Superposition Graph Neural Network for offshore wind power prediction

Wind power prediction plays an important role in its utilization. Currently, in machine learning methods and other traditional methods, the prediction is always based on the time series of data nodes, and sometimes wind turbines near the predicted nodes are also applied. These methods have limitations in the utilization of the spatial features of the entire wind farm, and can only be used to predict a single wind turbine. Offshore wind farm data is more difficult to predict due to the more dispersed distribution of wind turbines and the intermittent nature of offshore winds. We proposed a data integration method, which can connect all wind turbines in a certain range of wind farms by their geographical locations and other related information to form a graph(one type of data structure), then superimpose these graphs in a certain period of time. Then, we proposed the SGNN(Superposition Graph Neural Network) for feature extraction, which can maximize the use of spatial and temporal features for prediction. In the four offshore wind farms used in experiments, the mean square error (MSE) of the method is reduced by 9.80% to 22.53% compared with current-advanced methods, and the prediction stability of the method has also been greatly improved.