Data Driven Based Model Fusion Method For Monthly Rainfall Prediction

Rainfall Prediction is an important parameter for water resources management,early warning of urban waterlogging and hydrological analysis.Not only can it provide decision-making basis for agriculture,water conservancy and other relevant departments,but also can reduce unnecessary losses,which is of guiding significance.Because rainfall system is affected by many factors such as temperature,pressure and humidity,which leads to a highly non-stationary and non-linear feature of the monthly rainfall series,so it is very difficult to accurately forecast the monthly rainfall.Therefore,the accurate prediction of monthly rainfall is still a problem.In order to further improve the accuracy of monthly rainfall forecasting,this paper proposes a data driven based model fusion method for monthly rainfall prediction.the common single model and the method of this paper are validated through using the measured data from six stations.The related research results and conclusions are as follows:1)This paper introduces the Model discussion based on rainfall forecasting and studies the predictability of rainfall time series;Firstly,two categories of rainfall forecasting models are explained,namely,explanatory model and data-driven model.In general,it is more difficult to construct an accurate explanatory model,while the data-driven method does not consider the external influencing factors and only uses the historical data to establish a model that overcomes the shortcomings of the previous model.More complex system is more applicable.Secondly,the data-driven model commonly used in rainfall forecasting and the predictability of rainfall time series are introduced.Predictive analysis of the measured data from six regions can be drawn from the results of power spectrum and the largest Lyapunov exponent.The time series of these six regions all have obvious chaotic characteristics,which shows that the collected data in this paper have Predictive.2)Study on the commonly used rainfall forecasting model;This paper introduces three currently used forecasting models,respectively ARIMA model,BPNN model and LSSVM model,and introduces the theory and modeling step in detail.From the prediction results,the three models have a good grasp of the trend of time series,but have a poor ability to fit the mutation values.The fitting effect of ARIMA model is the worst,and the fitting effect of BPNN model and LSSVM model is similar.For highly complex nonlinear systems such as rainfall,these three models have lower prediction accuracy.3)In order to further improve the forecast accuracy of monthly rainfall,the characteristics of the combined model and the necessity of using the combined model in the complex rainfall system are analyzed emphatically.Then,this paper proposes a new method of monthly rainfall forecasting,which is model fusion.Considering the multi-scale characteristics of monthly rainfall,the monthly rainfall series are divided into monthly subscales and annual subscales.According to the different characteristics of these two subsequences,the monthly rainfall was selected by adaptive fuzzy neural network system(ANFIS)and gray model(EGM)respectively,and then the data of the two single models were fused by the gray relational method.The results of the example analysis show that the prediction accuracy of the monthly rainfall based on the model fusion is higher than that of the corresponding single model and the three common models.It proves that the proposed model fusion Can improve the prediction accuracy of rainfall.The combined model proposed in this paper not only provides a new efficient method for monthly rainfall forecast,but also provides a new way of thinking for its similar research.The prediction methods involved in this paper have compiled corresponding procedures and apply for software copyright,which laid a solid foundation for practical application in the future.