Feature averaging of historical meteorological data with machine and deep learning assist wind farm power performance analysis and forecasts

A novel feature-averaging technique of meteorological input data from recent past hours assists wind power generation forecasts from individual wind farms by capturing recent changes in conditions. Seven machine learning models and one deep learning model are configured to evaluate that data averaging technique using published 13,140 hourly data records for each of seven wind farms over an 18 months period. The datasets involve just four historically forecasted weather variables averaged in 12-hourly intervals over the previous 48 h for each hourly power record. This generates 16 time-related input variables for each hourly record. The convolutional neural network (CNN) performs best in training and testing on a supervised learning basis. However, the Adaboost (ADA) model is most accurate for semi-supervised forecasting (RMSE averages 0.177). The ADA achieves forecasting accuracy of RMSE = 0.154 on a t + 1 to t + 30 h ahead basis, outperforming on average, a seasonal autoregressive integrated moving average (SARIMA) model (RMSE = 0.183), trained with the univariate historical hourly power generated. The SARIMA model proved most accurate in forecasting t + 1 to t + 4 h ahead, whereas the ADA model provided more accurate forecasts for most of the hours in the t + 5 to t + 30 interval spread over a 12-month period. By adding additional variables the accuracy of the ADA model could potentially be further improved. Nevertheless, the results presented highlight the effectiveness of applying the proposed time-averaged feature selection technique.