An Innovative Method for Improving Spatial Resolution in Near-Term Precipitation Predictions

Due to the influence of marine climate, precipitation within a short-term range exhibits complex spatial distribution and rapid temporal changes. Current prediction models have significant errors in forecasting substantial changes in precipitation and require high training time costs. To address this, this paper proposes a new Graph Convolutional Neural Network(GCNN) model for capturing global spatial features. Firstly, given the strong local spatial characteristics of short-term precipitation, we construct a basic prediction model using the concept of graph convolutional neural networks to ensure effective extraction of spatial correlations between signals, thereby improving the spatial resolution of the prediction model. Additionally, an Axial Attention Mechanism(AAM) replaces the traditional attention module to enhance the model's ability to capture global features while reducing computational complexity. Finally, we conducted multiple random predictions on a short-term precipitation dataset for a city in China. Experimental results show that compared to traditional prediction methods, our improved model increased the Mean Absolute Error(MAE) and Root Mean Square Error(RMSE) by 33% and 37% respectively, and also reduced the training time compared to traditional models. For the same dataset, the new model saved nearly 30 minutes in training time. Our method can enhance the application capabilities of meteorological information in intelligent decision-making in relevant fields.