Application of Seq2Seq Model Based on TCN-GRU to Multivariate Water Quality Time Series Prediction

Water pollution is still a common problem worldwide and needs to be solved urgently. Accurately predict changes in water quality can help decision makers better solve water pollution problems. However, water quality data are characterized by uncertainty, nonlinearity and complexity, the existing water quality prediction models have low prediction accuracy and poor generalization ability. In this paper, a seq2seq multivariate water quality prediction model combining TCN model, GRU model and self-attention mechanism is proposed. The traditional seq2seq model based on the basic architecture of recurrent neural network has limited ability to extract features, and it is difficult to effectively process large-scale multivariate water quality data. We use the TCN model as the encoder and the GRU model as the decoder, so as to effectively use the nonlinear feature processing capability of TCN model for time series data and the nonlinear fitting ability of the GRU model. In addition, we add a self-attention mechanism to the decoder, so that the model can effectively extract hidden related factors that affect water quality changes, further improved the predictive power of the model. To demonstrate the predictive power of the TCN-GRU seq2seq model, we conduct an experimental study on a real river water quality dataset, Experiments show that our model achieves optimal prediction results.