Population Prediction Using Hierarchical LightGBM Based on Feature Extraction from Electricity Indicators

Population prediction is crucial for urban planning but traditionally relies on time-consuming census methods such as demographic surveys and census data. This paper proposes a hierarchical LightGBM framework that leverages electricity consumption patterns as demographic proxies. We develop a comprehensive feature extraction pipeline capturing temporal, statistical, and behavioral indicators from household electricity data, including vacancy rates, temperature sensitivity, and consumption patterns. Our two-stage hierarchical LightGBM model first processes top-3 features for primary signals, then exploits complementary features for residual patterns, combining predictions through weighted ensembling. Evaluated on real-world data from a mega city in Western China, the model achieves state-of-the-art performance with an RMSE of 0.1804 and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.9746, outperforming traditional forecasting techniques. The framework demonstrates strong interpretability through feature importance analysis and provides actionable insights for urban resource allocation.