基于LSTM模型的排水系统流量预测研究  

Flow Prediction of Drainage System Based on Long Short Time Memory Model

作　　者：李双宇 张明凯 刘艳臣[3] 施汉昌[2,3] LI Shuang-yu;ZHANG Ming-kai;LIU Yan-chen;SHI Han-chang(College of Engineering,Peking University Beijing 100871,China;Beijing Institute of Collaborative Innovation,Beijing 100094,China;School of Environment,Tsinghua University,Beijing 100084,China)

机构地区：[1]北京大学工学院,北京100871 [2]北京协同创新研究院,北京100094 [3]清华大学环境学院,北京100084

出　　处：《中国给水排水》2022年第5期59-64,共6页China Water & Wastewater

基　　金：国家水体污染控制与治理科技重大专项(2017ZX07103007)。

摘　　要：排水系统流量预测对于城市水安全、污水厂优化运行具有重要意义。与需要复杂建模和大量地理信息数据的传统水文水力学模型不同,机器学习可以通过数据驱动实现排水系统的流量预测预警。结合流量数据的时序性,分别在单变量(流量)、双变量(流量和降雨)的情况下,采用5种长短期记忆神经网络(LSTM)模型(Vanilla LSTM、Stacked LSTM、Bidirectional LSTM、CNN LSTM、ConV LSTM)对江苏省无锡市某污水处理厂的进水流量进行预测。结果表明,Bidirectional LSTM最优的实验参数条件是:隐藏层单元数为250,训练轮数为200,训练集样本数为250;在同等条件下,Bidirectional LSTM相较其他4种方法可以更有效地预测未来流量;相比仅输入流量变量,在增加降雨变量后,可以提升近20%的流量预测精度。Flow prediction of drainage systems is of great significance for urban water safety and optimal operation of wastewater treatment plants. Different from traditional hydrological models which need complex modeling and a large amount of geographic information data, machine learning can realize flow prediction and early warning of a drainage system through data driving. In combination with the time sequence of flow data, five long short time memory(LSTM) models(Vanilla LSTM, Stacked LSTM,Bidirectional LSTM, CNN LSTM and ConV LSTM) under the conditions of single variable(flow) and double variable(flow and rainfall) were applied to predict the inlet flow of a wastewater treatment plant in Wuxi City, Jiangsu Province. In the parameter selection experiment, the optimal parameter condition of Bidirectional LSTM was that the number of LSTM hidden layer units, training epochs and training set samples were 250, 200 and 250. Under the same condition, Bidirectional LSTM predicted the future flow more effectively than the other four methods. Compared with simulation with flow as the only variable, its accuracy of flow prediction was improved by nearly 20% after adding rainfall as another variable.

关 键 词：排水系统 LSTM模型 流量预测 时间序列 最优实验参数 

分 类 号：TU992[建筑科学—市政工程]