Research On Seasonal Prediction Of Winter Temperature In North China Based On Deep Learning

In this thesis,we construct North China winter temperature(NCDT)seasonal prediction models based on convolutional neural network(CNN),convolutional long and short-term memory network(Conv LSTM),and graph convolutional network(GCN)in order to predict the winter temperature(NCDT)in North China.The CNN-based model(CNNM)is trained on historical simulation of six CMIP5 models during 1852-1991 and tested on ERA5 data during1995-2017.Prediction of NCDT time series preliminary shows the potential of deep learning to improve the NCDT prediction skill,and the interpretability analysis of CNNM proves that CNNM can find physically meaningful relationships in data.Based on GCN,a GCN-based model(GCNM)with the structure of Encoder-Coupler-Decoder is further constructed for NCDT spatio-temporal seasonal prediction.GCNM is trained on historical simulations of eight CMIP5 models during 1854-1980 and tested on ERA5 data during 1982-2021.The main conclusions are presented as follows:(1)CNNM trained on poor quality CMIP5 data shows strong generalization ability on ERA5 data: CNNM is able to improve NCDT seasonal prediction skills,the average prediction skills during 1-6 months in adavance(0.57)improved by 50% compared with SEAS5.(2)Interpretability analysis of CNNM is carried out,and the mechanism of NCDT change is discussed from the perspective of deep learning.CNNM shows that SST anomalies in the Northeast Pacific in summer is the indicators of NCDT changes.It is verified by numerical model that SST anomalies in the northeast Pacific in summer have impacts on NCDT,and the mechanism may be as follows: In summer,the warming of the northeast Pacific causes westerly anomalies in the western and central tropical Pacific and last until December.The westerly anomalies causes the warming of the eastern Pacific in winter,and then induces sinking motion in the western Pacific and then induces anticyclonic circulation.The warm and wet southerly winds in the western part of the circulation leads abnormally high of NCDT.This indicates that the connection between data established by CNNM is of physical significance.(3)The interdecadal variation of GCNM prediction skills may be attributed to the interdecadal variation of the relationship between NCDT and its' impact factors;and the GCNM prediction error distribution is shaped by the distribution of temperature variability.SST is an important precursor signal for seasonal forecasting,and GCNM prediction skills decreased by14.02% on average after removing SST.It is reasonable to use GCN to approximate the interaction of variables in GCNM,and the attention mechanism introduced realizes the weighting of time and variables,which improves the physical properties of the model.