基于FFB-LSTM的供水计量区超短时水量预测方法研究  

A flow fusion-based long short-term memory neural network for ultra-short-term water demand forecasting in district metering areas

作　　者：王晓婷[1,2] 吴以朋 刘书明 吴雪[1] WANG Xiaoting;WU Yipeng;LIU Shuming;WU Xue(School of Environment,Tsinghua University,Beijing 100084,China;JD.com,Inc,Beijing 100176,China)

机构地区：[1]清华大学环境学院,北京100084 [2]京东集团,北京100176

出　　处：《给水排水》2023年第3期133-139,共7页Water & Wastewater Engineering

基　　金：国家自然科学基金(51879139)。

摘　　要：为实时监测管网运行状态、及时捕捉管线漏损,需开展供水计量区的超短时需水量预测。然而,供水区域和时间粒度的减小均会带来需水量数据波动性的增加,导致预测难度增大。在此背景下,以多维度水量融合提高信息利用度为学术思想,以长短时记忆神经网络(LSTM)为实现手段,提出基于多维度水量融合的LSTM预测算法(FFB-LSTM),预测我国南方某真实独立计量区的超短时需水量。与传统LSTM、ANN模型对比,结果表明,所提出的FFB-LSTM在MAPE、MSE、MAE三个指标上均优于传统模型,能够高精度的预测计量区超短时需水量,为供水行业计量区的超短时需水量预测工作提供了有效范例。Ultra-short-term water demand forecasting in district metering areas(DMA)plays an important role in water distribution systems'real-time operation and leaks detection.However,both a decrease in water supply area and a decrease in time granularity increase the volatility of the data and make forecasting more difficult.To solve the hard forecasting problem,this paper uses flow fusion to improve information utilization.A flow fusion-based long short-term memory neural network(FFB-LSTM)algorithm is developed using the LSTM model as the realization frame,by exploring the multi-scale flow fusion mechanism of the pipe network.To evaluate the forecasting performance,the proposed method was applied to a real-life DMA and compared with the LSTM and artificial neural network(ANN)model.The results show that the proposed method is superior to the traditional method in the three indicators of MAPE,MSE,and MAE,indicating that the proposed method can predict the ultra-short-term water demand in DMA with high accuracy.It provides an effective example for ultra-short-term water demand forecasting in DMAs of the water supply industry.

关 键 词：供水管网 分区计量 超短时需水量预测 深度学习 长短时记忆单元 

分 类 号：TU991[建筑科学—市政工程]