净水技术2025，Vol.44Issue(9)：175-185,11.DOI:10.15890/j.cnki.jsjs.2025.09.022
基于Transformer及其耦合模型的分钟级用水量预测新方法
New Method for Minute-Level Water Consumption Prediction Based on Transformer and the Coupling Models
摘要
Abstract
[Objective]Accurate short-term water consumption forecasting is a crucial prerequisite for optimized water supply scheduling and holds significant importance for achieving carbon emission reduction in the water supply industry.However,existing research predominantly focuses on hourly or daily-scale predictions,with limited in-depth exploration of minute-level water consumption forecasting.This study aims to develop a high-precision minute-level water consumption prediction model to enhance water supply scheduling efficiency and support low-carbon water management.[Methods]Based on smart water meter monitoring data from a residential community in Suzhou,this study selected water consumption time series with a 6-minute time step as the research subject.The transformer deep learning model was employed for prediction and comparatively analyzed with backpropagation neural network(BPNN)and long short-term memory(LSTM)networks.To further enhance prediction accuracy,the empirical mode decomposition(EMD)data preprocessing method was introduced to construct an EMD-Transformer hybrid prediction model.Evaluation metrics include the coefficient of determination(R2),mean absolute percentage error(MAPE),and root mean square error(RMSE).[Results]The transformer model achieved an R2 of 0.925,representing a 13.9%and 6.0%improvement over BPNN(R2=0.812)and LSTM(R2=0.873)respectively.Its MAPE reached 10.89%,showing significant reductions compared to BPNN(MAPE=16.24%)and LSTM(MAPE=12.75%).The EMD-Transformer hybrid model demonstrated further enhanced predictive performance,elevating R2 to 0.964 while reducing MAPE to 7.50%and decreasing RMSE by 23.6%.Analytical result indicated that EMD decomposition effectively extracted multiscale features from water consumption time series,enabling the model to better capture usage patterns across different temporal scales.[Conclusion]The transformer architecture demonstrates superior performance in minute-level water consumption prediction,with the EMD-Transformer hybrid model elevating prediction accuracy to practical application standards.This achievement provides reliable technical support for real-time optimization of water supply systems,while its method ological framework can be extended to load forecasting in other urban utility domains.The implementation of minute-level precision forecasting is projected to significantly reduce energy consumption in water supply systems,offering substantial practical value for promoting green and low-carbon development in the water industry.关键词
分钟级用水量预测/长短期记忆网络(LSTM)/Transformer模型/EMD-Transformer耦合模型/预测性能Key words
minute-level water consumption prediction/long short-term memory(LSTM)/Transformer model/EMD-Transformer coupling model/predicted performance分类
建筑与水利引用本文复制引用
刘康乐,林涛,孙军益,张雪,沈月生..基于Transformer及其耦合模型的分钟级用水量预测新方法[J].净水技术,2025,44(9):175-185,11.基金项目
江苏省住房和城乡建设厅科技项目:基于优化调度和智能运控算法的供水节能关键技术研究与示范(2022ZD033) （2022ZD033）
