基于贝叶斯图卷积神经网络的风电场内多风机风速概率预测

Wind speed is characterized by randomness and volatility, which makes the point forecast results cannot quantify the uncertainty of wind speed. Therefore, probabilistic wind speed forecasting is a hot topic in current wind speed forecasting research. In wind farms with large turbine intervals and irregular rows, the local wind speed of each turbine has significant differences. However, the traditional wind speed probabilistic forecasting models tend to treat the wind speeds in an area as the same value, which is not conducive to the detailed control of wind farms. To solve the above problems, this paper proposes a probabilistic wind speed forecasting model for multiple wind turbines based on Bayesian graph convolutional neural network (BGCNN). Firstly, the original graph model of wind field was constructed based on the historical data of wind speed and wind direction of each turbine. After that, the wind field graph model was inputted into BGCNN, which can parameterize the graph model to change the graph information and extract the multi-turbine spatial correlation. In this process, the long and short-term memory network (LSTM) was plugged into the back of GCNN module in the BGCNN model, so that the feature matrix containing spatial information among turbines was inputted into the LSTM for temporal feature extraction, and the wind speed forecast results of multiple turbines considering the time-space two-dimensional correlation were obtained. Finally, the Monte Carlo method was used to sample multiple sets of model output values to achieve probabilistic wind speed forecasting for multiple turbines. Simulation results on the wind farm data show that, multiple turbines in a wind farm present different graph structures under different epochs, indicating that the proposed model can fully consider various spatial relationships among different turbines. For point forecasting results, which are obtained by calculating the mean of the sampling results, the mean absolute error of the proposed model is 1.092 m/s for all 100 wind turbines, and the root mean squared error among them is 1.530 m/s. There are both smaller than the comparison model, BCNN-LSTM and Q-CNN-LSTM. For probabilistic forecasting, the reliability and sharpness of the proposed model are not all the best at different confidence level, such as in the case of 30%, 50% confidence level, etc. Nevertheless, the average forecast interval coverage bias index and average sharpness of the forecast results are 3.10% and 0.119, respectively, which are lower than the two comparison models. For turbine number 33, compared with the single turbine forecasting model, the probabilistic forecasting results of the proposed model are the best, with the smallest forecast interval coverage bias index, sharpness and Pinball score, which are 1.68%, 0.123 ～ 0.431 m/s, respectively. The following conclusions can be drawn from the simulation analysis: (1) In this paper, the clustering method is used to classify the output features of the GCNN-LSTM model, which enables the BGCNN to be used in regression problems and to extract the spatial features among wind turbine data well. (2) Compared to the comparison models, the model in this paper achieves higher forecast accuracy in both point forecasting and probabilistic forecasting, and the forecast intervals calculated from the forecast results have small deviations and concentrated distribution. (3) For a single wind turbine in a wind farm, the forecast accuracy of this paper's model is higher than that of the probabilistic forecasting model only for a single wind turbine, which reflects that the reasonable and effective extraction of spatial features is conducive to the reduction of forecast errors.