Data-aware forecast of harmful algal blooms with model error

Harmful algal blooms (HABs) adversely affect human health, lake biotic life, surrounding ecosystems, and water use. Effective HAB prevention strategies require a coordinated effort by all stakeholders and need a comprehensive assessment of historical HAB events and underlying factors, short-term forecasting, and on-site lake management. We contribute to this effort by introducing a modeling framework for short-term HAB forecast. The framework relies on a mechanistic (process-based) cyanobacteria-growth model with error and deploys two alternative approaches, Kalman filter (KF) and Gaussian processes (GP), to estimate the model error from low-cost daily observations of bacteria concentration, temperature, phosphorus, nitrogen, and irradiance. We deploy the two variants of our framework to forecast HABs in Cheney Reservoir, South Central Kansas, for which high-frequency in situ historical data are available. Both approaches yield accurate three- and seven-day forecasts, with the prediction accuracy decreasing with the forecast duration. The GP variant yields more accurate predictions than its KF counterpart, but has wider confidence intervals and is computationally more expensive. Both variants significantly outperform a model-free (statistical) approach based on autoregression of daily observations. The KF and GP models have higher predictive accuracy than a statistical model and machine learning techniques (a recurrent neural network and a decision-tree algorithm) trained on small datasets: they reduce relative error by at least 50% and increase the coefficient of determination by 70%.