基于深度学习的地表水总磷浓度预测模型研究

黎如昊; 严惠华; 周泉

引用本文:	黎如昊,严惠华,周泉.基于深度学习的地表水总磷浓度预测模型研究[J].环境监控与预警,2025,17(2):41-50
	LI Ruhao,YAN Huihua,ZHOU Quan.Research on a Deep Learning-Based Prediction Model for Surface Water Total Phosphorus Concentration[J].Environmental Monitoring and Forewarning,2025,17(2):41-50

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

←前一篇|后一篇→

过刊浏览高级检索

本文已被：浏览 96次下载 16次	码上扫一扫！
分享到：微信更多字体:加大+\|默认\|缩小-
基于深度学习的地表水总磷浓度预测模型研究
黎如昊¹，严惠华¹，周泉²
1. 广东省生态环境监测中心，广东广州 510308；2.生态环境部华南环境科学研究所河口与海岸生态环境研究中心，广东广州 530535

摘要:

提出一种基于数据稳态调控和深层时序解析的模型（DSC-SDP-TF），专注于解决水质数据的多尺度动态变化和复杂时变特性，从而提高水质数据的长期预测能力。该模型基于Transformer模型，搭建了数据稳态调控模块（DSC），提高模型面对异质数据的稳定性，搭建时序深层解析模块（SDP），提升模型对于多重周期和长期依赖的捕捉能力。基于广东省地表水自动监测网络6个地表水监测站点的水质监测数据，对总磷质量浓度进行预测。结果表明：（1）DSC-SDP-TF模型在6个站点上的1 d滚动预测的平均均方根误差（RMSE）、平均绝对误差（MAE）和平均百分比误差（MAPE）分别为 0.005 3 mg/L、0.004 4 mg/L和4.20%，1 d和5 d预测的平均误差均＜0.01 mg/L；（2）DSC-SDP-TF模型在6个站点上的5 d滚动预测平均RMSE相比于Transformer模型降低36.66%，MAE降低38.37%，MAPE降低37.08%；（3）对比了数据稳态调控模块和时序深层解析模块对模型性能的贡献，表明这2个模块均有效增加了模型的预测能力。

关键词: 总磷多尺度动态变化复杂时变特性深度学习数据稳态调控时序深层解析预测模型

DOI：DOI：10.3969/j.issn.1674-6732.2025.02.006

分类号:X824

基金项目:热带海洋环境国家重点实验室（中国科学院南海海洋研究所）开放课题（LTO2221）

Research on a Deep Learning-Based Prediction Model for Surface Water Total Phosphorus Concentration

LI Ruhao¹， YAN Huihua¹， ZHOU Quan²

1.Guangdong Ecological Environment Monitoring Centre， Guangzhou， Guangdong 510308， China； 2. South China Institute of Environmental Sciences， Ministry of Ecology and Environment， Guangzhou， Guangdong 530535， China

Abstract:

A model based on Data Stabilization Control and Sequential Deep Parsing Transformer（DSC-SDP-TF） is proposed， focusing on addressing the multi scale dynamic changes and complex time-varying characteristics of water quality data to improve the long-term prediction capability of water quality data. The model is based on the Transformer architecture and includes a Data Stabilization Control（DSC） module to enhance the models stability in dealing with heterogeneous data， and a Sequential Deep Parsing（SDP） module to improve the model's ability to capture multiple cycles and long-term dependencies. The model is used to predict the total phosphorus concentration in water quality monitoring data from six surface water monitoring stations of the Guangdong Province surface water automatic monitoring network. The results show：（1） The DSC-SDP-TF model achieves an average Root Mean Square Error（RMSE） of 0.005 3 mg/L， Mean Absolute Error（MAE） of 0.004 4 mg/L， and Mean Absolute Percentage Error（MAPE） of 4.20% for 1 day rolling forecasts at the six stations， with both 1 day and 5 day predictions having an average error of less than 0.01 mg/L；（2） The average RMSE of the 5 day rolling prediction of the DSC-SDP-TF model at the six stations is reduced by 36.66% compared to the Transformer model， with MAE reduced by 38.37% and MAPE reduced by 37.08%；（3） A comparison of the contributions of the DSC and SDP structures to the model's performance demonstrates that both structures effectively improve the model's predictive capability.

Key words: Phosphorus Multi-scale dynamic changes Complex temporal variations Deep learning Data stabilization control Sequential deep parsing Prediction model Prediction model