湖泊藻类生长的时滞驱动特征解析:融合随机采样与物理约束的可解释机器学习
CSTR:
作者:
作者单位:

1.云南省生态环境科学研究院;2.南京师范大学地理科学学院

作者简介:

通讯作者:

中图分类号:

基金项目:

国家自然科学(U2102207)、云南省科技厅重点研发计划(202503AT100008)、云南省基础研究计划项目(202501AT070436)、云南省生态环境科学研究院流域污染过程模拟新方法研究创新团队(2023-07CXTD)资助。


Deciphering Time-Lag Driving Characteristics of Algal Blooms in Lakes: An Interpretable Machine Learning Approach with Random Sampling and Physical Constraints
Author:
Affiliation:

1.Yunnan Academy of Ecological and Environmental Sciences,Kunming;2.Nanjing Normal University

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 附件
  • |
  • 文章评论
    摘要:

    驱动因子对湖泊藻类生长的驱动存在时滞特性,现有方法难以在高维交互因子的驱动条件下同时精准识别时滞区间并阐释其驱动特征。为此,本研究提出了一种融合随机采样与物理约束建模的可解释机器学习框架,旨在系统解析藻类生长的时滞驱动特征。该框架首先通过随机采样构建多时滞特征集,利用XGBoost模型与SHAP值分析,自适应地识别驱动因子影响藻类生长的最优时滞窗口。进而,在模型训练中引入现有物理约束项,并通过分析约束模型下的特征依赖图(PDP),定量揭示驱动因子的非线性作用特征。本研究以中国云南高原不同营养状态的两个湖泊(重度富营养化的滇池和轻度富营养化的洱海)为案例进行验证。结果表明:(1)该框架对叶绿素α浓度变化具有较好的模拟性能,验证集对滇池与洱海的R2达到0.88以上,MAPE低于14.6%,构建的模型对滇池和洱海分别可实现提前4天(R2=0.79, MAPE=13.3%)和6天(R2=0.82, MAPE=16.8%)的Chl-a高精度预测。(2)高藻情期水体理化与气象因子的时滞普遍大于低藻情期,主要归因于水体理化因子对藻类生长的累积驱动特征及高藻情期对光照的遮蔽效应。(3)滇池营养盐对叶绿素α的驱动时滞快于洱海,体现富营养条件下营养物质的即时消耗特征。(4)重度富营养化的滇池,其藻类生长主要受到水体理化特征(pH、DO)、营养物质(TN)与可氧化有机物(CODMn)的共同驱动;而处于轻度富营养化状态的洱海,则主要受到水体理化(水温、pH、DO)的驱动。本研究构建的框架为从“识别时滞”到“阐释特征”的探索提供新的可行路径,有助于发展更具过程理解与可解释性的藻华预测模型,并为针对不同湖泊类型制定差异化、前瞻性的生态管理策略提供科学依据。

    Abstract:

    The response of algal growth to driving factors exhibits notable time-lag effects. Existing methods face challenges in accurately identifying time-lag intervals and elucidating the associated driving characteristics under high-dimensional interactions among multiple factors. To address this gap, this study proposed a novel interpretable machine learning framework that integrated random sampling with physically constrained modeling to systematically analyze the time-lag driving characteristics of algal growth. The framework first constructed a time-lag feature set via random sampling, then employed XGBoost and SHAP analysis to adaptively identify the optimal time-lag window for each driving factor. Subsequently, physical constraints were incorporated into model training, and partial dependence plots (PDP) derived from the constrained model were used to quantitatively reveal the nonlinear effects of the drivers. The proposed framework was validated using two lakes of contrasting trophic statuses in the Yunnan Plateau, China: the hyper-eutrophic Dianchi Lake and the mildly eutrophic Erhai Lake. The results demonstrated that: (1) The framework achieved high predictive performance for chlorophyll-a concentrations, with R2 values exceeding 0.88 and MAPE below 14.6%. High-accuracy Chl-a forecasting with lead times of 4 days (R2 = 0.79, MAPE = 13.3%) and 6 days (R2 = 0.82, MAPE = 16.8%) can be realized for Dianchi and Erhai, respectively. (2) During high-algal-bloom periods, the time lags of hydrochemical and meteorological factors were generally longer than those during low-bloom periods, primarily attributable to the cumulative effects of hydrochemical factors on algal growth and light attenuation caused by dense algal blooms. (3) Nutrients responded faster to chlorophyll-a in Dianchi than in Erhai, reflecting the rapid consumption of nutrients under eutrophic conditions. (4) In hyper-eutrophic Dianchi, algal growth was co-driven by hydrochemical factors (pH, DO), nutrients (TN), and oxidizable organic matter (CODMn). In contrast, in mildly eutrophic Erhai, the primary drivers were hydrochemical factors (water temperature, pH, DO). In summary, this framework offered a robust pathway from time-lag identification to mechanistic interpretation, facilitating the development of process-aware and explainable algal bloom forecasting models, and providing science-based support for tailored, proactive ecological management strategies across different lake types.

    参考文献
    相似文献
    引证文献
引用本文
相关视频

分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2026-03-03
  • 最后修改日期:2026-07-10
  • 录用日期:2026-07-17
  • 在线发布日期: 2026-10-15
  • 出版日期:
文章二维码
您是第    位访问者
地址:南京市江宁区麒麟街道创展路299号    邮政编码:211135
电话:025-86882041;86882040     传真:025-57714759     Email:jlakes@niglas.ac.cn
Copyright:中国科学院南京地理与湖泊研究所《湖泊科学》 版权所有:All Rights Reserved
技术支持:北京勤云科技发展有限公司

苏公网安备 32010202010073号

     苏ICP备09024011号-2