基于短时傅里叶变换和两阶段深度迁移学习的多频段振荡源定位OA北大核心

With the large-scale integration of new energy,represented by wind and solar power,into the system and the use of new governors of high-power turbine units,the oscillations in the new power system have expanded from the traditional low frequency oscillation to multi-frequency band oscillation.Accurately locating the oscillation source is a key means to suppress the expansion of adverse effects.Thus a novel location method based on short-time Fourier transform(STFT)and two-stage deep transfer learning is proposed.In this method,the active power measurement signals of all generators are converted into time-frequency representation matrices by STFT processing,and the matrices are transformed into feature images by linear mapping,so that the location problem is transformed into an image classification problem.The feature images are then fed into a ResNet50-based two-stage classifier.The first stage is used to determine the type of oscillation,while the second stage is used to locate the source.Transfer learning integrated with image knowledge learning is adopted to further improve the training efficiency and localization accuracy.Simulation results for the New England system with wind power and the Hubei power grid show that,compared to the support vector machine,decision tree and single-step transfer learning method,the proposed method has higher accuracy and robustness in the presence of noise.