A Multiple Linear Regression Approach for Storm Surge Predictions in the New York/New Jersey Bight

Storm surge is the meteorological forced portion of the sea level that is dynamically generated by winds and pressure perturbations associated with the passage of tropical and extra-tropical cyclones. Often numerical hydrodynamical models (NHM) are forced with meteorological inputs (i.e. NAM or WRF) to generate spatial forecasts of storm surge. Since storm surge is directly forced by wind and pressure from the atmosphere, the accuracy of storm surge forecasts hinges on the accuracy of meteorological inputs. As such, it has been shown that probabilistic forecast guidance involving ensembles results in more skillful predictions of storm surge compared to deterministic forecasts. However NHM's are computationally expensive to run, especially in ensemble coupled atmosphere-ocean simulations. This study proposes a statistical method to predict storm surge using inputs from atmospheric models.

A multi-linear regression (MLR) approach was used to train the statistical model for storm surge using two different atmospheric reanalysis datasets (NARR and CFSR) and observed storm surge data at three stations along the NY/NJ Bight (Atlantic City, New Jersey, The Battery, New York, and Montauk Point, New York). The odd years between 1979-2012 are used to train the 3-hourly storm surge predictions during the cool season months (Oct 1-March 31), while the even years were used as verification. The predictors of the MLR represent prolonged surface wind fetch and sea-level pressure perturbations. This approach also works well for more extreme events, such as hurricane Sandy (2012). The MLR predicted a 2.56 m peak surge at The Battery using the NARR, while the observed peak was 2.68 m. Overall, the MLR is shown to be approximately unbiased. MLR predictions are also shown to have similar mean absolute error compared to an operational hydrodynamical model (SIT-NYHOPS) between 2010-2014 at The Battery, New York.

Since it is inexpensive to run the MLR approach, it is used to generate ensemble storm surge predictions (http://nybightstormsurges.weebly.com) at The Battery. The NCEP Short-Range Ensemble Forecast System (SREF) is used to force the statistical model to generate 9 members at 0900UTC and 0300UTC daily. Basic metrics are calculated, such as mean and spread, which a forecaster can relate to the ensemble predictions of wind and pressure on this same web page. Other applications include applying MLR to global climate model data for long-term (~100 y) 3-h projections of storm surge at selected stations in the NY/NJ Bight.