Deep Learning for Offshore Wind Intensity Forecasting in the South Atlantic: A ConvLSTM Approach Using ERA5 Data

Resumo

Accurate wind forecasting is crucial for optimizing the efficiency and operational planning of offshore wind energy projects. In Brazil, there is growing interest in offshore wind energy as companies seek to diversify their renewable energy portfolios. Numerical weather prediction (NWP) models, although widely employed, often require extensive computational resources and significant time to provide rapid short-term forecasts, limiting their applicability for near real-time decision-making. In addition, the scarcity of in situ measurements and the limited frequency of remote sensing data in the South Atlantic, particularly in the region of interest leads to a dependency on data generated by NWP models, such as ERA5, for training deep learning models. To address this challenge, this work proposes and evaluates a deep learning-based forecasting framework utilizing Convolutional Long Short-Term Memory (ConvLSTM) networks trained exclusively on ERA5 reanalysis data. ERA5 provides comprehensive atmospheric data globally, and in this work a sample of eight years of data was used, with intervals of 3 and 6 hours. This study explores various input configurations, including different lookback periods ranging from 24 to 72 hours to forecast future wind intensity for periods between 24 to 48 hours ahead. The predictive performance of the ConvLSTM model is compared when using wind intensity alone versus models incorporating additional atmospheric variables such as temperature and pressure. Model performance was rigorously assessed using Root Mean Square Error (RMSE), evaluating predictions of wind intensity. The preliminary analysis demonstrates promising performance of the ConvLSTM model in capturing complex spatial-temporal interactions. This research contributes valuable insights into effective ConvLSTM configurations and underscores the model’s potential advantages over traditional NWP approaches, particularly regarding rapid forecasting capabilities and computational efficiency, indicating their potential for improving short-term offshore wind forecasting in the South Atlantic region.