研究报告

  • 任树顺,高萌,王煦雯,余镒琦,陈纪新,陈能汪.基于3种时间序列模型的九龙江河流库区藻华预测[J].环境科学学报,2022,42(11):172-183

  • 基于3种时间序列模型的九龙江河流库区藻华预测
  • Algal bloom prediction in the Jiulong River Reservoir based on three types of time series models
  • 基金项目:国家自然科学基金(No.51961125203);中央高校基本科研业务费(No.20720200121);福建省环保科技项目(No.2021R009)
  • 作者
  • 单位
  • 任树顺
  • 厦门大学环境与生态学院,福建省海陆界面生态环境重点实验室,厦门 361102;厦门大学,近海海洋环境科学国家重点实验室,厦门 361102
  • 高萌
  • 罗德岛大学,海洋学研究生院,纳拉甘西特,美国罗德岛 02882
  • 王煦雯
  • 中国海洋大学环境科学与工程学院,青岛 266100
  • 余镒琦
  • 厦门大学环境与生态学院,福建省海陆界面生态环境重点实验室,厦门 361102;厦门大学,近海海洋环境科学国家重点实验室,厦门 361102
  • 陈纪新
  • 厦门大学,近海海洋环境科学国家重点实验室,厦门 361102
  • 陈能汪
  • 厦门大学环境与生态学院,福建省海陆界面生态环境重点实验室,厦门 361102;厦门大学,近海海洋环境科学国家重点实验室,厦门 361102
  • 摘要:湖库富营养化和有害藻华是全球性生态环境问题,藻华预测与早期预警是保障湖库水源地供水安全的关键技术.如何基于高频水生态在线监测数据进行藻华的实时动态预测成为水生态管理领域的重大需求.本研究以福建省九龙江江东库区(水源地)为例,利用3年连续观测的逐时平均总叶绿素a浓度数据,对比研究了SARIMA、Prophet和LSTM(长短期记忆神经网络)3种时间序列模型在藻华(日平均叶绿素a大于15 μg?L-1)预测方面的效果.结果表明:①时间序列模型要求参数少,灵活性强,能清晰反映水质特征和未来变化趋势,可弥补传统藻类监测预警方法的局限性;②基于深度学习框架的LSTM模型,具有独特的迭代优化算法,对藻类非线性变化特征的识别和预测能力较强,其总叶绿素a逐日预测和7日预测效果均显著优于SARIMA模型和Prophet模型;③输入数据长度会在一定程度上影响模型预测效果,最优的输入数据时间长度为7 d;输入数据频率对预测效果也有影响,在预测非藻华日时,小时数据的预测效果优于日频率数据;在预测藻华日时,两种频率数据 无显著差异,但日频率数据能更准确识别藻华日特征.总结起来,基于LSTM模型实现总叶绿素a浓度的短期预测,可为九龙江河流库区藻华早期预警和供水安全保障提供技术支撑.
  • Abstract:Eutrophication and harmful algal blooms in lakes and reservoirs are global eco-environmental issues. The prediction and early warning of algal blooms are the key techniques for securing the safe drinking water supply. How to predict algal blooms in a real-time dynamic way based on high-frequency water ecology monitoring data has become a major demand in the field of aquatic ecosystem management. Taking Jiangdong reservoir of Jiulong River (i.e., drinking water source of Xiamen in Fujian Province) as a case study, this study developed and compared the performance of three types of time series models of SARIMA, Prophet, and LSTM (long-term and short-term memory neural network) in predicting algal bloom (defined as daily average chlorophyll-a is greater than 15 μg?L-1), using the three-year continuously observed hourly mean total chlorophyll-a concentration data. The results show that: ①the time series model requires few parameters and has strong flexibility, which reflect the water quality characteristics and future trends, and can overcome the limitations of traditional methods of algae monitoring and early warning; ②The LSTM model based on the deep learning framework has a relatively strong ability to identify and predict the nonlinear variation characteristics of algae, due to its unique iterative optimization algorithm; the LSTM performance on daily prediction and seven-day prediction of total chlorophyll-a are both better than SARIMA model and Prophet model; ③The length of input data will affect the prediction performance of the models to some extent. The optimal length of inputs in this study was identified as 7-days. The frequency of input data also has an impact on the prediction performance. When predicting non-algal bloom days, the prediction ability to use hourly data is better than that of using daily data. When predicting algal bloom days, there is no significant difference between the two-frequency data, but the daily data can more accurately capture the characteristics of algal bloom. In summary, the short-term prediction of total chlorophyll-a concentration based on the LSTM model can provide technical support for early warning of algal bloom and water supply security in the Jiulong River Reservoir.

  • 摘要点击次数: 118 全文下载次数: 126