关键词: 领域先验知识, 时空神经网络, Madden-Julian振荡预报, 预训练模型

Abstract: Aiming at the problem that current artificial neural network methods could not  accurately forecast the Madden-Julian oscillation (MJO), a seasonal climate phenomenon, we proposed a domain prior knowledge-based spatio-temporal neural network model. Firstly, we integrated domain prior knowledge into data preprocessing according to the characteristics of climate circulation data. Secondly,  a pretraining-finetuning architecture was adopted, model data from the subseasonal-to-seasonal scale were used for pretraining model, and we completed  finetuning by using the ERA5 reanalysis data. Finally, through spatio-temporal modeling, a framework  combining convolutional neural networks and long short-term memory networks  was selected to embed the prior knowledge into the pretraining process and optimize the forecasting. Experimental results show that the proposed model can achieve accurate MJO forecasting for  23 d, and its performance is superior to other artificial neural network methods and domestic numerical forecasting methods.

Key words: domain prior knowledge, spatio-temporal neural network, Madden-Julian oscillation forecasting, pretraining model