A long-term prediction model for key water quality based on transformer with parallel attention mechanism and adaptive spectral enhancement
摘要
Accurate prediction of water quality parameters such as dissolved oxygen, temperature, salinity and pH is essential to ensure healthy marine ecosystems and sustainable aquaculture management. Water quality parameters reflect the ecological status and biogeochemical evolution of marine systems, serving as critical indicators for water quality regulation and health risk assessment. However, under the influence of global climate change and anthropogenic disturbances, the complexity of multivariable interactions and temporal dependencies in these high-dimensional, non-stationary datasets has increased significantly, making long-term forecasting a persistent challenge. However, global climate change and anthropogenic disturbances have exacerbated the complexity of multivariate interactions and temporal dependencies in these high-dimensional, non-stationary datasets, making long-term prediction particularly challenging. To address these challenges, we propose a Transformer-based model for long-term multivariate time series forecasting. First, an adaptive spectral enhancement module (ASEM) is introduced to dynamically suppress noise in the frequency domain while preserving critical information. Second, to better capture inter-variable correlations, the model incorporates a multi-source representation and intra-token parallel attention (MSR-TPA) mechanism, which dynamically adjusts feature importance and enhances local feature extraction. Furthermore, a multi-scale gating and inter-token parallel attention (MSG-TPA) mechanism is designed to simultaneously model multi-scale local features within variables and long-range dependencies across tokens. The model has been validated using eight real marine ranching datasets. On the Shandong Liugong Island dataset, it reduces the mean absolute error by 25.93%, the root mean square error by 40.78%, the R