Accurate forecasts of water-quality indicators are essential for agriculture, aquaculture, and industry. We compare classical statistical models (ARIMA, ETS), machine-learning baselines (Random Forest, XGBoost), and Transformer-based models on two multivariate datasets from China and England. We propose Parallel ETSFormer, which runs multiple encoder branches in parallel and aggregates their level/growth/seasonal outputs to mitigate depth-wise error accumulation in limited-data settings. Across experiments, the Parallel ETSFormer especially reduces tail errors (P95AE, HAE) while remaining competitive on average metrics (MAE, RMSE). We also report robustness-focused metrics to support operational decision-making.