Predicting ammonia nitrogen in surface water by a new attention-based deep learning hybrid model

Ammonia nitrogen (NH3-N) is closely related to the occurrence of cyanobacterial blooms and destruction of surface water ecosystems, and thus it is of great significance to develop predictive models for NH3-N. However, traditional models cannot fully consider the complex nonlinear relationship between NH3-N and various relative environmental parameters. The long short-term memory (LSTM) neural network can overcome this limitation. A new hybrid model BC-MODWT-DA-LSTM was proposed based on LSTM combining with the dual-stage attention (DA) mechanism and boundary corrected maximal overlap discrete wavelet transform (BC-MODWT) data decomposition method. By introducing attention mechanism, LSTM could selectively focus on the input data. BC-MODWT could decompose the input data into sublayers to determine the main swings and trends of the input feature series. The BC-MODWT-DA-LSTM hybrid model was superior to other studied models with lower average prediction errors. It could maintain NASH Sutcliffe efficiency coefficient (NSE) values above 0.900 under the lead time up to 7 days, and the area under the receiver operating characteristic (ROC) curve could reach 0.992. The hybrid model also had higher prediction accuracies at the peak spots, indicating that it was capable of early warning when sudden high NH3-N pollution occurred. The high forecasting accuracy of the suggested hybrid method proved that further improving LSTM model without introducing more complex topologies was a promising water quality prediction method.

Automated summary

This study proposed a new hybrid model based on LSTM neural network, combining the dual-stage attention mechanism and boundary corrected maximal overlap discrete wavelet transform data decomposition method. Compared with other models, the hybrid model showed higher prediction accuracy and provided early warning for NH3-N pollution.