Explainable deep learning for multi-horizon water quality forecasting: cross-continental transferability for environmental decision support

Forecasting river water quality is central to proactive environmental decision-making, yet skill is rarely benchmarked against simple references or evaluated across river networks that differ structurally in climate and pollution regime. Using 48 years of records from 141 stations of a Brazilian regulatory network (CETESB) and 3,782 to 6,696 European stations of the Global River Water Quality Archive, we benchmarked an attention-based long short-term memory (LSTM) network for multi-horizon forecasting against a random forest and persistence. The LSTM was more accurate for dissolved oxygen (one-step R²log , the coefficient of determination on the log (1+x) scale, 0.659 versus 0.605, or 0.639 for a tuned random forest), but differences were small for the other nutrient and oxygen parameters. Heavily censored trace metals were reframed as classifiers of concentrations at or above time-invariant harmonized reporting limits; zinc and nickel were predicted no better than by station-specific detection frequencies, and exceedances of lead, cadmium and mercury could not be forecast usefully (PR-AUC ≤ 0.07 ). Attention weights and Shapley-value-based attributions disagreed on the temporal allocation of importance (Spearman ρ =0.37 for dissolved oxygen). Zero-shot transfer to Europe was parameter-specific: nearly intact for nitrate ( R²log 0.555 versus 0.598 in-domain) but a failure for dissolved oxygen ( $$-0.576$$ ) that neither a global offset nor removing the station embeddings ( $$-0.169$$ ) fully corrected. Because dissolved oxygen was overestimated at low-oxygen European sites, an alert rule needed an additional 1.67 mg L ⁻¹ of margin to flag 90% of violations. Predictive accuracy and decision-relevant robustness are related but distinct properties of forecasting models intended for transfer across monitoring networks.