Filtros adaptativos no domínio da frequência para correção de previsão climática da temperatura da superfície do mar global

Abstract

This dissertation presents results of the application in the frequency domain of Recursive Least Squares (RLS) adaptive filters for the correction of global climate prediction errors of the Brazilian Earth System Model (BESM). The BESM consists of computational models aimed at generating climate change scenarios, just like the ECMWF (European Centre for Medium-Range Weather Forecasts) which also develops and produces global weather predictions such as ERA5 (ECMWF Reanalysis v5). The performance of BESM is measured with reference to the ERA5 observational dataset, and the difference between climate variables generally increases over time. BESM prediction is corrected using ERA5 observational data as a reference. The main objective is to correct the global climate prediction errors of the BESM model. Filters are designed using the previous prediction sequences for a cell (a small region of the globe where correction is applied) or in the frequency coefficients obtained via two-dimensional Discrete Cosine Transform (DCT-2D) in a cell neighborhood. DCT is a mathematical technique for analyzing periodic signals in the frequency domain with a sum of cosine functions. The filters are designed using the Recursive Least Squares (RLS) adaptive algorithm. Filters of different orders are tested, and neighborhoods of different sizes are considered. The correction methodology is also tested considering corrections seasonally and annually, separately. The performance of BESM prediction correction is evaluated using the Root Mean Square Error (RMSE) and Largest Absolute Error (LAE). The proposal is applied to the prognostic variable of Sea Surface Temperature (TSM) for regions of dimension M × M, with M = 8, 16, 32, and filters of length L = 4. Among the various zones analyzed, such as the Pacific Ocean, the Atlantic Ocean, the Indian Ocean, and the Arctic and Antarctic Oceans, the best performance of the proposed method is achieved in the first three zones, while the worst is observed in the last two zones. Finally, the best results are obtained for correction by seasons. Compared to the Mean and Variance Correction (CMV) method, the use of RLS filters generates better results for seasonal adjustment. However, in some specific geographic locations, CMV sometimes outperforms adaptive filtering when it comes to annual adjustment.