Unveiling the mystery of eutrophication prediction for large freshwater lake through Temporal Convolutional Network hybrid model

Eutrophication in large freshwater lakes poses a growing threat to aquatic ecosystems and drinking water safety, representing a major global environmental concern. However, most existing eutrophication prediction studies rely on a single water quality parameter or construct separate models for multiple parameters, lacking comprehensive predictions of water eutrophication that combine multiple water quality parameters. To address this, a hybrid prediction framework integrating feature selection, mode decomposition, deep learning, and intelligent optimization was proposed. This framework used a composite eutrophication index, derived from key water quality indicators, as a unified target to predict and assess trophic status holistically. The Relief method was employed to identify dominant influencing factors from multivariate water quality parameters. Variational Mode Decomposition (VMD) was then applied for multi-scale decomposition of time series, and the resulting modes were fed into a Temporal Convolutional Network (TCN) for prediction. Key model hyperparameters were optimized using the White Shark Optimizer (WSO). Comparative experiments against multiple baseline methods, including Random Forest (RF), Long Short-Term Memory (LSTM), and XGBoost models, demonstrated that the proposed Relief-VMD-WSO-TCN model achieved superior predictive accuracy and stability. Specifically, across three monitoring sections in Poyang Lake, the proposed model attains R² values ranging from 0.853 to 0.942 and Mean Absolute Percentage Error below 1.6 %, outperforming all comparison models. This study provided an effective solution for accurately predicting eutrophication status in complex aquatic environments.