Deep-width Learning Model And Its Application In Precipitation Predictio

With the development of science,technology and social progress,higher-accuracy and ultra-stability of prediction models are the fundemantal requirements of various fields.In recent years,the Recurrent Neural Network(RNN)technology in deep learning,especially its combination architecture(Recurrent Neural Network-Convolutional Neural Network,RNNCNN)with another deep learning technology CNN,has shown strong advantages in prediction.However,RNN-based models always are tangled with time-lag problem in prediction tasks,even the popular CNN-RNN architecture does not solve the problem.Noting that the Broad Learning System(BLS)with the advantages of direct weight calculation,fast and efficient operation,by integrating the advantages of broad learning and deep learning,the new deep-broad prediction architecture is proposed in this paper.Through applying the new models to precipitation forecast problem,validity of the new models are verified.The research contents and innovations of this paper are listed as follows:1.Combining the advantages of broad learning and deep learning,the deep-broad prediction architecture is introduced.For the single-factor prediction problem,the RNN-BLS prediction model is given;furthermore,taking into account that impacts of outliers and other factors on the prediction results are ignored in the single-factor prediction model,with the help of Weighted Broad Learning System(WBLS),which can weaken the the predominance of the adverse effects of outliers,the RNN-WBLS multivariate prediction model is established.2.On the basis of anaylsis of the characteristics of the monthly precipitation series data,taking advantages of the long-term memory function of Long Short-Term Memory Network(LSTM)in the RNN series with noise-eliminating function of the Complete Ensemble Empirical Mode Decomposition(CEEMD),the single-factor monthly precipitation forecast model based on CEEMD-LSTM-BLS is presented.Five representative stations with different geographical and precipitation characteristics in Hubei Province are selected for case analysis.Comparing with multiple predictive models,the results show that for all selected stations and different time steps,the new model performs best on all the evaluation indicators selected.Even in contrast with the state-of-the-art CNN-LSTM arichtecture,the new model reduces the Root Mean Square Error(RMSE)and Mean Absolute Error(MAE)by about 22.97% and24.15%,respectively,and increases the Decisive Factor(R2＿score)by about 8.43 %.In particular,the new model not only solves the time-lag problem,but also significantly improves the fitting effect of precipitation data for heavy precipitation and dry months in the series.In terms of computational efficiency,LSTM-BLS does not decrease compared with LSTM.3.According to the characteristics of daily precipitation series data,six factors-air pressure,temperature,humidity,wind speed,sunshine and precipitation are selected as the input of the model,and the LSTM-WBLS multi-factor daily precipitation forecast model is established.Case study demonstrates that LSTM-WBLS performs best on all selected evaluation metrics.Compared with the popular CNN-LSTM,LSTM-WBLS reduces about34.69% and 38.28% on RMSE and MAE,respectively,and increases about 6.98% on R2＿score.Particulally,the new model not only greatly improves the fitting degree,but also solves the hysteresis problem.In terms of computational efficiency,LSTM-WBLS does not decrease compared with LSTM.