| Visibility time forecasting refers to the prediction of future visibility from historical weather data.As a crucial meteorological indicator,the accuracy of visibility time series forecasts directly affects urban road systems,maritime transportation,and residential life,among others.Therefore,accurate forecasting of visibility has long been an important practical issue in the field of meteorology.The main difficulties in current visibility time prediction tasks are the ability of long sequence time-series forecasting and the accuracy of low visibility forecasting.Current methods based on recurrent neural network approaches are limited by the model structure,which makes it difficult to obtain long-range dependencies on the time series;low visibility occurs less frequently than high visibility and has a smaller sample size,making forecasting more difficult,yet crucial for accurate prediction of low visibility weather.At the same time,the formation of visibility is greatly influenced by geographic location,topography and subsurface,and the prediction method needs to consider spatial location and environmental differences.In order to solve the above problems and improve the accuracy of visibility time series forecasting,the following research works are carried out in this paper:(1)To address the problem that recurrent neural networks are not effective in long time series forecasting,this paper proposes to use a selfattention mechanism for visibility time series forecasting,so that it can capture the complex nonlinear relationships between visibility and other meteorological features,improve the accuracy of long time forecasting,and introduce temporal features through temporal coding.(2)To solve the problem of low visibility prediction accuracy,this paper uses a multitasking framework to focus the model on low visibility samples to improve the model’s early warning capability for extreme weather such as low visibility,in response to the data imbalance phenomenon of very few low visibility categories in visibility data.(3)To address the differences between the subsurface and meteorological change patterns of different meteorological stations,this paper proposes a multi-site visibility time series forecasting framework based on the pre-training method of contrastive learning,which is based on the heterogeneity between different stations for contrastive learning to obtain high-dimensional spatial-temporal features with more complete spatial-temporal information,further improving the accuracy of visibility forecasting.In summary,this paper proposes a comprehensive solution for visibility time series forecasting based on the characteristics of visibility time series data,the imbalance of data class distribution and the differences among different sites,which effectively improves the accuracy of visibility forecasting and provides a better guarantee for sea,land and air traffic safety and social production. |
