Robust long-term water quality forecasting via decomposition-enhanced patching and asymmetric attention fusion

Accurate long-term water quality forecasting is critical for ecosystem management but remains challenged by non-stationary data patterns and complex multivariate dependencies. Existing methods often struggle with error accumulation and introduce noise through symmetric modeling of covariates. To address these issues, we propose WaterLTS, a novel long-term water quality forecasting model. WaterLTS first decouples non-stationary series into trend and seasonal components using a moving average strategy. It then employs a Period-aware Patching mechanism, guided by Fast Fourier Transform (FFT), to align semantic segments with intrinsic physical cycles. Crucially, we introduce a Dual-Attention Asymmetric Fusion Encoder that treats covariates strictly as contextual guidance. By establishing a directed local–global interaction where target patches query global covariate tokens, this design effectively filters irrelevant noise while capturing essential environmental driving forces. Experiments on real-world United States Geological Survey (USGS) datasets for Dissolved Oxygen and Total Phosphorus demonstrate that WaterLTS achieves an average error reduction of 5.40% compared to state-of-the-art baselines. The model exhibits superior stability across extended horizons, offering a robust solution for long-term environmental monitoring.