Research on theLSTM River Dissolved Oxygen Prediction Model Optimized by Attention Mechanism
Zhou Quan1,2, Hu Xuanming3, Wang Dongkun3, Zhang Wucai1,2, Chen Zhongying1,2, Wang Jinpeng1,2, Wang Pengyang3*, Ren Xiuwen1,2*
1. South China Institute of Environmental Sciences, Ministry of Ecology and Environment, Guangzhou, Guangdong 510530
2. Key Laboratory of Water Environment Simulation and Pollution Control, Ministry of Ecology and Environment, Guangzhou, Guangdong 510530
3. University of Macau, National Key Laboratory of Smart City IoT, Macau Special Administrative Region 999078
Dissolved oxygen(DO) is one of the important indicators for measuring the water environment, and predicting the DO of water bodies is an important means to study the ecological status of water environments.
In recent years, with the development of artificial intelligence technology, data-driven models have been widely applied in the field of water environment protection. Considering the strong nonlinearity and significant non-stationarity of dissolved oxygen time series data, this paper proposes a river dissolved oxygen prediction model based on a dual-stage attention weight optimization mechanism using Long Short-Term Memory(LSTM) (DAIW-LSTM model).
Results show:
① The model’s encoder includes a spatial attention mechanism optimized with dual-stage weights, and the decoder includes a temporal attention mechanism optimized with dual-stage weights. By comparing with baseline models, the accuracy and reliability of this model in time series prediction are validated.
② The analysis of the dual-stage attention weight optimization mechanism reveals the intrinsic mechanism of the model in optimizing feature weights. The second stage optimizes and corrects the preliminary weights from the first stage, enabling the model to more accurately distinguish the weights of various features, thus improving the prediction accuracy of the model.
③ The input of upstream water quality indicators as new feature variables has a positive impact on the prediction performance of the DAIW-LSTM model, but an excessive number of upstream features does not necessarily enhance the model’s prediction performance.
Zhou Quan, South China Institute of Environmental Sciences
The first author of this paper, Zhou Quan, is a senior engineer at the South China Institute of Environmental Sciences, Ministry of Ecology and Environment, mainly engaged in research on water pollution prevention and control, numerical simulation, and the application of artificial intelligence in environmental protection. He has led 7 scientific research projects, including the National Natural Science Foundation, National Environmental Protection Special Projects, and Guangxi Key Research and Development Program. He has also led over 10 project outcomes for technology transfer. He is a member of the expert group for the key sea area comprehensive management battle of the Ministry of Ecology and Environment, and an industry mentor at the Chinese Academy of Environmental Sciences. He has published over 20 papers and holds over 10 authorized invention patents. He has received the second prize of the Environmental Protection Science and Technology Award.
Wang Pengyang, University of Macau
The corresponding author of this paper, Wang Pengyang, is an assistant professor at the National Key Laboratory of Smart City IoT, University of Macau. He mainly engages in research on spatiotemporal data mining, especially in the representation, prediction, and anomaly detection of spatiotemporal data based on the Internet of Things. He has published over 50 papers in top journals and conferences in artificial intelligence, including 35 in CCF-A, and holds 2 US patents. He has received the “AI Chinese New Star Top 100” award, the SIGKDD 2018 Best Paper Nomination Award, and the SIGSPATIAL 2020 Best Paper Nomination Award. His work on automatic urban planning has been reported by media such as “Synced Review” and “UCF Today,” and his series of works on urban vitality have been reported by the US National Science Foundation.
Ren Xiuwen, South China Institute of Environmental Sciences
The corresponding author of this paper, Ren Xiuwen, is a senior engineer at the South China Institute of Environmental Sciences, Ministry of Ecology and Environment. He mainly engages in research on water pollution prevention and control and water environment management. He has led over 10 specialized outcomes for national water projects, the second national pollution source census, and environmental public welfare projects, and has led over 30 project outcomes for technology transfer. He is a member of the expert group for the key sea area comprehensive management battle of the Ministry of Ecology and Environment and a member of the River Basin Water Ecological Environment Protection Committee of the Guangdong Provincial Environmental Society. He has published over 20 papers and holds over 10 authorized invention patents. He has received the second prize of the Environmental Protection Science and Technology Award.
