, deploying deep learning models in real-world home energy management systems is often hindered by strict edge-hardware constraints and a lack of model interpretability. This study proposes a zero-lag, edge-optimized deep learning framework evaluated on high-volatility, minute-resolution smart meter data. A pareto-optimal deployment strategy is established by mathematically quantifying the trade-off between predictive performance, deterministic latency, and static memory footprint. Results from a rigorous 5-run cross-validation protocol show that the proposed hybrid long short-term memory and gated recurrent unit architecture achieves peak accuracy, while the standalone gated recurrent unit model delivers statistically equivalent predictive power with a minimal 38.75 kilobyte footprint. To overcome the ‘black-box’ nature of deep learning, a dual-validation interpretability framework is evaluated. By upgrading standard analysis of variance to the non-parametric Kruskal–Wallis test and cross-examining it against Shapley additive explanations, it is proven that the recurrent architectures achieve superior performance not by memorizing noise, but by dynamically exploiting high-frequency temporal transients. These findings provide a fully reproducible blueprint for deploying interpretable, hardware-efficient artificial intelligence in smart grids.