基于径流特性分解的月径流集成预测模型研究

Integrated Monthly Runoff Prediction Model Based on Runoff Characteristic Decomposition

揭示混沌径流序列中的规律特性可使预测径流的可解释性、精度大幅提升。针对中长期径流序列的周期性、趋势性特征,收集洪泽湖流域吴家渡站1959~2019年实测月径流资料,提取径流周期成分和趋势成分,依据各成分的径流特性,选取契合物理特性规律的极限梯度下降(XGBoost)预测模型进行趋势成分预测,选择善于捕捉混沌规律的长短期记忆神经网络(LSTM)进行残差成分预测,构建了一种基于径流特性分解的XGBoost-LSTM集成预测模型,采用该模型对洪泽湖流域吴家渡站月径流序列进行预测,并将预测结果与XGBoost、LSTM、随机森林、BP等单一预测模型进行比较。结果表明,基于特性成分提取的XGBoost-LSTM集成模型的预测精度高于单一径流预测模型,能够利用径流序列规律特性,充分发掘预测模型优势,有效提升径流预测精度。

Revealing the regular characteristics in the chaotic runoff sequence significantly enhances the interpretability and accuracy of predicting runoff.In addressing the periodic and trend features of medium to long-term runoff sequences,the observed monthly runoff data from the Wujiadu station in the Hongze Lake basin were collected during the years 1959 to 2019.Runoff periodic components and trend components were extracted.Based on the runoff characteristics of each component,the Extreme Gradient Boosting(XGBoost)prediction model,aligning with the rules of physical characteristics,was chosen for trend component prediction.The Long Short-Term Memory neural network(LSTM),known for its proficiency in capturing chaotic patterns,was selected for residual component prediction.A prediction model,integrating XGBoost and LSTM and based on runoff characteristic decomposition,was constructed.This model was employed to forecast monthly runoff sequences at the Wujiadu station in the Hongze Lake basin.The predicted results were compared with single prediction models such as XGBoost,LSTM,Random Forest,and BP.The results indicate that the predictive accuracy of the XGBoost-LSTM ensemble model,based on characteristic component extraction,surpasses that of single runoff prediction models.It can utilize the regular characteristics of runoff sequences,fully exploit the advantages of the prediction model,and effectively improve the accuracy of runoff prediction.