Research On Runoff Forecast Of Jinghe River Basin Based On Machine Learning Hybrid Model

Efficient and accurate runoff prediction is an important basis for basin water resources planning,flood control,and drought resistance.The formation of runoff is influenced by various factors such as meteorological conditions,land use,and topography,making the runoff data present highly complex nonlinear characteristics,which increases the difficulty of accurately predicting runoff.The emergence of machine learning models provides a feasible solution to this problem.In this paper,the monthly runoff time series from Zhangjiashan hydrological station in the Jinghe River Basin from 1984 to 2017,a total of 34 years,were selected as the research object,and a machine learning-based runoff prediction coupling model was constructed,combined with factors such as meteorology and circulation.At the same time,time-frequency analysis methods and optimization algorithms were used to improve the model’s prediction accuracy.The main research contents and results are as follows.(1)From 1984 to 2017,the monthly runoff at Zhangjiashan station in the Jinghe River Basin showed a significant decreasing trend in all months except for October and December.Mann-Kendall abrupt change analysis showed that the runoff in January and May had abrupt changes in 1991,while the runoff in April,June,and November had abrupt changes in 1994,and the runoff in July and August had abrupt changes in 2005 and 2002,respectively.Morlet wavelet analysis showed that February,June,August,and October had 4 scale periods,while January,March,May,July,September,November,and December had 3 scale periods,and April had 2 scale periods.The study aims to construct a coupled model for runoff prediction based on machine learning,combined with meteorological and circulation factors,and to improve the prediction accuracy of the model by using time-frequency analysis methods and optimization algorithms.(2)To address the issue of low prediction accuracy caused by using only streamflow as a single-factor input in the prediction model,the Pearson correlation coefficient method was used to select rainfall with lag times of 0 and 1 month,and temperature with lag times of0 and 1 month,which have higher correlation with streamflow,as the prepared input variables.The random forest method was used to rank the importance of atmospheric circulation factors that affect streamflow,and the top 4 factors,EASM,SOI,WPSHI,and NAOI,were selected as prepared input variables for the prediction model.(3)To analyze the impact of input data on the prediction model,single-month runoff data and multiple-month runoff data with different meteorological and circulation factors added were used as input to predict monthly runoff using LSTM and Prophet models.The results show that the prediction accuracy of the multi-input model is significantly higher than that of the single runoff input model.Among them,the rainfall-temperature-atmospheric factor-LSTM coupling prediction model has the highest fitting degree R2(0.815),which is0.163 higher than the single runoff input LSTM prediction model;the fitting degree R2 of the rainfall-LSTM coupling prediction model is higher than that of the temperature-LSTM model,which is about 0.059 higher.Similar results were obtained in the Prophet model simulation,and the best prediction performance was obtained by the rainfall-temperature-atmospheric factor-Prophet model(0.846),followed by the rainfall-Prophet,temperature-Prophet,and Prophet models.Under the same input conditions,the Prophet coupling model performs better than the LSTM coupling model in terms of prediction performance.(4)Regarding the non-linear characteristics of the inflow input in the prediction model,two decomposition methods,CEEMD and DWT,were used to decompose the original inflow sequence into multiple sub-sequences as inflow input items,and different precipitation,temperature,and atmospheric circulation impact factor input items were added.Multiple coupled models were constructed based on two different machine learning methods,LSTM and Prophet.Among them,the fitting degree R2 of CEEMD-precipitation-temperatureatmosphere-LSTM is 0.897,and the fitting degree R2 of DWT-precipitation-temperatureatmosphere-LSTM is 0.861,both higher than the prediction model of the undecomposed inflow sequence.The fitting degree R2 of CEEMD-precipitation-temperature-atmosphere factor-Prophet model is 0.856,and the fitting degree R2 of DWT-precipitation-temperatureatmosphere factor-Prophet model is 0.906.The CEEMD decomposition method performs better in improving the prediction accuracy of the LSTM model,while the DWT decomposition method performs better in improving the prediction accuracy of the Prophet model.Furthermore,it was found that the prediction effect of the inflow could be significantly improved by decomposing it and incorporating it into the model.(5)To address the issue of selecting optimal hyperparameters in LSTM,whale optimization algorithm and Bayesian optimization algorithm were introduced.The results showed that both algorithms can improve the prediction accuracy,with R2 values of the models greater than 0.8.Among them,the Bayesian optimization algorithm had a better effect.The best-performing coupled model was CEEEMD-rainfall-temperature-atmosphere-BOLSTM,with an R2 value of 0.943,further improving the model’s prediction performance.