Joint probability prediction of multi-site wind power based on multi-model collaborative heterogeneity

In the global drive toward the "carbon neutrality" goal, the large-scale deployment of renewable energy has become a key strategic pathway for driving the clean and low-carbon transformation of energy infrastructure. However, the output of wind turbines is not only affected by local meteorology, but also has nonlinear interactions with spatial factors such as the wake effect of surrounding units and complex terrain. The refined prediction framework integrating spatio-temporal correlation can effectively depict the multi-scale coupling mechanism of wind power systems and significantly enhance the multi-dimensional representation ability of the model, which is of crucial significance for research and application in this field. The wind power prediction system proposed in this study, utilizes differential calculation and spatial recursive segmentation strategies to achieve high-precision data interpolation and outlier elimination. At the optimization level, a five-dimensional and one-dimensional hybrid search paradigm is adopted, integrating the two-stage logic of exploration and development, and taking into account both search completeness and efficiency. When extracting features, the complex spatio-temporal correlation patterns are captured by combining the mapping transformation of the weight matrix with the filtering convolution. Furthermore, the loss function, combined with the weighted absolute error strategy, optimizes the configuration of bandwidth and kernel function to achieve deterministic prediction and uncertain quantitative analysis of wind power in multiple dimensions and multiple sites.