Temperature Prediction Model Research Based On BiLSTM And GCN

Weather phenomena have an important impact on human production and socio-economic development that cannot be ignored.The rapid development of observation technology in recent years and the dramatic increase in the scale and dimensionality of weather data have posed new challenges to weather prediction.With the gradual maturity of deep learning methods,more and more deep learning models are used for temperature prediction.Based on meteorological time series data,this topic investigates the characteristic factors of temperature change,and by comparing and analyzing various time series prediction models,a temperature prediction model based on GCNBiLSTM is constructed,and temperature simulation prediction for Nanjing and other cities is carried out,and the performance is evaluated.The paper takes the land surface area of the Chinese region as the study area,and based on the meteorological time series data from meteorological stations,the following research work is mainly carried out:(1)Considering the influence of other meteorological characteristics on temperature change,the multidimensional time series of humidity,sunshine,precipitation,wind speed,maximum temperature,near earth pressure and time attribute are introduced to replace the single temperature time series as the experimental influence factor;At the same time,considering the historical meteorological data,the time dimension of the input data is expanded by means of sliding time window.The temperature prediction model based on DFN and LSTM deep learning network is constructed.Taking Nanjing,Jiangsu Province as an example,the temperature of each station is predicted,and the accuracy of the model is evaluated;(2)Considering the influence of surrounding areas on the temperature change in the study area,using the characteristics of GCN spatial feature extraction,a temperature prediction model is constructed based on GCN and BiLSTM network.Taking Nanjing,Jiangsu Province as an example,the temperature of each station is predicted and the accuracy is evaluated;(3)DFN,LSTM model and traditional Arima method model are compared with GCN-BiLSTM model.Carry out the temperature prediction comparison experiment of each model in Nanjing in the next 14 days,and the temperature prediction comparison experiment of each model in Harbin,Beijing,Jinan,Qingdao,Nanjing,Hangzhou,Guangzhou and Shenzhen in the next day,compare the performance index of each model,and analyze the temporal and spatial differences of GCNBiLSTM model.The experimental results show that GCN BiLSTM performs best in comparison with other models.In the actual temperature prediction,the average RMSE is 1.96,which is 1.40,0.91 and1.01 lower than DFN,LSTM and ARIMA models respectively;The average MAE was 3.11,which was 0.99,0.59 and 1.47 lower than that of DFN,LSTM and ARIMA methods,respectively.It is found that the error of each model will increase with the extension of prediction time,among which the ARIMA model will rise sharply,while the change of GCN BiLSTM is relatively stable.At the same time,the experiment shows that there are spatial differences in GCN-BiLSTM,which has higher accuracy in areas with stable temperature fluctuation and small temperature difference.