三种水质动态预测模型在米山水库的应用与对比

黄林显; 张明芳; 钱永; 邢学睿; 邢立亭; 韩忠

引用本文:	黄林显,张明芳,钱永,等.三种水质动态预测模型在米山水库的应用与对比[J].灌溉排水学报,0,():-.
	HUANG Linxian,ZHANG Mingfang,QIAN Yong,et al.三种水质动态预测模型在米山水库的应用与对比[J].灌溉排水学报,0,():-.

【打印本页】【下载PDF全文】【查看/发表评论】【EndNote】【RefMan】【BibTex】

过刊浏览高级检索

本文已被：浏览 187次下载 0次
分享到：微信更多字体:加大+\|默认\|缩小-
三种水质动态预测模型在米山水库的应用与对比
黄林显¹, 张明芳², 钱永³, 邢学睿⁴, 邢立亭¹, 韩忠⁵
1.济南大学水利与环境学院;2.威海市水文中心;3.中国地质科学院水文地质环境地质研究所;4.山东正元地质资源勘查有限责任公司;5.山东省第六地质矿产勘查院

摘要:

水质的准确预测对科学进行水生态环境保护和水资源管理规划具有重要的意义。但水质动态由于受人类活动、降雨、气温和水动力条件等多种因素的影响而具有复杂的随机性特征,造成预测时容易产生较大的误差。为验证不同模型进行水质预测时的实用性,分别构建了季节性差分自回归滑动平均(Seasonal Autoregressive Integrated Moving Average, SARIMA)、霍尔特-温特(Holt-Winters)和长短时记忆(Long Short-Term Memory,LSTM)神经网络三种预测模型；利用米山水库2012-2018年月电导率数据进行模型训练,并利用2019年月电导率数据进行验证,以此考察三种预测模型的准确性和稳定性。结果表明：SARIMA和Holt-Winters模型建模过程相对简单,但由于仅能考虑水质数据的时序演化趋势,造成预测精度较低；LSTM神经网络模型能同时考虑水质数据的时序演化趋势及不同时刻之间的前后依赖关系,且具有较强的非线性映射能力,因此预测精度最高。总体上来看,LSTM神经网络预测模型仅在电导率值突变处误差相对较大,但整体预测效果较为理想,因此在水质预测中更加具有推广应用价值。

关键词: 时间序列模型;LSTM模型;电导率;水质预测;米山水库

DOI：

分类号:P641.2

基金项目:国家自然科学基金(42272288)；山东省高校院所创新团队项目(2021GXRC070)；河北省/地调局地下水污染机理与修复重点实验室开发基金(SK202103KF01)

Application and Comparison of Three Water Quality Prediction Models in Mishan Reservoir

HUANG Linxian¹, ZHANG Mingfang², QIAN Yong³, XING Xuerui⁴, XING Liting¹, HAN Zhong⁵

1.School of Water Conservancy and Environment,University of Jinan;2.Weihai Hydrological Center;3.Institute of Hydrogeology and Environmental Geology,Chinese Academy of Geological Sciences;4.Shandong Zhengyuan Geological Resources Exploration coltd;5.No Institution of Geology and Mineral Resources Exploration of Shandong Province

Abstract:

Accurate prediction of water quality is of great significance to scientific water ecological environment protection and water resource management planning. However, due to the influence of various factors such as human activities, rainfall, temperature, and hydrodynamic conditions, water quality dynamics have complex random characteristics, resulting in large prediction errors. To verify the practicability of different models for water quality prediction, three prediction models of SARIMA, Holt-Winters, and LSTM neural network were respectively constructed. The monthly conductivity data of Mishan Reservoir from 2012 to 2018 was used for model training, and the monthly conductivity data of 2019 was used for verification to examine the accuracy and stability of the three prediction models. The results show that the modeling process of SARIMA and Holt-Winters models is relatively simple, but because they can only consider the time series evolution trend of water quality data, the prediction accuracy is low. The LSTM neural network model can simultaneously consider the time-series evolution trend of water quality data and the before-and-after dependencies between different moments and has strong nonlinear mapping capabilities, so the prediction accuracy is the highest. Generally speaking, the LSTM neural network prediction model has a relatively large error only at the sudden change in the conductivity value, but the overall prediction effect is relatively ideal, so it has more application value in water quality prediction.

Key words: Time series model; LSTM model; conductivity; water quality prediction; Mishan reservoir