Abstract: Excessive nitrogen, phosphorus, and permanganate in aquatic environments can lead to significant watershed pollution. Accurately predicting the levels of these indicators is crucial for effective pollution control. However, existing models often lack precision, and the selection of input factors lacks a mathematical basis. In this study, we propose a RF-BiLSTM hybrid network model focusing on the Yongjiang watershed as a case study. Leveraging the ability of RF (random forest) to extract optimal water quality index characteristics and the capacity of BiLSTM (bidirectional long-short-term memory) to capture temporal data patterns, our model employs dimensionality reduction followed by prediction to forecast TN, TP, and COD_Mn concentrations. Additionally, we conduct comparative analyses with benchmark models such as CNN, LSTM, BiLSTM, and RF-LSTM within the deep learning framework. Results demonstrate that our proposed model achieves lower mean absolute percentage errors (MAPE) for TN, TP, and COD_Mn at 4.33%, 6.781%, and 7.384%, respectively, outperforming other benchmark models. These findings indicate the high accuracy and practical utility of our predictions, offering valuable technical support for water pollution management.

Key words: water quality prediction, feature selection, random forest, bidirectional long-short-term memory network, deep learning