Document Type : Original Article

Authors

• Mohammad Nemati
• Iraj Teimouri
• Shahrivar Roustaie

Department of urban and regional planning, Faculty of planning and environmental sciences, University of Tabriz, Tabriz, Iran.

Abstract

This study investigates the climatic and urban factors influencing residential electricity consumption in Tabriz, Iran, a city with a cold semi-arid climate, using remote sensing data and machine learning (Random Forest, LightGBM) and deep learning (LSTM, CNN) models. Bi-monthly data from 2008 to 2023, including LST (day/night), NDVI, NDBI, NDWI, DMSP-OLS/VIIRS Nighttime Lights, Precipitation, Relative Humidity, Solar Radiation, Sunshine Hours, and AQI (NO₂, O₃, Aerosol), were used. Electricity consumption data were obtained from the Tabriz Electricity Distribution Company. Variables were Z-score normalized and processed in Google Earth Engine. Four models—Random Forest (RF), LightGBM, LSTM, and CNN—were implemented in Google Colab. Performance was evaluated using R², RMSE, and MAE, with feature importance assessed via SHAP analysis. Random Forest outperformed others (R²=0.91, RMSE=0.29, MAE=0.17), effectively capturing nonlinear relationships. Key drivers were NDWI (0.228), Night Light (0.211), and Solar Radiation (0.189), highlighting water scarcity, urbanization, and radiative heating impacts. LightGBM showed moderate performance (R²=0.64), emphasizing NDBI and Precipitation, while LSTM and CNN underperformed (R²<0) due to limited data and non-sequential variables. SHAP and residual analyses confirmed RF’s robust, stable predictions with minimal bias compared to LightGBM’s scale-dependent errors and deep models’ irregular residuals. In Tabriz’s cold semi-arid climate, residential electricity consumption is driven by NDWI, urbanization (Night Light), and solar radiation. These reflect increased heating/cooling needs due to low precipitation, urban heat islands, and rising temperatures. Green infrastructure and passive solar design are recommended to mitigate demand. This approach suits data-scarce regions, with potential enhancements via household behavioral data integration.

Keywords

• Residential Electricity Consumption
• Remote Sensing
• Cold Semi-Arid Climate
• Machine Learning
• Deep Learning

Main Subjects

• Geography and Urban Planning