基于ISSA-BiLSTM的供水泵站流量预测

Flow prediction of water supply pumping stations based on ISSA-BiLSTM

摘要

为解决供水泵站流量预测精度不足、传统预测方法在处理非线性时间序列数据时存在的局限性问题，提出基于改进麻雀搜索算法（improved sparrow search algorithm，ISSA）优化双向长短期记忆网络（bidirectional long short-term memory，BiLSTM）的流量预测模型。采用 Logistic 混沌映射初始化和自适应 Levy 飞行策略改进麻雀搜索算法，以增强种群多样性并平衡算法的全局探索与局部开发能力；结合引入注意力机制的 BiLSTM 网络，利用 ISSA实现对隐藏层单元数、学习率及正则化系数等超参数的自动寻优。以某供水工程 19 个月的实际运行数据为例，基于皮尔逊相关性分析选取流量、压力、阀门开度等 13 个关键特征构建预测模型。结果表明：提出的 ISSA-BiLSTM 模型在均方根误差（root mean square error，E_RMS）、均绝对误差 (mean absolute error，E_MA)、平均绝对百分比误差（mean absolute percentage error，E_MAP）和决定系数（R²）上均表现最优，分别达到了 28.75 m³/h、14.03 m³/h、0.12% 和 0.89；与前沿深度学习模型、集成学习方法以及其他优化算法模型相比，该模型显著降低了预测误差，具有更强的鲁棒性与泛化能力。研究可为供水泵站的精细化管理与优化调度提供可靠的技术支撑。

Abstract

The optimal scheduling and energy-efficient operation of urban water supply pumping stations are critically dependent on the precise prediction of water flow. Accurate forecasting is a cornerstone of intelligent water management, allowing proactive control of pump operations to stabilize pressure within the distribution network, reduce energy consumption, and improve the overall safety and reliability of the water supply system. However, flow data from these stations typically show pronounced non-linearity, complex periodicities, and random fluctuations, which are influenced by a variety of interconnected factors such as consumer behavior, meteorological conditions, and network pressure. Modeling long-term dependencies has proven to be a major challenge for traditional forecasting techniques, such as statistical models and traditional machine learning algorithms. Furthermore, the performance of advanced deep learning models is frequently limited by the challenge of hyperparameter optimization, which frequently relies on empirical adjustments and is prone to suboptimal local results. To address the stated challenges, a novel hybrid deep learning framework was designed for water supply flow prediction, which integrated an improved sparrow search algorithm (ISSA) with a Bi-directional long short-term memory (BiLSTM) network incorporating an attention mechanism. First, a logistic chaotic map was used to initialize the sparrow population. This was one of the major changes made to the traditional sparrow search algorithm (SSA) to improve its optimization capability. By ensuring a more consistent and varied initial search agent distribution, this technique successfully keeps the algorithm from entering local optima in its early phases. Secondly, an adaptive Levy flight strategy, coupled with dynamic cosine weights, was implemented to balance the algorithm's global exploration and local exploitation capabilities throughout the iterative process. This enhanced ISSA was then utilized to systematically and automatically optimize the critical hyperparameters of the BiLSTM network, including the number of hidden units, learning rate, and regularization coefficients. A comprehensive dataset gathered over 19 months from a large-scale water supply pumping station was used to train and validate the model. Rigorous data preprocessing was performed, involving anomaly detection using an Isolation Forest algorithm and feature selection based on Pearson correlation analysis, which identified thirteen key operational features such as flow rates, pressure levels, and valve opening percentages. The suggested ISSA-BiLSTM model performed better, according to the empirical evaluation. The model achieved a root mean square error (E_RMS) of 28.75 m³/h, a mean absolute error (E_MA) of 14.03 m³/h, a mean absolute percentage error (E_MAP) of 0.12%, and a coefficient of determination (R ²) of 0.89 on the test dataset. A comprehensive comparative analysis was performed against a wide range of baselines, including advanced deep learning variants (Transformer, TCN), ensemble learning models (XGBoost, LightGBM), and other metaheuristic-optimized models (GA-BiLSTM, GWO-BiLSTM). The results revealed that the proposed model significantly outperformed these competitors, reducing the prediction error by approximately 62.7% compared to the standard LSTM model and showing statistically significant improvements over other optimized variants. The results confirm that the suggested framework offers a very reliable and accurate way to predict flow in water supply pumping stations, providing a substantial technical basis for the development of intelligent operational scheduling and fine-grained management in contemporary urban water supply systems.

关键词

改进型麻雀搜索算法 / 供水泵站 / 双向长短期记忆网络 / 流量预测 / 参数优化

Key words

improved sparrow search algorithm / water supply pumping station / bidirectional long short-term neural network / flow prediction / parameter optimization

参考文献

[1] Zakariazadeh A, Ahshan R, Abri R A, et al. Renewable energy integration in sustainable water systems: A review[J]. Cleaner Engineering and Technology, 2024, 18: 100722. DOI: 10.1016/j.clet.2024.100722.

[2] 景唤, 王永智, 宋志红, 等. 引江济淮工程（河南段）泵站系统长期服役风险评估[J]. 南水北调与水利科技（中英文）, 2023, 21(6): 1174-1183. DOI: 10.13476/j.cnki.nsbdqk.2023.0116.

[3] 张新, 陈致远, 袁静, 等. 基于泵阀联合调度的技术供水系统节能优化研究[J]. 水利水电技术（中英文）, 2024, 55(11): 87-97. DOI: 10.13928/j.cnki.wrahe.2024.11.008.

[4] 袁卓异, 李勇刚. 基于集成 LSTM 的泵站供水流量智能预测方法[J]. 湖南科技大学学报（自然科学版）, 2023, 38(1): 68-75. DOI: 10.13582/j.cnki.1672-9102.2023.01.009.

[5] 高怡然, 卢龙彬, 桑国庆, 等. 基于输入因子影响分析的 LSTM 梯级泵站站前水位实时预测模型[J]. 南水北调与水利科技（中英文）, 2025, 23(1): 157-165. DOI: 10.13476/j.cnki.nsbdqk.2025.0016.

[6] 杨利纳, 李文竹, 刘心. 基于灰色遗传 BP 神经网络的校园区间需水预测研究[J]. 水资源与水工程学报, 2019, 30(3): 133-138. DOI: 10.11705/j.issn.1672-643X.2019.03.20.

[7] Liu Chanjuan, Xu Jing, Li Xi'an, et al. Water resource forecasting with machine learning and deep learning: A scientometric analysis[J]. Artificial Intelligence in Geosciences, 2024, 5: 100084. DOI: 10.1016/j.aiig.2024.100084.

[8] Hasan M, Mifta Z, Papiya S J, et al. A state-of-the-art comparative review of load forecasting methods: Characteristics, perspectives, and applications[J]. Energy Conversion and Management: X, 2025, 26: 100922. DOI: 10.1016/j.ecmx.2025.100922.

[9] 欧阳文宇, 叶磊, 王梦云, 等. 深度学习水文预报研究进展综述Ⅰ: 常用模型与建模方法[J]. 南水北调与水利科技（中英文）, 2022, 20(4): 650-659. DOI: 10.13476/j.cnki.nsbdqk.2022.0067.

[10] 欧阳文宇, 叶磊, 顾学志, 等. 深度学习水文预报研究进展综述Ⅱ: 研究进展及展望[J]. 南水北调与水利科技（中英文）, 2022, 20(5): 862-875. DOI: 10.13476/j.cnki.nsbdqk.2022.0087.

[11] 代金汕, 葛恒军, 阚永庚, 等. 泵站单机组优化调度组合改进粒子群算法[J]. 南水北调与水利科技（中英文）, 2024, 22(5): 978-986. DOI: 10.13476/j.cnki.nsbdqk.2024.0097.

[12] Akay B, Karaboga D, Akay R. A comprehensive survey on optimizing deep learning models by metaheuristics[J]. Artificial Intelligence Review, 2022, 55(2): 829-894. DOI: 10.1007/s10462-021-09992-0.

[13] 袁建华, 谢斌斌, 何宝林, 等. 基于 DTW-VMD-PSO-BP 的光伏发电功率短期预测方法[J]. 太阳能学报, 2022, 43(8): 58-66. DOI: 10.19912/j.0254-0096.tynxb.2020-1405.

[14] 张薇薇, 赵平伟, 王景成. 基于长短时神经网络的城市需水量预测应用[J]. 净水技术, 2019, 38(增刊 1): 257-260. DOI: 10.15890/j.cnki.jsjs.2019.s1.064.

[15] 徐冬梅, 张一多, 王文川. 基于小波包分解的 LS-SVM-ARIMA 组合降水预测[J]. 南水北调与水利科技（中英文）, 2020, 18(6): 71-77. DOI: 10.13476/j.cnki.nsbdqk.2020.0116.

[16] 张伟, 邓彬彬, 仇建春, 等. 基于 ESMD-FE-AJSO-LSTM 算法的水闸深基坑变形预测[J]. 南水北调与水利科技（中英文）, 2024, 22(2): 378-387. DOI: 10.13476/j.cnki.nsbdqk.2024.0039.

[17] 赵菲菲, 张青青, 张宇, 等. 基于贝叶斯网络的黄河径流预测[J]. 南水北调与水利科技（中英文）, 2021, 19(3): 511-519. DOI: 10.13476/j.cnki.nsbdqk.2021.0054.

[18] Dong Wei, Yang Qiang. Data-driven solution for optimal pumping units scheduling of smart water conservancy[J]. IEEE Internet of Things Journal, 2020, 7(3): 1919-1926. DOI: 10.1109/JIOT.2019.2963250.

[19] 崔文浩. 基于流量预测的城市排污泵站调度系统的研究与实现[D]. 镇江： 江苏大学, 2024. DOI: 10.27170/d.cnki.gjsuu.2024.001376.

[20] Farzad A, Gulliver T A. Log message anomaly detection with fuzzy C-means and MLP [J]. Applied Intelligence, 2022, 52(15): 17708-17717. DOI: 10.1007/s10489-022-03300-1.

[21] 张可, 张政, 金伟. 数据-知识融合的水利工程建设安全风险灰色因子分解机预测模型[J]. 水利水电技术(中英文), 2024, 55(1): 134-143. DOI: 10.13928/j.cnki.wrahe.2024.01.012.

[22] Fu Guangtao, Jin Yiwen, Sun Siao, et al. The role of deep learning in urban water management: A critical review[J]. Water Research, 2022, 223: 118973. DOI: 10.1016/j.watres.2022.118973.

[23] Miao Jiawei, Tao Haicheng, Xie Haoran, et al. Reconstruction-based anomaly detection for multivariate time series using contrastive generative adversarial networks[J]. Information Processing & Management, 2024, 61(1): 103569. DOI: 10.1016/j.ipm.2023.103569.

[24] Pang Guansong, Shen Chunhua, Hengel A V D. Deep anomaly detection with deviation networks[C]// Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. New York, USA: Association for Computing Machinery, 2019: 353-362. DOI: 10.1145/3292500.3330871.

[25] 谢小鹏, 王银河, 熊鹏宇, 等. 基于贝叶斯优化 BiLSTM 的锂电池 SOH 预测[J]. 电源技术, 2025, 49(5): 1013-1018. DOI: 10.3969/j.issn.1002-087X.2025.05.017.

[26] Wei Qing, Yang Ju, Fu Fangbing, et al. Dynamic classification and attention mechanism-based bidirectional long short-term memory network for daily runoff prediction in Aksu River basin, northwest China[J]. Journal of Environmental Management, 2025, 374: 124121. DOI: 10.1016/j.jenvman.2025.124121.

[27] Lin Peijie, Zhang Xiangxin, Gong Longcong, et al. Multi-timescale short-term urban water demand forecasting based on an improved PatchTST model[J]. Journal of Hydrology, 2025, 651: 132599. DOI: 10.1016/j.jhydrol.2024.132599.

[28] 汪昌港, 耿峻, 赵鹏, 等. 基于模糊聚类和孤立森林算法的大坝监测数据粗差识别方法[J]. 水力发电, 2025, 51(12): 96-100.

[29] 陈嘉兴, 刘扬, 刘晓茜, 等. Logistic 混沌映射与差分进化改进人工蜂群优化水下定位[J]. 工程科学与技术, 2025, 57(1): 57-67. DOI: 10.12454/j.jsuese.202300763.

[30] 程翠娜, 奉松绿, 莫礼平. 基于正余弦优化算子和 Levy 飞行机制的和声搜索算法[J]. 数据采集与处理, 2023, 38(3): 690-703. DOI: 10.16337/j.1004-9037.2023.03.017.

[31] 陈湛峰, 李晓芳. 基于注意力机制优化的 BiLSTM 珠江口水质预测模型[J]. 环境科学, 2024, 45(6): 3205-3213. DOI: 10.13227/j.hjkx.202306024.

[32] Hu Kui, Cheng Yiwei, Wu Jun, et al. Deep bidirectional recurrent neural networks ensemble for remaining useful life prediction of aircraft engine[J]. IEEE Transactions on Cybernetics, 2023, 53(4): 2531-2543. DOI: 10.1109/TCYB.2021.3124838.

[33] Chang Fi John, Chen Pin An, Lu Ying Ray, et al. Real-time multi-step-ahead water level forecasting by recurrent neural networks for urban flood control[J]. Journal of Hydrology, 2014, 517: 836-846. DOI: 10.1016/j.jhydrol.2014.06.013.

[34] Xue Jiankai, Shen Bo. A novel swarm intelligence optimization approach: Sparrow search algorithm[J]. Systems Science & Control Engineering, 2020, 8(1): 22-34. DOI: 10.1080/21642583.2019.1708830.

[35] 付华, 刘昊. 多策略融合的改进麻雀搜索算法及其应用[J]. 控制与决策, 2021, 37(1): 87-96. DOI: 10.13195/j.kzyjc.2021.0582.

[36] Chen Hui, Wang Xu, Ge Bin, et al. A multi-strategy improved sparrow search algorithm for coverage optimization in a WSN[J]. Sensors (Basel, Switzerland), 2023, 23(8): 4124. DOI: 10.3390/s23084124.

[37] Zhou Jiangguo, Xu Zhongtian, Wang Shiguo. A novel hybrid learning paradigm with feature extraction for carbon price prediction based on Bi-directional long short-term memory network optimized by an improved sparrow search algorithm[J]. Environmental Science and Pollution Research International, 2022, 29(43): 65585-65598. DOI: 10.1007/s11356-022-20450-4.

[38] Fei Teng, Wang Hongjun, Liu Lanxue, et al. Research on multi-strategy improved sparrow search optimization algorithm[J]. Mathematical Biosciences and Engineering: MBE, 2023, 20(9): 17220-17241. DOI: 10.3934/mbe.2023767.

[39] Zhou Shengwen, Guo Shunsheng, Du Baigang, et al. A hybrid framework for multivariate time series forecasting of daily urban water demand using attention-based convolutional neural network and long short-term memory network[J]. Sustainability, 2022, 14(17): 11086. DOI: 10.3390/su141711086.

[40] Huang Zhenhan, Tanaka F. MSPM: A modularized and scalable multi-agent reinforcement learning-based system for financial portfolio management[J]. PLoS ONE, 2022, 17(2): e0263689. DOI: 10.1371/journal.pone.0263689.

基金资助

国家自然科学基金面上基金项目(52379087)

重庆市水利科技项目(CQSLK-2024013)