A comparison of LSTM-based Ensemble models for wind speed prediction in the Atlantic Ocean using optimised parameters

Abstract

Climate change represents one of the biggest challenges for the planet in this century. Its impact has caused significant variations in long-term wind patterns. The consequences can be devastating for people living near the coasts, especially in the regions around the Atlantic Ocean, which have experienced some of the most intense tropical cyclones worldwide. This paper studies the application of ensemble models for improving the performance of wind speed prediction of Atlantic storms. We have applied two ensemble models based on Long Short-Term Memory (LSTM), one also using Autoregressive Integrated Moving Average (ARIMA) and another also using Random Forest (RF) and Support Vector Regression (SVR). The models are used in four scenarios involving different combinations of variables considering their correlations. We examine the validation loss, training loss, R squared, NSE, KGE and BF using optimised parameters for each scenario and find that the combination of pressure, latitude and longitude are the reliable predictors for the models. As a result, the ensemble model LSTM -ARIMA can reduce the RSME of LSTM by 12% and the MAE by 11% but it does not fit completely the testing data while the LSTM-RF-SVR shows a reduction of more than 50% for RSME and MAE which indicates an improvement of LSTM performance as well as its capacity to predict complex patterns. From our research, we have demonstrated the optimization of parameters for the ensemble model using Deep Learning and Traditional Machine Learning to achieve improved prediction and avoid bias of wind intensity.