Deep learning decodes multi-horizon dynamics and probabilistic risks of harmful algal blooms.

Harmful algal blooms are expanding globally across freshwater ecosystems due to climate change and accelerating eutrophication, posing escalating threats to public health and economic stability. Robust early warning requires predictive frameworks that are accurate at hourly resolution, uncertainty-aware, and mechanistically interpretable. Existing process-based models are limited by sparse parameterization, while statistical and deep-learning approaches typically operate at daily resolution, produce only point predictions, and rely on threshold-based classifications or post-hoc explanations that fail to capture high-frequency temporal dynamics or quantify predictive uncertainty, thereby hindering proactive intervention and masking hidden ecological risks. Here we present BloomNet, an intrinsically interpretable deep neural network architecture configured for multi-horizon quantile forecasting, resolves these limitations by integrating future environmental covariates with historical observations from a hyper-eutrophic lake. Evaluated on a four-year hourly record, BloomNet achieves exceptional predictive stability across 24-, 48-, and 72-h horizons (