Reasearch On Guangxi Temperature Prediction Model Based On Neural Network

Traditional temperature prediction methods are becoming increasingly difficult to deal with the explosive growth of time series observation data and the complex and interrelated problems of multiple observation factors.Neural network technology can effectively use the correlation of input meteorological elements over time,but not all meteorological elements are highly correlated with changes in temperature.Too many low-correlation elements will cause a decrease in model prediction accuracy.Therefore,this dissertation focuses on the research on Guangxi temperature prediction model based on neural networks,mainly using ground observation meteorological element data,building different neural networks and improving models to predict temperatures in multiple cities in Guangxi.The main work is as follows:(1)In view of the massive multivariate meteorological time series data,an RF-GRU model combining random forest algorithm with feature selection characteristics and gated recurrent unit network suitable for time-series modelling is proposed.The model is used to predict the temperature of Nanning city for the next 14 days.The common evaluation indicators of regression models,including the absolute mean error and the root-mean-square error,are used to evaluate different models.The results show that the RF-GRU model outperforms BP neural network,LSTM and GRU models,which indicates that the RF-GRU model has strong temperature prediction ability.(2)In view of the problems that GRU networks have time-dependent data and unidirectional transmission methods that lead to insufficient predictive ability,an RF-BiGRU model is further proposed based on random forests and bidirectional gate recurrent unit networks.The average temperatures for the next 14 days in three representative cities of Guangxi,namely Nanning,Beihai and Guilin,are predicted using different models.The results show that the RF-BiGRU model performs better than RF-GRU,BP neural network and GRU models,which suggests that the RF-BiGRU model can solve the time-dependent problem better through a two-way propagation mechanism.(3)In view of the problem that the performance of the network model may be degraded due to the lack of rationality of artificially determined neural network hyperparameters,an improved PSO-RF-GRU model is proposed based on particle swarm optimization algorithm.The average temperatures for the next 14 days in Nanning,Beihai and Guilin are predicted using different models.The results show that the PSO-RF-GRU model outperforms BP neural network,GRU,RF-GRU and RF-BiGRU models in terms of absolute mean error and root-mean-square error.These indicate that the PSO-RF-GRU model has higher temperature prediction accuracy and stronger generalization ability.