The Use of Attention-Enhanced CNN-LSTM Models for Multi-Indicator and Time-Series Predictions of Surface Water Quality

Deep learning (DL) has recently been applied to surface water quality prediction, whereas its online monitoring data consists of multiple indicators and time series, which are challenging for prediction models due to complex temporal dependencies and inter-indicator mechanisms. Convolutional neural network (CNN) and long short term memory (LSTM) can be used for indicator and temporal domains respectively, but still lack the ability to represent complex patterns in surface water quality. Since attention mechanisms are designed to effectively focus on the most crucial information, spatial attention mechanism (SAM) and temporal attention mechanism (TAM) are suitable for dealing with the above multi-indicator and time series issues. This work incorporates SAM and TAM into the CNN-LSTM model to form 4 DL models for water quality prediction including CNN-LSTM, SAM-enhanced CNN-LSTM, TAM-enhanced CNN-LSTM, and the CNN-LSTM enhanced by both attention mechanisms. Four surface water online monitoring sites are used as case studies to examine the models in predicting three water quality indicators including dissolved oxygen (DO), ammonia nitrogen (NH₃-N), and total organic carbon (TOC). According to the case results of the 4 models after training with similar training epochs, the prediction accuracies of attention-enhanced models are better than the CNN-LSTM model, and the model with both attention mechanisms generally achieves the best performance among the 4 models. The prediction NSE of DO by the four models are 0.817, 0.948, 0.952, and 0.967 respectively in a representative case Jiujiang. The results demonstrate that spatial and temporal attention can analyze correlations from multiple indicators and time series of water quality data respectively, to improve the accuracy of surface water quality prediction.