Acta Scientiarum Naturalium Universitatis Pekinensis ›› 2022, Vol. 58 ›› Issue (2): 337-344.DOI: 10.13209/j.0479-8023.2021.126

Previous Articles     Next Articles

Real-Time River Water Quality Prediction Model Based on Spatial Correlation and Neural Network Model

ZHANG Yang1, XIAN Huiting2, ZHAO Zhijie1,†   

  1. 1. College of Environmental Science and Engineering, Peking University, Beijing 100871 2. Guangzhou Urban Drainage Monitoring Station, Guangzhou 510010
  • Received:2021-03-30 Revised:2021-05-16 Online:2022-03-20 Published:2022-03-20
  • Contact: ZHAO Zhijie, E-mail: zhaozhijie(at)pku.edu.cn

基于空间相关性和神经网络模型的实时河流水质预测模型

张阳1, 冼慧婷2, 赵志杰1,†    

  1. 1. 北京大学环境科学与工程学院, 北京 100871 2. 广州市城市排水监测站, 广州 510010
  • 通讯作者: 赵志杰, E-mail: zhaozhijie(at)pku.edu.cn
  • 基金资助:
    广州市水务局典型流域考核断面水质达标关键技术与应用项目(GZCPJ/ZD-2020-38)资助 

Abstract:

Based on the high frequency water quality online monitoring data, the spatial correlation of water quality data was used to construct a neural network model to realize the real-time prediction of river water quality. The model was applied to the Baini River Basin in Guangzhou, and the water quality parameters of dissolved oxygen and ammonia nitrogen were predicted and analyzed to verify the effect of the model. According to different prediction time, six water quality prediction models were built, and the results showed that the model predicting dissolved oxygen 6 hours in advance had better prediction effect, while the model predicting ammonia nitrogen 24 hours in advance had better effect. The average absolute errors of the better trained model for real-time water quality prediction of dissolved oxygen and ammonia nitrogen were 0.43 mg/L and 0.29 mg/L, respectively, and the root mean square errors were 0.71 mg/L and 0.36 mg/L, respectively. At 95% confidence level, the prediction interval coverage rates were 96.6% and 97% respectively. The model can be used as the early warning of abnormal water quality events. At the same time, the sensitivity analysis of the input items by the model can be used to analyze the pollution sources to help the basin identify the main sources of pollutants. 

Key words: neural network, real-time water quality prediction, spatial correlation, pollution source analysis

摘要:

基于高频水质在线监测数据, 利用水质数据的空间相关性构建神经网络模型实现对河流水质的实时预测。应用此模型对广州市白坭河流域的水质参数溶解氧和氨氮进行预测和分析, 验证水质模型的效果。根据预测时间的不同, 搭建 6 种水质预测模型, 结果显示溶解氧提前6小时预测的模型具有更好的预测效果, 氨氮提前24小时内预测的水质模型效果较好。训练较好的模型对溶解氧和氨氮实时水质预测的平均绝对误差分别为0.43和0.29 mg/L, 均方根误差分别为0.71和0.36 mg/L。在95%置信度水平下, 预测区间覆盖率分别为96.6%和97%。该模型可以作为水质异常事件的预警, 同时可以借助模型对输入项的敏感性分析, 进行污染源解析, 帮助流域识别污染物主要来源。

关键词: 神经网络, 实时水质预测, 空间相关性, 污染源解析