LSTM-TCN: dissolved oxygen prediction in aquaculture, based on combined model of long short-term memory network and temporal convolutional network

Abstract

Dissolved oxygen (DO) is an important water quality monitoring parameter of great significance in aquaculture. Accurate prediction of dissolved oxygen can help farmers to take necessary measures in advance to ensure the healthy growth of cultured species. The characteristics of multivariate and long-term correlation of water quality time series in the traditional methods make it difficult to achieve the expected prediction accuracy. To solve this problem, we propose the combined prediction method LSTM-TCN (long short-term memory network and temporal convolutional network). After the preprocessing of time series, the LSTM extracts the features of the series in time dimension, and then combines with TCN to build the fusion prediction model. In this study, we have carried out the DO predictions of LSTM and TCN algorithms separately, followed by the analysis of DO prediction, based on CNN-LSTM and LSTM-TCN combined models. The effects of attention mechanism and window size of historical time on the prediction results were also investigated. The experimental results show that the combined method has high accuracy in dissolved oxygen prediction, and can capture better characteristics of historical data with increasing time window of the historical dissolved oxygen sequence. The LSTM-TCN method achieves better prediction performance, with evaluation index values of MAE = 0.236, MAPE = 3.10%, RMSE = 0.342, and R² = 0.94.

Keywords: Attention mechanism; Combined model; Dissolved oxygen prediction; Long short-term memory network; Temporal convolutional network; Time window.