An accurate forecasting model for key water quality factors based on Transformer with multi-scale attention mechanism

Abstract

The prediction of water quality parameters is vital for sustainable aquaculture. Dissolved oxygen (DO), a key factor influencing the health and growth of aquatic organisms, is challenging to predict due to its non-linearity and significant time lag. This study proposed a DO time-series prediction model based on Transformer architecture. A dynamic interpretable time-series decomposition strategy was proposed to extract the key feature information of the DO. A multi-scale decomposition attention mechanism was then designed to better understand the nonstationary characteristics in the time series and capture key features at different scales. Finally, the multi-scale temporal fusion attention mechanism reduced the loss of key information by integrating information from different scales to comprehensively capture complex patterns and dynamic changes in the data. Experimental results show that the prediction performance of the proposed model on six datasets including BaffleCreek is better than that of seven deep learning models.