Interval prediction of TN based on the bidirectional long short-term memory- residual block-bayesian optimization model

Abstract Based on multi-station water quality data in Guangzhou section of the Pearl River Basin, a bidirectional long short-term memory - residual block-Bayesian optimization model (Bidirectional long short-term memory - residual block-Bayesian optimization model) is designed by combining BI-LSTM, residual network and Bayesian optimization. The results show that compared with the reference model, the model converges faster and the prediction accuracy is higher. To further investigate the impact of socioeconomic and land use factors on water quality, a random forest algorithm is employed to quantify the relative importance of land use composition and landscape pattern indices in influencing TN concentrations. The results reveal that variables such as land use intensity, landscape fragmentation, and specific land cover types substantially affect TN levels, indicating a strong correlation between anthropogenic activities and nitrogen pollution. This integrated modeling approach not only improves prediction accuracy but also provides important insights into the spatiotemporal mechanisms underlying water quality variation. The findings offer valuable support for data-driven decision-making in watershed management and targeted pollution mitigation strategies in rapidly urbanizing catchments.