An interpretable hierarchical neural network insight for long-term water quality forecast: A study in marine ranches of Eastern China

Abstract

Accurate forecast of water quality parameters is important for water quality monitoring and water quality regulation. However, the increasingly complex marine water environment makes improving the accuracy of forecasting water quality parameters challenging. In this study, a new interpretable deep learning method for forecasting water quality parameters is proposed. This method inputs the decomposed feature data into different stacks for data processing. Several parallel structure stacks are designed to capture the features of decomposed sequences. A new attention mechanism based on the combination of recent and long-term historical data is proposed, as well as an enhanced double residual temporal convolutional network block module. In this study, dissolved oxygen data obtained from eight marine ranches along the coast of Shandong Peninsula are used. The results showed that the MAE, RMSE and MAPE of our model Forecast results were 33.48%, 33.33% and 29.26% lower than those of other algorithms on average and R2 was 6% higher on average. Our model exhibited the highest degree of fitting between the predicted value and the observed value, with the best linear fitting and the smallest error. Our work provides a valuable framework for investigating the cause and influence of water quality Forecast in marine ranches.

Keywords

• Deep learning
• Self-attention mechanism
• Temporal convolutional network
• Dissolved oxygen
• Forecast