AutoWP: Automated wind power forecasts with limited computing resources using an ensemble of diverse wind power curves

Forecasting the locally distributed Wind Power (WP) generation is crucial for future energy systems, demanding scalable WP forecasting models to keep pace with the increasing number of smart grid applications. Therefore, we propose AutoWP, which is a weighted ensemble of WP curves that represent different site conditions. This representation is achieved by using a diverse set of WP curves from Original Equipment Manufacturers (OEMs) and optimizing their contribution to the weighted sum of curves using the least squares method. AutoWP is advantageous since physical limitations in WP generation are implicitly reflected in the considered WP curves, and the optimization to find the ensemble weights requires only a small amount of data and computational effort. Furthermore, AutoWP uses a rule-based and filter-based approach for training data cleaning to identify and filter samples representing abnormal operation due to shutdowns or partial load operation. It is shown on a real-world data set that AutoWP achieves competitive day-ahead forecasting performance against other methods based on WP curve modeling and outperforms the autoregressive Deep Learning (DL) method Temporal Fusion Transformer (TFT), which is based on the attention mechanism.