Advancing weather predictions for offshore wind farm maintenance through deep learning

Abstract

Historical met-ocean data are widely used in the decision support tools to evaluate different operations & maintenance (O&M) strategies for offshore wind energy. Although effective, they are often very limited, which may not be able to represent prolonged offshore weather conditions at the wind farm sites. This hinders their application to the O&M planning. In this paper, a deep learning approach is proposed to build up the stochastic weather generator, in which the long short-term memory neural network is leveraged to simulate wind and wave time series data. The neural network is trained using the (limited) historical met-ocean dataset to accurately capture the statistical characteristics of wind and wave conditions. The results demonstrate the effectiveness of the proposed stochastic weather generator in delivering both open-loop and closed-loop forecasting for wind speed and significant wave height data, thereby supporting both corrective and preventive maintenance activities. The case study reveals that the open-loop forecast excels in short-term hourly met-ocean parameter predictions, while the closed-loop forecast proficiently captures the met-ocean patterns within a predefined window. Although the closed-loop forecast for wave parameters generally follows the measurements trend, it diverges from the actual measurements; a discrepancy likely due to the complex spatial-temporal dynamics of waves not completely captured by the LSTM model. The proposed LSTM model, considered as a complementary but connected solution, is able to enhance the utility of the limited historical data in O&M planning.