A.V. Kalyuzhnaya - Post-graduant Student, National Research University of Information Technologies, Mechanics and Optics
А.V. Boukhanovsky - Dr. Sc. (Eng.), Professor, National Research University of Information Technologies, Mechanics and Optics

This paper offers generalized approach to ensemble forecasting in terms random spatial-temporal field and its utilization for storm surges forecasting in Saint-Petersburg. Traditionally, multi-model ensemble forecast could be represented as weighted sum of ordinary dynamical forecasts. Approach described in this research makes possible to combine ensemble members with different forecast times and spatial resolution, and to take into account differences of grid data. The main character of flood forecast is a sea level, that-s why ensemble forecast expression was subdivided into two parts for level and currents. These expressions were represented as linear dynamical systems in each point of field. For investigation of offered method, numerical experiments were realized. For this purpose we took expression for sea level in point Kronstadt. As members of ensemble were used two alternative hydrodynamic models: two-dimensional model BSM, three-dimensional baroclinic model Baltp. As meteorological data sources atmospheric forecasts GFS, HIRLAM and regional forecasts built using WRF model, were taken. For numerical experiments we used problem-oriented environment based on CLAVIRE platform. For identification parameters of dynamical system model results and sea level data for December 2011 were utilized. Coefficients identification for each next forecast is performed with moving data sample with different lengths (from 2,5 to 5 days). Results showed, that variability of error of ensemble forecast with different lengths of moving window negligible compared to error eliminated due to the procedure of ensemble forecasting itself. Also, efficiency of different model combinations was estimated. Results demonstrate possibility of flexible choice of data sources among available variants keeping adequate efficiency; it helps to improve forecast quality even having a minimal set of available models. These results provide improvement in system fail-safety due to capability for work with incomplete set of ensemble members.