Liu Xin, Sang Xuefeng, Chang Jiaxuan, Zheng Yang
School of Water Conservancy, North China University of Water Resources and Electric Power, Zhengzhou, 450046 China.
Research Office for Water Resources Management, China Institute of Water Resources and Hydropower Research, Beijing, 100038 China.
Water Resour Manag (Dordr). 2021;35(12):4021-4041. doi: 10.1007/s11269-021-02927-y. Epub 2021 Sep 23.
The water supply in megacities can be affected by the living habits and population mobility, so the fluctuation degree of daily water supply data is acute, which presents a great challenge to the water demand prediction. This is because that non-stationarity of daily data can have a large influence on the generalization ability of models. In this study, the Hodrick-Prescott (HP) and wavelet transform (WT) methods were used to carry out decomposition of daily data to solve the non-stationarity problem. The bidirectional long short term memory (BLSTM), seasonal autoregressive integrated moving average (SARIMA) and Gaussian radial basis function neural network (GRBFNN) were developed to carry out prediction of different subseries. The ensemble learning was introduced to improve the generalization ability of models, and prediction interval was generated based on student's t-test to cope with the variation of water supply laws. This study method was applied to the daily water demand prediction in Shenzhen and cross-validation was performed. The results show that WT is superior to HP decomposition method, but maximum decomposition level of WT should not be set too high, otherwise the trend characteristics of subseries will be weakened. Although the corona virus disease 2019 (COVID-19) outbreak caused a variation in water supply laws, this variation is still within the prediction interval. The WT and coupling models accurately predict water demand and provide the optimal mean square error (0.17%), Nash-Sutcliffe efficiency (97.21%), mean relative error (0.1), mean absolute error (3.32%), and correlation coefficient (0.99).
特大城市的供水会受到居民生活习惯和人口流动性的影响，因此每日供水数据的波动程度剧烈，这给水需求预测带来了巨大挑战。这是因为每日数据的非平稳性会对模型的泛化能力产生很大影响。在本研究中，使用了霍德里克 - 普雷斯科特（HP）和小波变换（WT）方法对每日数据进行分解，以解决非平稳性问题。构建了双向长短期记忆（BLSTM）、季节性自回归积分移动平均（SARIMA）和高斯径向基函数神经网络（GRBFNN）来对不同子序列进行预测。引入集成学习以提高模型的泛化能力，并基于学生t检验生成预测区间，以应对供水规律的变化。将本研究方法应用于深圳的每日用水需求预测并进行了交叉验证。结果表明，小波变换优于HP分解方法，但小波变换的最大分解层数不宜设置过高，否则子序列的趋势特征会被削弱。尽管2019年冠状病毒病（COVID - 19）疫情导致供水规律发生了变化，但这种变化仍在预测区间内。小波变换及耦合模型准确地预测了用水需求，并提供了最优的均方误差（0.17%）、纳什 - 萨特克利夫效率（97.21%）、平均相对误差（0.1）、平均绝对误差（3.32%）和相关系数（0.99）。
