Village electricity consumption prediction using CNN-DNN framework integrating construction land morphology: A Case study of 285 villages in cold regions

With the deepening implementation of China’s rural revitalization strategy, village electricity consumption (VEC) prediction has become a crucial component of energy planning. However, traditional methods face challenges including data acquisition difficulties and slow prediction speeds. This study focuses on construction land and develops a deep learning framework integrating morphology and socioeconomic data for VEC prediction. Using 285 villages in cold regions of China as research subjects, a CNN-DNN dual-model framework was developed: CNN automatically extracts nine morphology parameters from high-resolution satellite imagery, combines with three socioeconomic data to construct feature sets, and DNN realizes VEC prediction. This method not only avoids the time and cost limitations of traditional field surveys, but also solves the data acquisition challenges in rural areas, making large-scale VEC assessment possible. SHAP analysis reveals that residential land area (RLA) and permanent population (PP) are the dominant predictive factors, exhibiting nonlinear threshold effects. The model demonstrates excellent performance (R² = 0.938, RMSE = 815,821 kWh), identifying three types of VEC patterns: high-consumption villages (15 %), medium-consumption villages (64 %), and low-consumption villages (21 %), showing significant spatial differentiation characteristics. This framework provides scientific tools and methods for rapid assessment of VEC and formulation of differentiated electricity policies in cold regions of China.