The Runoff Forecasting Of Heihe River Based On SARIMA And SVR Hybrid Model

The runoff forecasting is an important direction of hydrological research.It refers to the research and prediction of the runoff trends of rivers through the establishment of mathematical models.The forecast results can be widely used in flood control,drought prevention,environmental improvement,reservoir dispatch,hydropower station operation,shipping management,water resources allocation and management.However,due to various factors such as the weather system,basin surface,and human activities,the dynamics of the hydrological system have been strengthened,and its scale and complexity have increased.This poses a great challenge to the research.Runoff prediction methods can be divided into two categories: first,causal models,which are modeled according to the physical mechanisms of hydrological processes;second,data-driven models which do not consider the mechanism of runoff formation are modeled by mining the inherent changes in data.The causal model has great limitations.It does not only require detailed research on the hydrological process,but also require specific models under different conditions.It also needs high requirement for the data acquisition and research experience.Driven by the rapid development of artificial intelligence and machine learning technology,the data-driven model has become a research hotspot in recent years.It has the advantages of simplicity,good prediction accuracy,and wide applicability.This paper uses data-driven models to predict the runoff of the Heihe River.The hybrid model SARIMA-RandomForest-SVR with better accuracy is established in the end.Its goodness-of-fit is 0.832.Firstly,the time series analysis method was used to establish the SARIMA model to forecast the monthly runoff data.Then,in the process of residual processing,a mode of supervised learning is used.A total of 41 variables including lagged period values(lag time from 1 to 37),a sliding average of the residual values of the lag period,and a sliding standard deviation are used as candidate input variables.The current residual value is used as output variable in the establishment of support vector regression model.In order to improve the efficiency and accuracy of the model,random forest algorithm was used to select features.And then the candidate input variables were analyzed.Thus,20 factors with higher importance scores were selected as the final input variables of the model.Afterwards,a Support Vector Regression(SVR)model was built to fit the residual sequence.In this process,simulated annealing algorithm was used to perform parameter optimization.Finally,the predicted values of the residual were integrated with the predicted values of the time series model to obtain the final forecast result.Compared with the single model and other hybrid models,we can find that the hybrid model SARIMA-RandomForest-SVR established in this paper can make full use of the advantages of each single algorithm to achieve a good prediction effect.