Spatio-temporal wind speed forecasting with approximate Bayesian uncertainty quantification

The prediction of short- and long-term wind speed has great utility for the industry, especially for wind energy generation. Deep neural networks can be used to tackle this task by modeling the spatio-temporal behavior of the wind. In this work, Bayesian spatio-temporal wind speed forecasts are performed based on measurements collected from wind turbine data acquisition systems and predictions from widely used global climate forecasting models. Moreover, the resulting predictions are complemented by the quantification of the corresponding uncertainty, extracted via approximate Bayesian inference techniques. Such uncertainty is a valuable information in practical scenarios, such as turbine maintenance. The proposed solution is evaluated using real data collected from a wind farm in the South of Brazil. Different combinations of models and approximations are compared based on the achieved metrics and graphs of uncertainty calibration. The conducted experiments indicate that the use of recurrent convolutional neural networks (ConvLSTM) with the Deep Ensembles strategy provides the best results in terms of predictive distribution, which has the potential of assisting maintenance and operation in wind farms. The final results can be viewed as a novel way for wind farm performance teams to extract Bayesian wind speed forecasts with spatial information.