面向实时控制的排水系统深度LSTM神经网络模型  

Deep LSTM Neural Network Model for Real-time Control of Urban Drainage System

作　　者：杨萌祺 徐智伟 王一茗 曾思育[1,2] 杜鹏飞 董欣[1,2] YANG Meng-qi;XU Zhi-wei;WANG Yi-ming;ZENG Si-yu;DU Peng-fei;DONG Xin(School of Environment,Tsinghua University,Beijing 100084,China;State Key Joint Laboratory of Environment Simulation and Pollution Control,Beijing 100084,China)

机构地区：[1]清华大学环境学院,北京100084 [2]环境模拟与污染控制国家重点联合实验室,北京100084

出　　处：《中国给水排水》2023年第1期105-110,共6页China Water & Wastewater

基　　金：国家自然科学基金资助项目(51778327)。

摘　　要：如何得到兼顾运算时间和预测效果的排水系统预测模型是排水系统实时控制领域亟需解决的问题。针对这一难点,以非线性映射能力较强且运算速度较快的长短时记忆(LSTM)神经网络为基础,构建了面向实时控制的城市排水系统深度LSTM神经网络模型,并以苏州市福星片区为案例区域,验证该模型的预测效果和计算效率。结果显示,该模型对18个泵站站前液位预测结果的纳什效率系数均在0.5以上,且在不同降雨情景下均能得到较好的拟合结果;与机理模型相比,该模型能节约99.7%的计算时间,可显著提高排水系统预测模型的实时性。An urgent problem in the context of real-time control of drainage system is to establish a predicting model which balances operation time and prediction effect. To solve this problem, a deep long short term memory(LSTM) neural network model for real-time control of urban drainage system was constructed, which had strong nonlinear mapping ability and fast operation speed. The prediction performance and operation efficiency of the model were verified in Fuxing area of Suzhou City. The Nash-Sutcliffe efficiency coefficient of the prediction results of the water level in front of 18 pumping stations was above 0.5, and good fitting results were obtained under different rainfall scenarios. Compared with the mechanism model, the proposed model saved 99.7% of the operation time and significantly improved the real-time performance of the drainage system prediction model.

关 键 词：城市排水系统 实时控制 长短时记忆(LSTM)神经网络 深度学习 泵站站 

分 类 号：TU992[建筑科学—市政工程]