Abstract:
|
In influenza forecast, error structure in state space emerges naturally from the nonlinear dynamics of epidemic models. Although a number of data assimilation methods have been applied to forecasting seasonal outbreaks of influenza, the value of error structure between the states and observations has been largely overlooked. Here we introduce a procedure for diagnosing structural errors using the breeding method and then correct these errors to further improve forecast quality. Using a humidity-driven influenza model, we show that correcting structural errors can effectively recover perturbed trajectories given other non-perturbed states are accurate. To examine its efficacy in a model-data assimilation influenza prediction system, we generate retrospective ensemble forecasts in conjunction with the structural error correction for synthetic observations of influenza in New York City. Remarkably, in combination with the data assimilation technique that constrains model state variables close to the truth, the correction procedure substantially improves the forecast accuracy of outbreak peak timing several weeks in advance of the actual peak.
|