Machine learning-based energy consumption models for rural housing envelope retrofits incorporating uncertainty: A case study in Jiaxian, China

Rural housing envelope retrofits significantly affect energy consumption, yet traditional simulation-based assessments are often time intensive and repetitive. This study presents a novel framework that integrates uncertainty analysis (UA) and machine learning (ML) to increase the accuracy and efficiency of rural housing envelope retrofits, with a focus on key envelope components such as exterior walls, roofs, and windows. Reference models were established via field surveys and monitoring data, and multiple energy datasets were generated via UA. Data within 75%, 90%, 95%, and 100% intervals were used to train the ML models. Sensitivity analysis, employing standardized regression coefficients, random forests, and the treed gaussian process, identified key factors such as cooling and heating setpoint temperatures, infiltration, and exterior wall construction and roof construction. These factors serve as inputs for ML models built with neural networks and random forests. The coverage ratio and evenness were used to assess residual distributions. Results indicate that a more uniform residual distribution within the 95% interval balances data volume and prediction accuracy, reducing error variability and improving model performance. This study demonstrates that uncertainty-informed datasets and ML enhance the reliability and generalizability of energy consumption predictions, providing a scalable approach for optimizing rural housing energy retrofits. • Combining dynamic simulation of uncertainty with machine learning. • Data screening for uncertainty results using 75%, 90%, 95% and 100% intervals. • Constructing prediction models using BP neural network and random forest. • Residual analysis across different intervals using coverage ratio and evenness • Build predictive models that adapts to the diverse characteristics of rural house.