Descomposición multiescala y atención de reglas difusas: Un marco transferible entre cuencas para la predicción a...

Accurate long-term forecasting of river water quality is pivotal for ecosystem management; however, process-based models are often limited due to fragmented historical records lacking long-term continuity and complex, non-stationary interactions among environmental drivers. To address the trade-off between predictive accuracy and inter-basin generalizability, we propose a transferable deep learning framework integrating physical signal decomposition with fuzzy logic. Specifically, we employ multichannel singular spectrum analysis to isolate distinct trends from stochastic noise. The framework integrates a Transformer architecture with adaptive fuzzy mechanisms, designed to manage uncertainty and capture complex spatiotemporal dependencies. Evaluated on in-situ observations from 149 monitoring stations across diverse Chinese river basins from 2007 to 2022, the model achieved an overall average Nash-Sutcliffe efficiency (NSE) of 0.75±0.19 for 26-step-ahead (approximately six-month) forecasts. It demonstrated effective predictive capability for indicators such as dissolved oxygen and pH, and improved stability for event-driven pollutants like COD_Mn and ammonia nitrogen. Comparative analysis reveals that the proposed framework outperforms state-of-the-art baselines, achieving an average NSE improvement of 24.9%. Notably, the cross-region modeling strategy yielded a 28.4% performance gain in NSE compared to independent single-station models, validating the framework's ability to transfer learned structural priors across heterogeneous regions. These findings suggest that integrating structural decomposition with soft fuzzy constraints offers a promising solution for water quality forecasting in data-scarce basins.