Improving the forecast of wind speed and significant wave height using neural networks and gradient boosting trees

Abstract

This research aims to enhance the accuracy of significant wave height (SWH) and wind speed (WSP) predictions over extended forecast lead times. Two forecasts are used as test cases of the methods developed: the European Centre for Medium-Range Weather Forecasts (ECMWF) and the Global Ensemble Forecast System – Wave (GEFSWAVES). Each model considers 15 and 16 days of forecast, respectively. To improve the predictions of the ECMWF and GEFSWAVES, the residuals representing the differences between those forecasts and the observations from buoys near the coast of the Iberian Peninsula are estimated using machine learning (ML) algorithms. The estimations are then used to correct the predictions of the numerical models. Five ML models are used: long short-term memory (LSTM), gated recurrent units (GRUs), XGBoost, LightGBM, and CatBoost. The results demonstrate that the last three models significantly outperform the first two, achieving notable improvements in forecasting the two variables across all forecast days, with reduced success as the forecast times increase. CatBoost and LightGBM produce by far the best results, with impressive improvements in prediction ability for both variables according to most metrics, except the bias. LSTM and GRUs have not been very successful in improving SWH's forecast. • This research aims to enhance the accuracy of significant wave height and wind speed forecasts over extended lead times. • Two forecasts are used ECMWF and GEFSWAVES with 15 and 16 days of forecast, respectively. • The residuals from buoys near the coast of the Iberian Peninsula are estimated using machine learning algorithms. • Five ML models are used: long short-term memory (LSTM), gated recurrent units (GRUs), XGBoost, LightGBM, and CatBoost. • CatBoost and LightGBM produce by far the best results, with impressive improvements in prediction ability.

Topics & Concepts