| The inter-turn short circuit fault in excitation winding is the most common type of generator faults.When the fault is mild,it has little impact on the operation of the generator.However,when the fault is serious,it will increase the excitation current,reduce the generator’s reactive power,and cause a series of problems such as generator vibration.In this paper,a deep learning based synchronous generator excitation winding inter-turn short circuit fault early warning method is proposed.By diagnosing the excitation winding mild inter-turn short circuit fault,the fault warning effect is achieved.(1)An early warning model,which is composed of Convolution Neural Network(CNN)and Deep Belief Network(DBN),of short circuit faults in the field winding based on the classification mode is constructed.The model uses CNN as the extractor to extract the input data features,and DBN as the classifier to classify the extracted feature data.The generator’s operating state is divided into the normal state and the field winding short-circuit fault state.The appropriate network structure is selected by the control variable method,and the trained model is verified by Receiver Operating Characteristic curve.(2)Long Short-Term Memory-Convolution Neural Network(LSTM-CNN)and Gated Recurrent Units-Convolution Neural Network(GRU-CNN)fitting model are constructed.The excitation current is selected as fault judgment indicators,and the Improved Particle Swarm Optimization method is used to determine the optimal structure of the network.The model with better performance is selected and the offset between the model output value and the actual measured value is used to realize the inter-turn short circuit fault warning of the excitation winding.(3)The visual interface of the early warning model for inter-turn short circuit fault in excitation winding is built.The interface includes four parts: fitting model training part,classification model training part,fitting type fault warning part and classification type fault warning part.The fitting model training part is used to build two fitting deep learning networks,LSTM-CNN and GRU-CNN.The classification model training part is used to build a deep learning network which is composed by CNN and DBN.The fitting type fault warning part and classification type fault warning part can call the trained fitting model or classification model.Importing real-time data after setting the relevant parameters,these parts can realize the diagnosis of the inter-turn short circuit slight fault of the excitation winding of the generator.The above models are verified by experimental data,and all have a high fault diagnosis accuracy rate,and have good guiding value for engineering applications. |
