Previsão da maré meteorológica com rede neural aplicada à estimativa do nível do mar na Baía de Guanabara – RJ

Abstract

Sea level (SL) in coastal regions is affected by several meteorological and oceanographic factors, especially in the southern and southeastern regions of Brazil. Conventional forecasting, based only on the astronomical tide, disregards variations induced by forcings of non-astronomical origin, which can generate significant deviations from SL. In this study, we propose the use of a recurrent Long Short-Term Memory (LSTM) neural network to predict the Meteorological Tide (MT) in Guanabara Bay - RJ and to estimate the SL from the sum of the astronomical and meteorological components. Wind and Mean Sea Level Pressure data from the ERA5 reanalysis were used at four points on the Brazilian coast, in addition to sea level data provided by the Brazilian Navy. The MT was obtained by the difference between the observed level and the predicted astronomical tide and was subsequently filtered in the 3.1 to 31.5-day band. The model had its architecture and hyperparameters optimized through the Random Search method. The prediction of MT with the LSTM network showed good performance, with a correlation coefficient (r) of 0.84 and RMSE of 6.5 cm. The reconstruction of the entire times series, with the sum of the astronomical tide to the MT predicted by the network, reduced the error by 43% compared with conventional forecast, with an r of 0.97 and RMSE of ~8 cm. The results highlight the potential of LSTM to improve short-term coastal forecasts and enhance forecasting accuracy in higher energy events.