PDF下载

[1]杨利刚,郭帅,申屠华斌.基于深度学习的污水检查井液位峰值预测方法[J].中国给水排水,2025,41(19):163-168.
　YANGLi-gang,GUOShuai,SHENTUHua-bin.Prediction Method of Peak Liquid Level in Sewage Manhole Based on Deep Learning[J].China Water & Wastewater,2025,41(19):163-168.

点击复制

基于深度学习的污水检查井液位峰值预测方法

中国给水排水[ISSN:1000-4062/CN:12-1073/TU] 卷: 第41卷期数: 2025年第19期页码: 163-168 栏目: 出版日期: 2025-10-01

Title:: Prediction Method of Peak Liquid Level in Sewage Manhole Based on Deep Learning

作者:: 杨利刚¹，郭帅¹，申屠华斌²; （1.合肥工业大学土木与水利工程学院，安徽合肥 230009；2.中国电建集团华东勘测设计研究院有限公司，浙江杭州 311122）

Author(s):: YANG　Li-gang¹， GUO　Shuai¹， SHENTU　Hua-bin²; （1. School of Civil and Hydraulic Engineering， Hefei University of Technology， Hefei 230009， China； 2. Power China Huadong Engineering Corporation Limited， Hangzhou 311122， China）

关键词:: 污水检查井; 液位峰值预测; 深度学习; 长短期记忆（LSTM）神经网络; 时间序列预测模型

Keywords:: sewage manhole; peak liquid level prediction; deep learning; long short?term memory (LSTM) neural network; time series prediction model

摘要:: 污水检查井液位峰值预测可为提前发现排水管网潜在问题并采取预防措施以避免污水溢流提供重要的依据。相对于传统时序预测方法，深度学习具有更强大的拟合能力和特征学习能力。基于W市某区布设的4个在线液位计2023年1月—12月的每小时液位峰值数据，分别搭建基于随机森林（RF）、BP神经网络和长短期记忆（LSTM）神经网络3种方法的液位峰值预测模型，并采用均方根误差（RMSE）、平均绝对误差（MAE）、决定系数（R²）等指标评价不同预测模型的性能表现。结果显示，基于同一数据集所建立的3种预测模型的R2仅存在细微差异且均接近于1，3种预测方法均表现出良好的拟合程度；LSTM神经网络模型的MAE和RMSE值均最小，表明在预测误差和稳定性方面该模型优于另两种模型，LSTM神经网络模型具有更优的泛化能力。

Abstract:: The prediction of peak liquid levels in sewage manholes can serve as a critical reference for the early identification of potential issues within drainage networks, enabling proactive measures to prevent sewage overflow. Compared with traditional time series prediction methods, deep learning demonstrates superior capabilities in both function fitting and feature learning. Based on the hourly peak liquid level data collected from four on-line liquid level meters installed in a district of city W between January and December 2023, peak liquid level prediction models were developed using three different methodologies: random forest (RF), back propagation (BP) neural network, and long short-term memory (LSTM) neural network. The performance of these models was evaluated using key evaluation metrics, including root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R²). The R² values of the three prediction models, established based on the same dataset, were very similar and all approach 1, indicating a high degree of model fit for all three predictive methods. The MAE and RMSE values of the LSTM neural network model were the lowest among the three models, indicating that it outperformed the other two in terms of prediction accuracy and stability. Furthermore, the LSTM model demonstrated superior generalization capability.

更新日期/Last Update: 2025-10-01