Temperature is one of the most critical meteorological elements in people’s daily lives and production, significantly impacting agriculture, transportation, and other sectors. Error correction in multisite temperature predictions is crucial for the accuracy of temperature forecasting. While existing methods for temperature prediction revision have made progress, they still face challenges related to uneven and imprecise multisite predictions, and a failure to fully and effectively explore the correlation between sites. To address these issues, we propose a correlation disentangling and spatio-temporal cooperative optimizing network(CSTGCN). Specifically, by utilizing temporal cues, we decompose the homogeneous heterogeneous correlations of the input data to process different types of inter-variable relationships across multiple data sites. Analyze and explore the overall correlations among multisites in both temporal and spatial dimensions through the spatiotemporal attention mechanism, which process enhances the impact of crucial temporal and spatial nodes to stabilize the overall prediction accuracy. And mine the temporal and spatial local relationships among multisite data through spatiotemporal map convolution, further deepening the model’s comprehension of geographic information and temporal correlation and enhancing the alignment of predictions with actual climate change trends. Extensive experimental results demonstrate the effectiveness of our approach compared to existing methods.