SMA-LSTM模型在径流预报中的适用性研究

Application of Long Short-Term Memory Neural Network Optimized by Slime Mold Algorithm in Runoff Forecast

径流预报对于防洪、发电和生态调度等具有重要意义。以大渡河丹巴以上流域为研究区域,采用黏菌优化算法(SMA)对长短期记忆神经网络(LSTM)的隐藏层输出维度进行优化,构建SMA-LST M模型对未来10日径流过程进行预报,以探讨深度学习方法对流域径流预报的适用性。基于2012-2017年的日降雨量和日流量资料,构建了预见期为10天的逐日径流SMA-LSTM预报模型,以2018-2019年的资料进行模型验证;采用最大1日径流量相对误差和10日总径流量相对误差作为SMA-LSTM模型精度的评价指标,并与未优化的LSTM模型和新安江模型结果进行对比。结果表明:SMA-LSTM模型具有较高的模拟和预报精度,无论是在率定期还是验证期,两种指标均控制在±10%以内,且两种指标的绝对值平均都不超过7%;整体而言,SMA-LSTM模型精度更高,预报的径流过程与实测过程更为贴近。研究成果可供流域径流预报实际工作参考。

Runoff forecasting is of great significance for flood control,power generation and ecological regulation.In this paper,taking the basin above Danba of Dadu River as the research area,the slime mold optimization algorithm(SMA)is used to optimize the hidden layer output dimension of long-term and short-term memory neural network(LSTM),and the SMA-LSTM model is constructed to predict the runoff process in the next 10 days,so as to explore the applicability of deep learning method to watershed runoff prediction.Based on the daily rainfall and daily flow data from 2012 to 2017,a daily runoff SMA-LSTM prediction model with a forecast period of 10 days is constructed,which is verified by the data from 2018 to 2019;The maximum 1-day runoff relative error and 10-day total runoff relative error are used as the evaluation indexes of the accuracy of SMA-LSTM model,and compared with the results of non-optimized LSTM model and Xinxanjiang model.The results show that the SMA-LSTM model has high simulation and prediction accuracy.Both indexes are controlled within±10%in the periodic rate and verification period,and the average absolute value of the two indexes is not more than 7%;On the whole,SMA-LSTM model has higher accuracy,and the predicted runoff process is closer to the measured process.The research results can be used as a reference for the practical work of runoff prediction in the other areas.