Climate change increasingly threatens global wheat production, affecting growth, yield stability, and quality, and thereby underscoring the need for accurate yield prediction to support climate-resilient breeding. To address this challenge, we developed a new integrated framework that couples climate data correction and reconstruction with multiple-crop-model yield prediction. First, a multivariable correction super-resolution diffusion model (MCSDM) was designed to correct and reconstruct daily meteorological data from multiple CMIP6 global climate models, significantly reducing error and improving spatial resolution at the regional scale. The corrected and reconstructed datasets were then used to drive three crop models (WOFOST, AquaCrop, DSSAT), which were calibrated and validated using field experiments on representative wheat germplasms in Zhejiang Province, China. Subsequently, a multiple-crop-model weight distribution network (MCMWDN) was employed to integrate individual crop model outputs, thereby enhancing prediction robustness and accuracy. The results showed that the MCSDM substantially reduced errors in temperature, precipitation, radiation, and wind speed variables and improved resolution, while the MCMWDN achieved an R² of 0.823 and reduced the root-mean-square error by over 40% compared to single-model predictions. Yield responses varied widely among germplasms: Norin 61 maintained high and stable yields, whereas Festin and Jagger/W94-244-132 performed poorly, and Lumai 5 and Zhoumai 31 remained relatively stable under moderate scenarios. Across climate pathways, yields are projected to slightly increase under SSP1-2.6, remain stable under SSP2-4.5, and decline significantly under SSP3-7.0 and SSP5-8.5, particularly during 2071–2100. Overall, this study provides a robust and generalizable prediction framework, offering valuable guidance for breeding climate-resilient wheat varieties and informing sustainable agricultural management under future climate change.