Daily Runoff Forecasting Reseacher Based On Multiple Factors And Machine Learning

Predicting runoff is essential for efficient water management and mitigating flood damages,and it’s one of the fundamental research topics in hydrology.The advent of ML models has revolutionized runoff prediction by providing efficient,accurate,and reliable runoff forecasting based on straightforward inputs.At the initial stages of the technology,Most ML models adopted historical data as the primary forecasting inputs due to limited practical data and perfect forecasting solutions.Consequently,single-factor predictions models were popular.However,the development of technology and an increase in forecasting factors has paved the way for causationbased prediction models supported by empirical data.Models have since been able to incorporate new factors such as precipitation and temperature,which has significantly improved their accuracy.However,despite the models’ high accuracy,multi-factor models require further research to overcome the challenges of model training.The challenges relate to the unpredictability of random parameters and a lack of transparency of the prediction outcomes.In this regard,we constructed an ML model library using the daily runoff data from the Weihe and Hanjiang basins.We assessed the models’ usability based on their efficiency,generalization ability,and sensitivity.The main research achievements of this study are as follows:(1)Analyzed the changes in runoff,precipitation,temperature,humidity,and other climatic factors in watersheds regarding their statistical characteristics,trend,and mutation.Conducted a correlation analysis of hydro-meteorological elements,and completed building the model training samples.From this,it was found that: the inter-annual variation of runoff in the two basins was significant,the annual distribution was uneven,the multi-year change tended to be stable,and the runoff mutation in the Han River basin was significantly correlated with precipitation mutation.Precipitation,neighboring station runoff,and historical runoff had a high correlation with future runoff.(2)Compared the prediction accuracy of different prediction samples in the two basins and found that: each prediction sample achieved higher prediction accuracy in both basins.The impact of different driving factors on the prediction accuracy of different basins varied.In the Weihe River Huaxian Station,neighboring station runoff input significantly improved the model’s prediction accuracy,while the addition of the neighboring station runoff had limited impact on the model’s prediction accuracy in the Han River Shiquan Station,and the input of meteorological elements could improve the prediction accuracy.By comparing the prediction accuracy of different prediction models in the two basins,it was found that when the input information was limited,the prediction accuracy of different models was similar.With the depth-learning model and additional input data,the deep learning model which performed better in learning causal factors of hydrometeorology of the basin,was found to have higher predictive performance.(3)Evaluated the models’ usability in practical work from the perspectives of computational cost,generalization ability,and parameter sensitivity,and found that: deep learning models had stronger generalization capabilities than shallow learning models but require higher computational costs.All models were affected by model hyperparameters and sample lag period length,and parameter sensitivity greatly influenced the prediction accuracy and computational cost of each model iteration.The SVR model had higher parameter sensitivity,which needs to be determined before practical applications.(4)Conducted an attribution analysis of the flood prediction results of the model library at the Shiquan Station using the model’s explanatory methods and found that: the XGBoost,LSTM,and Seq2 Seq models focused more on input from neighboring stations and meteorological elements,and their predictions took into account some physical rules.The SVR model focused more on the autocorrelation statistical characteristics of historical runoff changes.Based on the attribution analysis of the XGBoost model’s prediction results on the entire runoff sequence at the Shiquan Station,it was found that the main driving factors were the runoff process at the Shiquan Station,the runoff process at the Hanzhong Station,and the precipitation at the Foping Station.Among these factors,in the dry and moderate water periods,the runoff process changed slowly,and the prediction results had a high correlation with the historical runoff,which played a dominant role in the prediction contribution.In contrast,in the flooding period,the runoff process was affected by neighboring station runoff and precipitation factors,and there was a joint effect among the driving factors.