基于DM-LSTM的城市降雨径流预测研究

Prediction of Urban Rainfall Runoff Based on DM-LSTM

智慧水务背景下,如何基于人工智能理论与技术深化城市降雨径流模型研究,是一项值得探索的课题。由于城市降雨径流时间分辨率高且样本特征分布不具有规律性,直接采用长短期记忆(LSTM)模型进行预测面临着挑战。基于此,提出用数据挖掘(DM)算法及规则对城市降雨径流时序数据集进行聚类和重构,并基于深度学习算法对LSTM模型的结构和参数进行优化,构建了DM-LSTM耦合模型,并用于研究区域的降雨径流模拟。结果表明,对于各类降雨事件,与LSTM模型相比,DM-LSTM耦合模型的均方根误差(RMSE)降低了2.1%~41.9%,纳什效率系数(NSE)提高了0.4%~56.4%,决定系数(R^(2))提高了0.3%~65.6%。DM-LSTM耦合模型不仅对各类降雨事件均表现出更好的预测性,而且模型运行时间仅为2.044 s,能够很好地满足城市降雨径流预测对实时性、准确性和稳定性的需求。

Under the background of smart water, how to deepen the research on urban rainfall runoff model based on artificial intelligence theory and technology is a topic worthy of exploration. Due to the high temporal resolution of urban rainfall runoff and irregular distribution of sample features, it is challenging to directly use long short term memory(LSTM) model for prediction. Based on this backround,data mining(DM) algorithm and rules were proposed to cluster and reconstruct the urban rainfall runoff time series data sets, the structure and parameters of LSTM model were optimized based on deep learning algorithm, and a DM-LSTM coupling model was constructed and applied to simulate rainfall runoff in the study area. For all kinds of rainfall events, compared with the LSTM model, the root mean square error(RMSE) of the DM-LSTM coupling model was decreased by 2.1%-41.9%, the Nash-Sutcliffe effciency(NSE) coefficient was increased by 0.4%-56.4%, and the coefficient of determination(R^(2)) was increased by 0.3%-65.6%. The DM-LSTM coupling model showed a better prediction performance for all kinds of rainfall events, and decreased the running time of the model to only 2.044 s, which could well meet the needs of real-time, accuracy and stability for urban rainfall runoff prediction.