Hybrid Prediction for Water Quality with Bidirectional LSTM and Temporal Attention

Abstract

Accurate prediction of water quality indicators can effectively prevent sudden water pollution events, and control pollution diffusion. Neural networks, e.g., long short-term memory (LSTM) and encoder-decoder network, have been widely used to predict time series data. However, as the water quality data increases, it becomes unstable and highly nonlinear. Accurate prediction of water quality becomes a big challenge. This work proposes a hybrid prediction method called VBAED to predict the water quality time series. VBAED combines Variational mode decomposition (VMD), Bidirectional input Attention mechanism, an Encoder with bidirectional LSTM (BiLSTM), and a Decoder with temporal attention mechanism and LSTM. Specifically, VBAED first adopts VMD to decompose the ground truth time series, and the decomposed results are used as the input along with other features. Then, a bidirectional input attention mechanism is adopted to add weights to input features from both directions. VBAED adopts BiLSTM as an encoder to extract hidden features from input features. Finally, the predicted result is obtained by an LSTM decoder with a temporal attention mechanism. Real-life data-based experiments demonstrate that VBAED obtains the best prediction results compared with other widely used methods.