Abstract:

Accurately and efficiently predicting lake water quality is vital for water resource protection, ecological balance, and economic development. We propose a combined prediction model for total nitrogen (TN) concentration in lakes, integrating modal decomposition, multidimensional feature selection, Temporal Convolutional Network (TCN), self-attention mechanism, bidirectional long short-term memory (BiLSTM), and bidirectional Gate Recurrent Unit (BiGRU). First, we apply variational mode decomposition to break down the original TN sequence into intrinsic mode functions (IMFs) of different frequencies. This step effectively reduces the complexity and non-stationarity of the original sequence. Next, we use the random forest algorithm to select highly correlated features for each IMF. Then, we feed the filtered feature matrix into the TCN-BiLSTM hybrid network equipped with a self-attention mechanism for modeling. This network extracts key temporal information from the hidden data. Finally, to enhance the model’s prediction accuracy, we employ the BiGRU network to learn the detailed features of the residual sequence. We then fuse the residuals with the model’s prediction results to obtain the final prediction value. We conduct an experimental analysis using the water quality data from the Duchang Monitoring Station in Poyang Lake. The results demonstrate that, compared with other models, our model significantly improves the prediction accuracy of TN concentration. Specifically, its mean absolute error (MAE) is 0.03 mg/L, root mean square error (RMSE) is 0.049 mg/L, and coefficient of determination (R²) is 0.992.

Key words: water quality prediction, total nitrogen, variational mode decomposition, temporal convolutional network, integrated prediction