Research On Islanding Detection Based On Deep Learning

Due to the continuous reduction of fuel energy and the increasing harm to the environment,people have turned their attention to renewable energy sources such as solar energy and wind energy,which are clean and have huge reserves.Distributed generation technology based on renewable energy is an important development in the future of the power industry.direction.Islanding is a fundamental problem in distributed generation systems.After the island is generated,it may threaten the safety of the grid maintenance personnel and cause damage to the electrical equipment during reclosing.Therefore,island detection is a protection function that distributed generation devices must possess.This paper first studies the causes of the island phenomenon,and analyzes the changes in the parameters of each microgrid when an island occurs,and selects the voltage waveform of the grid point to determine whether an island has occurred.In view of the existence of detection blind spots in the current passive islanding detection methods and the impact of active islanding detection methods on the operation of microgrids,an islanding detection method based on deep learning is proposed.Then,using MATLAB to build a micro grid model containing distributed photovoltaic,lithium battery energy storage systems,large power grids and loads.Due to the investment and withdrawal of some equipment of the microgrid,it will impact the voltage of the grid point and affect the determination of islanding detection.Therefore,this paper simulates the voltage waveforms and specific values of the grid point under different conditions such as grid-connected operation,photovoltaic start-up and islanding,and is used for the training of the island detection model.Finally,from the construction of network structure,selection of activation function,choice of loss function and optimization function,and learning rate hyperparameter adjustment strategy,an islanding detection model based on deep feed-forward neural network was built.Results show that when the waveform sampling period of the voltage at the grid point is50?s,the model can detect the occurrence of islanding at 75 ms.When the sampling period is25?s,it can detect the occurrence of islanding at 120 ms.In theory,the accuracy of the model to determine whether the system is in an island state has reached 100%.