Predicting Bioluminescent Algal Blooms by Conditioning Neural Forecasts on Reconstructed Dynamics

Forecasting rare events in nonlinear environmental systems is difficult because observations are sparse, important state variables are unobserved, and neural networks must infer both the underlying dynamics and the forecast from limited data. Episodic bioluminescent red tides along the Southern California coast provide a demanding real-world test of this challenge. These events arise from blooms of dinoflagellates such as Lingulaulax polyedra, whose chaotic dynamics limit the effectiveness of many conventional statistical forecasting approaches. Here we condition a convolutional long short-term memory (CNN–LSTM) model on state-space embeddings derived from Empirical Dynamic Modeling (EDM). These embeddings preserve local attractor geometry and provide a mechanistic multivariate dynamical coordinate system for machine learning. Applied to bloom prediction, this EDM-CNN-LSTM framework improves mean PR-AUC from 0.171 to 0.552 relative to a standard LSTM and shows more consistent convergence. More generally, the results suggest that EDM-conditioned neural models more reliably capture the underlying attractor structure, leading to improved robustness and predictive performance in chaotic, partially observed systems, such as those associated with algal blooms, harmful (HABs) and otherwise.