Multi-site collaborative forecasting of regional visibility based on spatiotemporal convolutional network

Regional visibility forecasting encounters challenges due to data imbalance, temporal non-linearity and the consideration of multi-scale spatial factors. To tackle these challenges, this study introduces a novel approach for collaborative multi-site visibility forecasting based on spatiotemporal convolutional networks. Firstly, we preprocess the ERA5 reanalysis dataset and ground observation dataset, standardizing the spatiotemporal dimensions. We employ correlation coefficient analysis to select relevant meteorological factors. Subsequently, we create a spatiotemporal convolutional network model (TCN_GCN), which combines the power of temporal convolutional network (TCN) and graph convolutional network (GCN). Additionally, a weighted loss function is incorporated, accounting for the distribution of visibility values. The model is trained with multi-site data, enabling it to learn spatiotemporal visibility patterns across various sites. This empowers the model to generate multi-site visibility forecasts, thereby significantly improving regional visibility forecasting accuracy. Using 50 meteorological stations in Fujian Province, China, as a case study, we assess the model's predictions using key metrics such as mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R²). The experimental results demonstrate that the inclusion of both temporal and spatial features leads to a substantial enhancement in model prediction performance. The TCN_GCN model outperforms other deep learning methods in multi-site visibility forecasting, highlighting its effectiveness and superiority in improving regional visibility forecasting accuracy.