Research On Weather Forecasting Method Based On Spatio-temporal Self-supervision

Weather forecasting is one of the core tasks in the field of meteorology and is closely related to human production and life.It plays a huge role in many fields such as transportation forecasting,agricultural production,tourism,etc.,thus ensuring the quality and steady development of the economy and society.Therefore,weather forecasting has been a popular topic of research for scholars at home and abroad.At present,mainstream weather forecasting methods can be divided into two categories:physical model-driven methods(i.e.,numerical weather forecasting,which is widely used in practical meteorological operations)and data-driven methods.Physical model-driven methods require supercomputers to simulate and extrapolate the complex physical evolution of the atmospheric system,which consumes huge computational resources and requires specialized domain knowledge to fine-tune the initialization settings of the established numerical models.In contrast,most data-driven weather forecasting methods use deep learning models to learn potential patterns of variability from large amounts of historical weather data.Compared with physical model-driven methods,data-driven methods not only reduce the computational cost by several orders of magnitude,but also eliminate the need to manually build numerical models of the physics used to describe the atmospheric motion processes,and therefore,data-driven weather forecasting methods are gradually attracting more and more research interests from scholars.However,many data-driven models tend to ignore the correlation in spatial dimension,or simply consider all weather elements within a weather station as a whole without exploring the complex correlation within weather data from a more fine-grained perspective,while some other data-driven methods tend to further improve the accuracy of weather forecasting by building complex neural network models.To alleviate these defects,this thesis first proposes a spatio-temporal contrastive selfsupervised weather forecasting method(STWF)based on a classical encoder-decoder structure that captures the spatio-temporal dependencies of weather data from the macroscopic weather station level,which includes a spatial-based self-supervised pre-training module and a timebased self-supervised pre-training module.These two modules exploit the spatio-temporal similarities in weather data through hand-designed contrastive self-supervised auxiliary tasks,and ultimately the method achieves the joint prediction of multiple weather elements in multiple stations over a future period of time.Extensive experimental results on three real weather datasets validate the effectiveness of STWF and the feasibility of the two self-supervised auxiliary tasks.To further improve the prediction accuracy of the model,this thesis proposes a weather forecasting method incorporating generative contextual self-supervised learning(STCWF)based on the method STWF.A generative contextual self-supervised pre-training module is introduced in this method,and a manually designed weather feature reconstruction auxiliary task with high mask ratio is used to train it,so as to capture the spatio-temporal dependencies of weather data at the weather element level and realize the modeling of contextual relationships among weather elements.Finally,by combining the generative contextual self-supervised auxiliary task with two contrastive self-supervised auxiliary tasks in STWF,an uncomplicated network obtain strong capability to capture latent representations for weather changes with time-varying.Thereafter,an effective encoder-decoder based fine-tuning framework is proposed,consisting of three self-supervised pretrained encoders.Finally,this thesis conducts comprehensive experiments on three real-world weather datasets,and the experimental results demonstrate the superiority of STCWF in weather forecasting tasks,and also validate the importance and necessity of the newly introduced generative contextual self-supervised auxiliary task.