In response to the lack of rigour in identification and prediction of risks related to energy-retrofitting prevision in urban renewal interventions, this research aims to formulate an advanced risk prediction framework that incorporates multi-source GIS information for the retrofit process of old residential areas during urban renewal, using a spatially regularized Random Forest–Deep Neural Network (RF–DNN) hybrid model with high accuracy. A multi-dimensional feature library is built, which covers geospatial features and building energy related attributes, as well as socioeconomic factors and environmental effects. The input structure is optimized by multi-scale spatial aggregation and dimension reduction for improved spatial interpretability, thereby achieving computational efficiency. The developed hybrid model can appropriately assess technical, economic and environmental risks in retrofit projects. Experimental results show that the spatially regularized RF–DNN hybrid model exhibits much better predictive performance, robustness and interpretability than traditional single-source and single-model methods. The system offers a data-driven decision-making tool for the urban renewal policy makers and project administrators to make more accurate and flexible risk assessment on large-scaled energy efficiency retrofit operation.