Dissolved oxygen prediction in the Dianchi River basin with explainable artificial intelligence based on physical prior knowledge

Abstract

Dissolved oxygen (DO) is a critical parameter for monitoring water quality. However, most existing deep learning models have overlooked the physical relationship between DO and other parameters during simulation, leading to simulated values that deviate from the actual physical laws. Moreover, the inherent opacity of deep learning models restricts their applicability. Here, we propose the prior knowledge-constrained bidirectional long short-term memory network (PKBiLSTM) model to simulate DO levels in the Dianchi Lake basin. Our results show that the PKBiLSTM model achieves an average Kling-Gupta efficiency coefficient (KGE) of 0.926, which represents a 3.35% and 2.38% increase compared to the gated recurrent unit (GRU) and categorical boosting (CatBoost) models, respectively. The experiments reveal that pH has the greatest effect on DO concentration within the range of 6.5–10. Furthermore, the primary factors affecting DO exhibit seasonal differences. The findings underscore the potential of our method to enhance the scientific management of watersheds.