基于LSTM的城市供水量时间序列预测研究

Research on Urban Water Supply Forecasting Based on LSTM

摘要: 近年来，城市化进程的持续推进与人口规模的不断扩张，使得城市供水系统的稳定运行面临日益严峻的挑战，精准的供水量预测因此成为城市供水管理中的核心问题之一。传统基于统计学的预测方法在处理供水量时间序列时，往往难以有效刻画其内在的复杂非线性特征，这在一定程度上限制了预测精度的进一步提升。针对这一问题，本文提出一种基于长短期记忆网络(Long Short-Term Memory, LSTM)的城市供水量预测方法。研究以某省会城市2019年度供水量监测数据为基础，通过数据清洗、特征构建与样本集划分等预处理步骤，构建适用于时序建模的数据结构，并增加双向长短期记忆网络(BiLSTM)进行对比实验。为验证所提方法的有效性，将LSTM、BiLSTM模型与线性回归(Linear Regression, LR)、支持向量机(Support Vector Machine, SVM)和随机森林(Random Forest, RF)等传统机器学习模型进行对比分析，并采用平均绝对百分比误差(MAPE)、均方根误差(RMSE)作为评价指标。实验结果显示，LSTM模型在测试集上的MAPE为2.62%、平均绝对误差为0.35 × 10⁴ m³/d，BiLSTM模型MAPE为2.28%、平均绝对误差为0.30 × 10⁴ m³/d，预测性能显著优于对比模型，能够有效捕捉城市供水量序列中的时序依赖特征。研究结果为城市供水系统的科学调度与智能管理提供可靠的数据支持与方法参考。

Abstract: In recent years, the continuous advancement of urbanization and the constant expansion of population size have posed increasingly severe challenges to the stable operation of urban water supply systems. Accurate water supply forecasting has thus become one of the core issues in urban water management. Traditional statistical forecasting methods often struggle to effectively capture the intrinsic complex nonlinear characteristics when processing water supply time series, which to some extent limits further improvement in prediction accuracy. To address this problem, this paper proposes an urban water supply forecasting method based on Long Short-Term Memory (LSTM) network. Using daily water supply data from a provincial capital city in 2019, a data structure suitable for time series modeling was constructed through preprocessing steps including data cleaning, feature construction, and sample set partitioning, with Bidirectional LSTM (BiLSTM) added for comparative experiments. To validate the effectiveness of the proposed method, the LSTM and BiLSTM models were compared with traditional machine learning models including Linear Regression (LR), Support Vector Machine (SVM), and Random Forest (RF). Mean Absolute Percentage Error (MAPE) and Root Mean Squared Error (RMSE) were adopted as evaluation metrics. Experimental results show that the LSTM model achieves a MAPE of 2.62% and a mean absolute error of 0.35 × 10⁴ m³/d on the test set, while BiLSTM achieves a MAPE of 2.28% and a mean absolute error of 0.30 × 10⁴ m³/d, demonstrating significantly superior performance compared to the benchmark models, effectively capturing the temporal dependence characteristics within urban water supply sequences. The research findings provide reliable data support and methodological reference for the scientific scheduling and intelligent management of urban water supply systems.