Short-term prediction of algal dynamics in freshwater under meteorological variability: insights from high-frequency vertical observations and hybrid modeling

Harmful algal blooms (HABs) increasingly threaten freshwater ecosystems under intensifying anthropogenic and climatic pressures. Accurate short-term forecasting of HABs remains challenging, particularly in shallow lakes where phytoplankton dynamics are strongly influenced by meteorological variability through its effects on mixing intensity, thermal structure, and light availability. These rapid, non-stationary processes play a critical role in the development and decay of algal blooms, yet they remain poorly resolved by conventional low-frequency monitoring and are often oversimplified in predictive models relying on temporal persistence alone.

In this study, we investigate seasonal algal dynamics in Lake Taihu, a large shallow lake in China, using high-frequency vertical profiling data acquired by an autonomous monitoring system, providing a high-resolution dataset comprising 3 to 5 readings per second, and moving 2–3 cm per dataset, including observations of water temperature, conductivity, dissolved oxygen, pH, colored dissolved organic matter, chlorophyll-a, phycocyanin, and underwater photosynthetically active radiation, together with concurrent meteorological forcing including wind, air temperature, atmospheric pressure, and precipitation. This unique combination enables the explicit characterization of diel to seasonal variability in vertical water-column structure under changing meteorological conditions.

To extract spatiotemporal patterns from these heterogeneous observations, we apply a hybrid deep learning framework that integrates convolutional, recurrent, and attention-based components to predict short-term vertical chlorophyll-a dynamics. Rather than relying purely on autoregressive persistence of biomass, the process-guided model (Phytoformer) is designed to learn the influence of physical drivers associated with wind-driven mixing, stratification, and light attenuation, thereby enhancing ecological interpretability and physical consistency. High short-term predictive skill based on biomass persistence does not necessarily imply an understanding of the environmental drivers that govern bloom intensification or decay. Feature relevance analyses further indicate that physical controls modulate phytoplankton dynamics beyond short-term state persistence, with distinct seasonal patterns.

Our work demonstrates the potential of integrating high-resolution vertical sensing with interpretable deep learning to improve short-term prediction and early warning of HABs across seasons. Ongoing work extends this hybrid modeling framework to deep stratified Wahnbach Reservoir in Germany, where HABs can bloom in specific depth layers under contrasting water quality regimes. This cross-system application aims to explore model generalizability and to identify how dominant physical drivers differ between shallow and deep lake environments.