Multi-indicator water quality prediction with attention-assisted bidirectional LSTM and encoder-decoder

Abstract

Accurate and real-time prediction of water quality not only helps to assess the environmental quality of water, but also effectively prevents and controls water quality emergencies. In recent years, neural networks represented by Bidirectional Long Short-Term Memory (BiLSTM) and Encoder-Decoder (ED) frameworks have been shown to be suitable for prediction of time series data. However, traditional statistical methods cannot capture nonlinear characteristics of the water quality, and deep learning models often suffer from gradient disappearance and gradient explosion problems. This work proposes a hybrid water quality prediction method called SVABEG, which combines a Savitzky-Golay (SG) filter, Variational Mode Decomposition (VMD), an Attention mechanism, BiLSTM, an ED structure, and a hybrid algorithm called Genetic Simulated annealing-based Particle Swarm Optimization (GSPSO). SVABEG first adopts the SG filter and VMD to remove noise and deal with nonlinear features in the original time series, respectively. Then, SVABEG combines BiLSTM, the ED structure and the attention mechanism to capture bi-directional long-term correlations, realize dimensionality reduction and extract key information, respectively. Furthermore, SVABEG adopts GSPSO to optimize its hyperparameters. Experimental results with real-life datasets demonstrate that the proposed SVABEG outperforms current state-of-the-art algorithms in terms of prediction accuracy.