Multi-scale electricity consumption prediction model based on land use and interpretable machine learning: A case study of China

Abstract

The prediction of electricity consumption plays a vital role in promoting sustainable development, ensuring energy security and resilience, facilitating regional planning, and integrating renewable energy sources. A novel electricity consumption characterization and prediction model based on land use was proposed. This model achieves land-use subdivision to provide highly correlated variables; exhibits strong interpretability, thereby revealing even marginal effects of land use on electricity consumption; and demonstrates high performance, thereby enabling large-scale simulations and predictions. Using 297 cities and 2,505 counties as case studies, the key findings are as follows: (1) The model demonstrates strong generalization ability (R² = 0.91), high precision (Kappa = 0.77), and robustness, with an overall prediction accuracy exceeding 80 %; (2) The marginal impact of industrial land on electricity consumption is more complex, with more efficiency achieved by limiting its area to either 104.3 km² or between 288.2 and 657.3 km²; (3) The marginal impact of commercial and residential land on electricity consumption exhibits a strong linear relationship (R² > 0.80). Restricting the scale to 11.3 km² could effectively mitigate this impact. Mixed commercial and residential land is advantageous for overall electricity consumption control, but after exceeding 43.5 km², separate layout considerations for urban residential land are necessary; (4) In 2030, Shanghai's electricity consumption is projected to reach 155,143 million kW·h, making it the highest among the 297 cities. Meanwhile, Suzhou Industrial Park leads among the 2,505 districts with a consumption of 30,996 million kW·h; (5) Identify future electricity consumption hotspots and clustering characteristics, evaluate the renewable energy potential in these hotspot areas, and propose targeted strategies accordingly.