With the continuous optimization of energy structures, wind power generation has become the dominant new energy source. The strong random fluctuation of natural wind will bring challenges to power system dispatching, so it is necessary to predict wind power. In order to improve the short-term prediction accuracy of regional wind power, this paper proposes a new combination prediction model based on convolutional neural network (CNN) and similar days analysis. Firstly, the least square fitting and batch normalization (BN) are used to preprocess the data, and then the recent historical wind power data set for CNN is established. Secondly, the Pearson correlation coefficient and cosine similarity combination method are utilized to find similar days in the long-term data set, and the prediction model based on similar days is constructed by the weighting method. Finally, based on the particle swarm optimization (PSO) method, a combined forecasting model is established. The results show that the combined model can accurately predict the future short-term wind power curve, and the prediction accuracy is improved to different extents compared to a single method.

When multiple scattered wind farms are connected to the power grid, the meteorological and geographic information data used for power prediction of a single wind farm are not suitable for the regional wind power prediction of the dispatching department. Therefore, based on the regional wind power historical data, this study proposes a combined prediction method according to data decomposition. Firstly, the original sequence processed by the extension methods is decomposed into several regular components by Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). All the components are classified into two categories: fluctuant components and smooth components. Then, according to the characteristics of different data, the long short-term memory (LSTM) network and autoregressive integrated moving average (ARIMA) model are used to model the fluctuant components and the smooth components, respectively, and obtain the predicted values of each component. Finally, the predicted data of all components are accumulated, which is the final predicted result of the regional ultra-short-term wind power. The feasibility and accuracy of this method are verified by the comparative analysis.