Use of turbine-level data for improved wind power forecasting

Abstract

Short-term wind power forecasting is based on modelling the complex relationship between the weather forecasts and wind farm power production. To date, efforts to improve wind power forecasts have focused on improving Numerical Weather Prediction and wind farm power curve models. However, utility-scale wind farms cover large areas meaning that a single power curve model may struggle to represent the collective behaviour of large numbers of wind turbines. Contemporary statistical techniques are capable of processing large volumes of data, enabling the assimilation of measurements from individual wind turbines to construct a more detailed representation of wind farm power generation. Here, three state-of-the-art techniques are applied to forecast wind farm power production 1) directly from numerical weather predictions, and 2) by aggregating forecasts for individual wind turbines. Furthermore, it is observed that some wind turbines are better predictors than others and an aggregation process based on conditional weighting is proposed. In case studies of two large wind farms in the UK, it is shown that wind farm power forecasts comprising a conditional weighted sum of turbine-level predictions are superior to a direct wind farm forecast for horizons up to 48 hours ahead. Specifically, performance of the best-performing benchmark, the gradient boosting machine, is improved by 12% at Clyde South wind farm and by 6% at Gordonbush.