Research On Power Quality Disturbances Detection Based On Deep Learning

Power quality is closely related to the national economy,and is the core of energy management of power plants and public utility companies.Continuous and stable electricity supply can reduce equipment damage accidents,thus improving the economic benefits of the company.Therefore,it is of very importance to accurately monitor all kinds of power quality disturbances in power system and improve the detection accuracy and speed of power quality disturbances to ensure the safety and stability of power system.According to the power quality standard and related researches,this work summarizes the mathematical model of power quality disturbance signals,and simulates a large number of experimental samples by M atlab.Two power quality disturbances detection models based on deep learning method are proposed,and the effectiveness of the proposed models is verified by simulation and actually measured signals respectively.The main contents of the thesis is as follows:(1)An end-to-end Simple Gated Recurrent Network(SGRN)model is proposed,which can quickly and accurately complete power quality disturbances detection without preprocessing and empirical feature extraction operations.The core of the proposed model is a group of new recurrent units,which are only composed of "forget gate" and "input gate",and can retain long-term memory.Moreover,the number of model parameters(i.e.memory cost)and disturbances detection speed are better than those of the Long-Short Term Memory(LSTM)model and Gate Recurrent Unit(GRU)model,which makes it easier to deploy in the resource constrained edge device microcontroller.(2)Combining One-Dimensional Convolutional Neural Network(1DCNN)and Denoising Auto-Encoder(DAE),a 1DCNN-DAE model is proposed to extract the features of power quality disturbances.The model is trained with the minimum reconstruction loss as the goal.The features of disturbances can be automatically ed in the encoder part,and the original signals can be reconstructed in the decoder part.The experimental results show that the model has a good reconstruction performance,and can provide new ideas for the research on the compression and transmission of the disturbance signals.(3)Combining 1DCNN-DAE with SGRN,a power quality disturbances detection model named 1DCNN-DAE-SGRN is proposed,which improves the disturbance detection performance of the SGRN method in a noisy environment and enhances the robustness of the algorithm.The model automatically extracts the disturbance features with good noise immunity in the auto-encoding stage,and then these features are input into SGRN for disturbances detection.The experimental results show that the detection accuracy of SGRN model will decrease significantly when the noise in the test set increases gradually,while the average detection accuracy of 1DCNN-DAE-SGRN model can still reach 98.74%.