Scholarlink

<p>Dissolved oxygen (DO) is a key indicator of reservoir water quality and ecosystem integrity. The dynamics of DO are governed by complex physical, chemical, and biological reactions such as thermal stratification, nutrient interactions, photosynthesis, respiration, and microbial decomposition of algae biomass, surface oxygen aeration, and sediment oxygen demand (SOD), making accurate simulation challenging. This study proposes an integrated machine learning framework for DO prediction in the Javeh Reservoir by combining ensemble learning models, explainable Artificial Intelligence, and network-based visualization. Six models (Random Forest, XGBoost, Hist Gradient Boosting, Extra Trees, MLP, and SVR) were optimized using five-day averaged water quality flux, hydrological, hydraulics, and meteorological data. The prediction results were evaluated using statistical measures, MAE, RMSE and Nash–Sutcliffe efficiency (NSE). Among the tested models, ET achieved the best predictive performance (MAE = 0.3465&#xa0;mg/L, RMSE = 0.5997&#xa0;mg/L, NSE = 0.9655), indicating excellent agreement between observed and predicted DO values. To enhance interpretability, SHapley Additive exPlanations (SHAP) was applied to the best-performing model. The results highlight air temperature, inflow fluxes of DO, phosphate, ammonium, and outflow discharge as dominant drivers of DO variability, reflecting combined thermal forcing and biogeochemical oxygen-demand processes. In addition, a Chord-diagram connectivity visualization was used to summarize DO linkages with top predictors, providing a complementary network-oriented view of model explanations. Overall, by replacing “black-box” simulation with highly interpretable models, the proposed framework ensures both accuracy and transparency, offering a practical, transferable tool for proactive reservoir water quality management.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable ensemble machine learning for dissolved oxygen prediction in a reservoir using SHAP and chord diagram analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Omid Yazdan Panah,
  • Motahareh Saadatpour

摘要

Dissolved oxygen (DO) is a key indicator of reservoir water quality and ecosystem integrity. The dynamics of DO are governed by complex physical, chemical, and biological reactions such as thermal stratification, nutrient interactions, photosynthesis, respiration, and microbial decomposition of algae biomass, surface oxygen aeration, and sediment oxygen demand (SOD), making accurate simulation challenging. This study proposes an integrated machine learning framework for DO prediction in the Javeh Reservoir by combining ensemble learning models, explainable Artificial Intelligence, and network-based visualization. Six models (Random Forest, XGBoost, Hist Gradient Boosting, Extra Trees, MLP, and SVR) were optimized using five-day averaged water quality flux, hydrological, hydraulics, and meteorological data. The prediction results were evaluated using statistical measures, MAE, RMSE and Nash–Sutcliffe efficiency (NSE). Among the tested models, ET achieved the best predictive performance (MAE = 0.3465 mg/L, RMSE = 0.5997 mg/L, NSE = 0.9655), indicating excellent agreement between observed and predicted DO values. To enhance interpretability, SHapley Additive exPlanations (SHAP) was applied to the best-performing model. The results highlight air temperature, inflow fluxes of DO, phosphate, ammonium, and outflow discharge as dominant drivers of DO variability, reflecting combined thermal forcing and biogeochemical oxygen-demand processes. In addition, a Chord-diagram connectivity visualization was used to summarize DO linkages with top predictors, providing a complementary network-oriented view of model explanations. Overall, by replacing “black-box” simulation with highly interpretable models, the proposed framework ensures both accuracy and transparency, offering a practical, transferable tool for proactive reservoir water quality management.