Inferring Daily Human Mobility Patterns from Electricity Consumption with Graph Neural Networks

With the rapid advancement of urbanization, accurately forecasting population mobility has become increasingly important for effective urban management and economic planning. This study investigates the spatiotemporal relationship between electricity consumption and inter-city population movement to enhance forecasting precision. We propose the Enhanced Population Flow Time Series Graph Convolutional Network (EPT-GCN), a novel model that integrates electricity consumption data with migration data from Gaode, while incorporating the geographical relationships among cities. EPT-GCN captures dynamic spatiotemporal dependencies in node features and consistently outperforms baseline models such as SARIMA, LSTM, and T-GCN across multiple forecasting horizons. Experimental results highlight electricity consumption time series as a key predictor for high-accuracy population flow estimation. This research demonstrates the potential of leveraging electricity data for improving urban mobility forecasts and offers valuable insights for data-driven urban governance and policy-making.