Comparative evaluation of deterministic and Bayesian neural networks for chlorophyll-a time series forecasting in...

Reliable forecasting of chlorophyll-a concentration is essential for early warning of harmful algal blooms in eutrophic lakes. However, most deep learning approaches provide only deterministic point estimates without quantifying predictive uncertainty. This study compared four neural network architectures (MLP, LSTM, and their Bayesian variants via Monte Carlo dropout) for multi-step chlorophyll-a forecasting in Dianchi Lake, China. Models were trained on 1461 daily observations of eight physicochemical variables (seven environmental predictors plus chlorophyll-a) spanning 2022 to 2025. The dataset was chronologically split into 70% training and 30% testing, with the final 20% of the training set withheld as a validation set. Bayesian-LSTM achieved the best performance across all forecast horizons, maintaining R²=0.678 at 7-day lead time while producing calibrated uncertainty estimates. Predictive uncertainty decomposition revealed a systematic shift from epistemic-dominated uncertainty at short lead times to aleatoric-dominated uncertainty at longer horizons. This shift highlights distinct management pathways: improving short-term forecasts via denser monitoring data versus addressing irreducible stochasticity for longer horizons. These findings demonstrate that integrating recurrent networks with MC dropout offers a practical, uncertainty-aware framework for probabilistic water quality forecasting in eutrophic lakes.