Research And System Integration Of Medium And Long Term Runoff Prediction Model Based On Machine Learning

Water resources are the source of life and the foundation of ecology.Medium-and long-term runoff forecasting is one of the important links in the overall planning of water resources.However,due to the influence of complex factors such as climate change and human activities,the runoff process has become more complex,which has brought great challenges to hydrological forecasting.In order to realize the efficient and reasonable use of water resources,it is necessary to continuously improve the medium and long-term forecasting methods to better plan the water resources.This paper uses Yichang Station as a research station to reveal the characteristics of monthly runoff and uses conventional prediction models: SARIMA model,SVR model,BP neural network,LSTM model for prediction.Furthermore,a modern statistical method is proposed: BOA-EEMD-LSTM model based on the Bayesian optimization principle for prediction,and the effectiveness of the proposed prediction model in medium and long-term runoff forecasting is proved by comparison.Based on the above research,design and develop a mid-and long-term runoff forecasting system.The main results of the research work are as follows:(1)Taking the Yichang Station as the main research object,analyze the characteristics of monthly runoff time series.By using the Spaearman's method,the Kendall rank correlation method,the Mann-Kendall trend test method,etc.,it is analyzed that the runoff of the hydrological station has no obvious trend over the years,and the overall change trend of the runoff over the years is not significant.Using MK test,ordered clustering method,sliding t test method and other methods,the significant mutation points are found in January1981 and December 2004.The wavelet analysis method is used to draw the conclusion that there is a first major cycle of 18 months of monthly runoff over the years and a second major cycle of 9 months.So as to provide scientific basis and data support for nonlinear runoff forecasting.(2)In order to analyze the application effect of the conventional mid-and long-term runoff prediction model in the study area,The SARIMA model,SVR model,BP neural network,and long short-term memory network were used to predict the monthly runoff process of Yichang Station.The application effects of each model and the advantages and disadvantages of each model were compared and analyzed.The results show that the certainty coefficient of SARIMA and LSTM prediction results can reach above 0.8,and the accuracy is relatively good,which can provide a reference for the mid-and long-term runoff prediction of Yichang Station,but the prediction accuracy can further improved.(3)At present,in the process of mid-and long-term runoff forecasting models,there are still problems such as strong subjectivity of model parameter selection and natural runoff time series noise redundancy.With the goal of improving the accuracy,this paper further proposes a modern statistical method based on the above-mentioned LSTM model method:based on Bayesian optimization algorithm(BOA)and integrated empirical mode decomposition(EEMD)combined with machine learning Medium and long-term runoff prediction model(BOA-EEMD-LSTM).The model predicts runoff according to the idea of "optimization-decomposition-prediction-synthesis".Aiming at the problem of timeconsuming and laborious selection of hyperparameters in machine learning,this paper uses a Bayesian hyperparameter optimization algorithm based on Gaussian process,which can realize rapid and automatic optimization of model hyperparameters.At the same time,this paper uses the EEMD algorithm to decompose the originally mixed runoff time series into single-component sub-sequences,so that the fitting performance of the prediction algorithm can be further exerted,and the algorithm's ability to predict runoff extremes is improved.Finally,the model proposed in this chapter is compared with the composite model EEMDBP,EMD-LSTM composite model and the conventional prediction model in Chapter 3,and the results show that the BOA-EEMD-LSTM model fits the runoff process curve well,and the forecast level can reach Grade A,which proves the validity of the BOA-EEMD-LSTM model.(4)In order to enhance the engineering practical value of the medium and long-term runoff forecast model,promote the digitalization,automation and informationization of hydrological forecasting operations.This paper develops a set of mid-and long-term runoff forecasting system,the system adopts the B/S architecture,the development model of frontend separation,and the front-end adopts the lightweight framework Vue.js.The back-end module is developed using the springboot framework based on java language.Effectively improve the development efficiency,reduce the complexity of system installation and deployment,enhance the scalability and portability of the system,and provide a good human-computer interaction page for engineering developers.