Interpretable spatiotemporal urban energy forecasting

Accurate and spatially detailed urban energy consumption forecasting is crucial for sustainable urban development. Existing methods often fail to capture the complex interplay of spatial and temporal factors influencing energy demand, hindering interpretability and limiting their effectiveness for targeted interventions. This paper presents a novel deep learning model for interpretable, multi-scale urban energy demand forecasting. Our approach leverages time series imaging to transform discrete energy consumption data into continuous spatial representations, generating energy consumption density maps. These maps are input to a deep learning encoder–forecaster architecture, enabling the model to learn intricate spatiotemporal dependencies. Crucially, by preserving the 2D spatial structure throughout the prediction process, our model offers enhanced interpretability compared to methods that reduce spatial information to 1D. We validate our model with real-world electricity data from Shanghai, demonstrating superior performance against traditional and state-of-the-art benchmarks across various spatial granularities and forecasting horizons. For a 7-day forecast, our model achieves a Mean Squared Error (MSE) of 6.032. The resulting interpretable forecasts, visualized as density maps, provide actionable insights for urban planners, policymakers, and utility operators, promoting energy efficiency and facilitating the integration of renewable energy sources into the urban fabric.