Research And Application Of Runoff Based On Ensemble Learning

The western region of my country has abundant water resources and a large number of water conservancy and hydropower projects.Most of the regions have agriculture,forestry and animal husbandry as the main economic industries.Runoff forecasting is very important for the regional economic development planning.Traditional runoff predictions are mostly based on mature hydrological cycle models,but there are many factors that affect runoff values,and there are complex relationships between these factors,and the prediction accuracy is highly dependent on other factors.In recent years,the more popular machine learning models rarely combine meteorological factors and other characteristics,and only make predictions based on the data rules of the runoff value itself,so there is still room for improvement in the prediction effect.Based on the above problems,this thesis mainly completes two parts,one is to propose a new runoff forecasting algorithm,and the other is to build a runoff forecasting platform based on the runoff forecasting algorithm.In the algorithm part,this thesis obtains the data of the influencing factors of the runoff value through various channels,and proposes the method of PCA and Prophet model prediction to construct the feature to obtain an effective feature sequence,which solves the problem of model overfitting caused by complex features and high collinearity.At the same time,an ensemble learning method based on the idea of Stacking is proposed,which uses LSTM,Prophet,and ridge regression model as the base class learner,and then uses the ridge regression model as the two-layer learner.It solves the problem that the traditional machine learning model has a poor fitting effect when the runoff value changes greatly during the wet season.In this thesis,the runoff prediction is carried out for Station A and Station B,which are densely populated and agriculturally developed in a certain watershed in the west,and use R2 and accuracy as evaluation indicators.From the results of the evaluation metrics,the ensemble learning method predicts well on both sites.Compared with the LSTM model that is currently widely used in time series data,in the prediction results of the integrated learning model for site A,the accuracy rate is increased by 3% when R2 is basically the same.In the prediction results for site B,the accuracy rate is flat.case,R2 improved by 2%.In the platform part,this thesis builds a runoff forecasting platform based on the runoff forecasting algorithm of ensemble learning and the related comparison algorithms involved in its experiments.According to the platform function structure,the platform is divided into three services: user service,algorithm service,and basic function service.The user service is responsible for information management and authority verification of users with different permissions,the algorithm service is responsible for completing the implementation of each algorithm,and the basic function service is responsible for the management of algorithm-related data.According to the technical architecture,the platform is divided into three layers: data layer,algorithm layer,and application layer.The data layer uses Mysql to store relational data and Redis to store cached data.The algorithm layer uses the Flask framework to encapsulate each algorithm and completes the communication through Http.The application layer completes the front-end work development based on the Vue framework,and completes the back-end development based on the Spring Boot framework.This platform takes the forecast of runoff value as the core work,and provides the query service of forecast value of runoff to the practitioners of agriculture,forestry and animal husbandry near the watershed site.At the same time,the algorithmrelated data involved in this thesis is also integrated into the platform to provide algorithm data query function,which can be obtained by researchers who need it.