Integrating causality with temporal fusion transformer for interpretable and probabilistic forecasting of riverine...

Accurate, interpretable, and uncertainty-aware riverine dissolved oxygen forecasting is essential for aquatic ecosystem management. Challenges include complex non-stationarity, model opacity, and the need for uncertainty quantification. This study develops a Causal-Temporal Fusion Transformer (C-TFT) framework integrating neural Granger causality for driver selection, wavelet decomposition for multi-scale feature extraction, and a TFT for probabilistic forecasting. This framework was validated using high-frequency monitoring data from three representative sites in the Yangtze River Basin: Upstream, Midstream, and Downstream. The results showed that the C-TFT framework performed excellently across all sites, achieving its best performance at the Downstream station, with a Nash-Sutcliffe efficiency of 0.9727 and an average quantile loss of 0.0510. A quantitative component analysis confirmed that each module, including causal filtering, wavelet features, and the introduction of static covariates in a multi-site training strategy, made a quantifiable contribution to the final model's accuracy. The framework's interpretability revealed spatial heterogeneity in drivers, identifying Surface Solar Radiation as a universal driver and Water Temperature as a key factor only for Downstream. The TFT's internal attention mechanism showed that its predictions rely on long-term historical patterns and that it dynamically adjusts its focus during periods of abnormal DO variability. These findings underscore the potential of the proposed C-TFT framework as a high-performance, transparent, and uncertainty-aware tool for intelligent water resource management.