Dissolved oxygen prediction in the Taiwan Strait with the attention-based multi-teacher knowledge distillation model

Abstract

This study investigates the prediction of dissolved oxygen (DO) in the Taiwan Strait, a crucial indicator for marine ecosystems. Accurate prediction of DO in coastal waters is critical for various coastal engineering activities. Despite the availability of numerous prediction methods, including numerical models, statistical models, and machine learning approaches, these often fall short when addressing complex dynamic marine areas like the Taiwan Strait. This challenge is partly due to the complexity of the environment and issues with data scarcity. The research focuses on seven buoy stations in the Taiwan Strait and proposes the Attention-based Multi-teacher Knowledge Distillation (AMKD) model, which integrates a multi-channel attention mechanism with a multi-teacher knowledge distillation algorithm, specifically enhancing the accuracy of DO prediction and adaptability to missing data. The model is capable of forecasting DO hourly for the next 24 h, effectively mitigating the randomness and instability associated with DO. Experimental comparisons with state-of-the-art prediction models demonstrate that our approach outperforms commonly used methods in terms of accuracy. Overall, the AMKD model presents a novel and effective solution for predicting DO in complex marine areas, with significant implications for future marine environmental monitoring and management.