基于CNN-LSTM模型的黄河水质预测研究

王军 高梓勋 朱永明

摘 要：水质预测是水资源管理和水污染防治的基础性、前提性工作，但黄河流域水质预测研究相对滞后。为了改善LSTM水质预测模型的性能、提高其泛化能力，根据水质变化具有周期性和非线性的特征，以黄河小浪底水库溶解氧含量为研究对象，构建了一种卷积神经网络CNN和长短时记忆网络LSTM结合的CNN-LSTM预测模型，经试验验证，该模型可以高效地提取水质特征信息并进行时间序列预测，预测误差比LSTM模型的更低，其预测值的平均绝对误差和均方根误差分别比LSTM模型的低19.72%和10.44%，对较大值和较小值的预测更为准确，且具有较好的泛化性能。

关键词：水质预测;长短时记忆网络;卷积神经网络;CNN-LSTM模型;小浪底水库;黄河

中图分类号：X832;TV882.1 文献标志码：A

doi：10.3969/j.issn.1000-1379.2021.05.018

Abstract： Water quality prediction is the basic and prerequisite work for the management of water resources and prevention and control of water pollution， but the research on water quality prediction in the Yellow River basin is relatively lagged behind. In order to improve the performance of the LSTM water quality prediction model and increase its generalization ability， according to the periodic and non-linear characteristics of water quality changes， taking the dissolved oxygen concentration of the Xiaolangdi Reservoir on the Yellow River as the research object， a combination of convolutional neural network CNN and length was constructed. The CNN-LSTM prediction model of the time memory network LSTM had been verified by experiments. The model can efficiently extract water quality feature information and perform time series prediction. The prediction error is lower than that of the LSTM model. The average absolute error of the predicted value and the root mean square error are 19.72% and 10.44% lower than that of the LSTM model respectively. The prediction of larger and smaller values is more accurate and it has better generalization performance.

Key words： water quality prediction; long and short-term memory network; convolutional neural network; CNN-LSTM model; Xiaolangdi Reservoir; Yellow River

黃河是我国第二长河、西北和华北地区的重要水源之一[1]，在我国经济发展、社会安定和生态环境保护等方面处于举足轻重的位置[2]。然而，随着生产工业化和人口城镇化速度加快等，黄河水质污染愈发严重[3]。水质预测是水资源管理和水污染防治的基础性和前提性工作，随着新一代信息技术的发展，越来越多的学者采用智能算法构建水质预测模型[4]，如李娜等[5]提出了一种灰色系统（GM）和新陈代谢原理结合的水质预测模型、Ju等[6]建立了预测水中氨氮含量的最小二乘支持向量机模型、刘洁等[7]将遗传算法和BP神经网络算法结合用于水质实时预测。由于水质预测涉及多种因素的非线性关系、需……