Climate prediction serves as a pivotal technological approach to addressing global climate change challenges and safeguarding socio-economic security. Accelerated industrialization and surging greenhouse gas emissions have led to a marked increase in extreme weather events, exposing the nonlinear and multi-scale coupled characteristics of the global climate system. Its inherent complexity and uncertainty pose significant challenges to prediction accuracy. For instance, in rapidly urbanizing regions like Beijing, the interaction between the urban heat island effect and the East Asian monsoon intensifies spatiotemporal uncertainties in local precipitation patterns. Devastating events such as the July 2021 "7·21" extreme rainfall and the 2023 Beijing-Tianjin-Hebei floods caused substantial economic losses, underscoring the urgent need for precise climate prediction. Concurrently, machine learning and neural network technologies have emerged as transformative paradigms for climate modeling. Their strength lies in extracting nonlinear relationships from vast historical observations and reanalysis data, such as the intricate linkages between satellite cloud imagery, ocean temperature fields, and precipitation dynamics. However, purely data-driven models are prone to sample bias and lack physical consistency constraints, making them inadequate for explaining the dynamical mechanisms behind extreme events like abrupt typhoon track shifts. Consequently, multi-model integration methods that synergize physical model mechanisms with data-driven advantages have become a research frontier.
To address these gaps, this study proposes a ​Multi-Model Fusion framework based on Model Averaging (MMF-MA), which harmonizes the strengths of machine learning, neural networks, and physics-based models. By embedding atmospheric physical equations as constraints into the loss function to minimize prediction errors, the framework dynamically calculates model-specific weights through cross-validation and integrates estimates from diverse models via weighted averaging. Unlike model selection strategies that rely on a single "optimal" model, the model averaging approach leverages information from all candidate models more effectively, thereby enhancing prediction accuracy. Focused on rainfall prediction in Beijing—a region characterized by complex microclimates—the framework is further extended to typhoon forecasting. Real-data simulations demonstrate that the proposed method consistently outperforms existing approaches in generating more accurate estimates.

 

Climate prediction,Model Averaging,machine learning,Neural Networks