| 摘要: |
| 总磷是广东省地表水自动监测网络的重要指标之一,其浓度超标会导致水体富营养化,造成藻类过渡繁殖从而破坏生态平衡。总磷浓度的准确预测对于地表水环境监管控制具有重要意义,其预测主要的难点在于难以捕捉水质数据复杂的动态时变性,本文提出一种新的模型(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,从而提高水质数据中长期预测能力。该模型基于Transformer架构,搭建了数据稳态调控模块(Data Stabilization Control)提高模型面对异质数据的稳定性,搭建时序深层解析模块(Sequential Deep Parsing Module),通过深层嵌入的动态时序解析提升模型对于多重周期和长期依赖的捕捉能力。在模型实验部分,本文搭建DSC-SDP-TF模型,基于广东省地表水自动监测网络六个地表水监测站点的水质监测数据,对总磷浓度进行预测,主要结论有,第一,在六个站点上的1日滚动预测平均RMSE、MAE和MAPE分别为 0.0053mg/L、0.0044????????????? mg/L和4.20%,5日滚动预测误差平均RMSE、MAE和MAPE分别为0.0082 mg/L、0.0069 mg/L和6.14%,1日预测和5日预测的平均误差均低于0.01mg/L；第二,DSC-SDP-TF模型在六个站点上5日滚动预测平均RMSE相比于Transformer模型降低36.66%,MAE降低38.37%,MAPE降低37.08%；第三,在消融实验部分,本文对比了数据稳态调控结构和时序深层解析结构对模型性能的贡献,结果表明,数据稳态调控结构和时序深层解析结构均有效增加了模型预测能力。 |
| 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 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Research on Total Phosphorus Concentration Prediction Model of Guangdong Province Surface Water Automatic Monitoring Network Based on Deep Learning Architecture |
|
yan hui hua,zhou quan
|
| Abstract: |
| Total phosphorus is one of the important indicators for the automatic monitoring network of surface water in Guangdong Province.Its excessive concentration can lead to eutrophication of water bodies, causing excessive reproduction of algae and disrupting ecological balance., causing algae overgrowth and disrupting ecological balance. Accurate prediction of total phosphorus concentration is significant for the regulation and control of surface water environments. The primary challenge in predicting total phosphorus lies in capturing the complex dynamic variability of water quality data. This paper proposes a new model (DSC-SDP-TF) aimed at addressing the multi-scale dynamic changes and complex temporal characteristics of water quality data, thereby improving the long-term prediction capabilities. The model is based on the Transformer architecture, incorporating a Data Stabilization Control module to enhance model stability when dealing with heterogeneous data, and a Sequential Deep Parsing Module to improve the model""s ability to capture multiple cycles and long-term dependencies through deeply embedded dynamic temporal analysis. In the experimental section, the DSC-SDP-TF model is constructed and applied to predict total phosphorus concentration using water quality monitoring data from six surface water monitoring stations in Guangdong Province""s surface water automatic monitoring network. The main conclusions are: first, the average RMSE, MAE, and MAPE for one-day rolling predictions across the six stations are 0.0053 mg/L, 0.0044 mg/L, and 4.20%, respectively, while the average RMSE, MAE, and MAPE for five-day rolling predictions are 0.0082 mg/L, 0.0069 mg/L, and 6.14%, with both one-day and five-day prediction errors averaging below 0.01 mg/L; second, compared to the Transformer model, the DSC-SDP-TF model reduces the average RMSE by 36.66%, MAE by 38.37%, and MAPE by 37.08% for five-day rolling predictions across the six stations; third, in the ablation experiments, the contributions of the Data Stabilization Control structure and the Sequential Deep Parsing structure to the model performance were compared, showing that both structures effectively enhanced the model""s prediction capability. |
| Key words: phosphorus multi-scale dynamic changes complex temporal variations deep learning data stabilization control sequential deep parsing |
