A universal hybrid modeling framework for MSWI temperature field prediction: fusion of physics-informed and data-driven transfer learning

Predicting the performance of municipal solid waste incineration (MSWI) accurately and quickly is crucial for achieving high efficiency and clean combustion, but it remains limited by variable conditions and on-site data. This study proposes an innovative hybrid transfer learning framework (BPNN-TL) that combines physics information with data-driven modeling to achieve fast and efficient prediction of combustion processes. A full-scale 3D transient computational fluid dynamics (CFD) model, rigorously validated with on-site data, serves as a high-credibility data generator. A novel phased knowledge transfer strategy is implemented: in the source domain, the BPNN is pre-trained using CFD simulation data to capture the underlying multiphasic coupling mechanisms; in the target domain, the model is fine-tuned using limited on-site data to bridge the simulation-reality gap and enhance generalization. Importantly, the BPNN-TL framework achieves superior prediction accuracy across all temperature points (RMSE<33.54 K, MAPE<3.85 %, and R2 > 0.983). Compared with conventional CFD, the BPNN-TL framework delivers predictions that align more closely with on-site data, while reducing the computational time per prediction by six orders of magnitude. Furthermore, the moisture content, volatiles, and load were identified as the most critical factors determining incineration performance by the mean influence value (MIV) analysis. This framework provides innovative tools for the prediction and optimization of MSWI processes, and offers new insights for clean energy utilization and environmental sustainability.