Multi-scale complex terrain wind power output prediction based on multi-feature refinement fusion

Reliable wind power output (WPO) prediction is a security guarantee for high penetration power systems from wind power. Currently, the utilization of artificial intelligence prediction models in WPO prediction does not fully exploit data features, leading to a lack of accuracy in output prediction. To improve the accuracy and robustness of WPO prediction, this study proposes a model based on K-Medoids-temporal convolutional networks-gated recurrent unit (K-Medoids-TCNGRU) for multi-feature refinement fusion, considering multi-scale complex terrain. Firstly, the raw data is divided into multiple spatial scales utilizing a multi-scale approach. Secondly, dual channel input of weather data and historical WPO data. Thirdly, weather data are divided by the K-Medoids clustering into two separate classes for prediction utilizing GRU and historical WPO for prediction utilizing TCN. Finally, the prediction results are fed into a fully connected neural network for decision-level fusion. In addition, this study replaces all dropout layers with Monte Carlo dropout layers and averages the multiple predictions to improve the certainty of the results. A dataset provided by a region of Natal, Brazil, is utilized to predict the WPO of the region for a day. Experimental results show that K-Medoids-TCN-GRU has a 17.4421 lower root mean square error than the highest performing GRU of the comparison model, validating the K-Medoids-TCN-GRU as more accurate and robust.