Attention-based recurrent neural networks for accurate short-term and long-term dissolved oxygen prediction

Abstract

Accurate prediction of dissolved oxygen is important for the intelligent management and control in aquaculture. However, due to the interference of external factors and the irregularity of its own changes, it is still a difficult problem, especially in long-term prediction. Also, most of the current researches only report good results in short-term prediction of dissolved oxygen. In this paper, we study the effectiveness of attention-based recurrent neural networks (RNN) on short-term prediction (including about 1 h, 2 h) and long-term prediction (including about 8 h, 24 h and 48 h) of dissolved oxygen. We systematically discuss and compare the application of the attention-based RNN method in dissolved oxygen prediction, including spatial attention (Input-Attn), temporal attention (Temporal-Attn), spatio-temporal independent attention (DARNN and GeoMAN), and spatio-temporal joint attention (Spatio-temporal-Attn). Specifically, we first analyze the popular methods for water quality prediction and attention-based RNN methods for time series prediction. Then, we develop the latest attention-based RNN into a multi-step prediction for long-term dissolved oxygen prediction. Next, we propose two attention-based RNN structures to capture temporal relationships separately and learn spatio-temporal relationships simultaneously, which have achieved comparable performance with the state-of-the-art methods. Finally, we compare five attention-based RNN methods and five baseline methods in a real-world dataset. Experimental results show that attention-based RNN can achieve more accurate dissolved oxygen prediction in both short-term and long-term prediction.