Research Of Harmonic Measurement Method Based On Neural Network

On the basis of introduction of the current harmonic measurement methods, the harmonic measurement methods based on artificial neural network (ANN) are investigated and the algorithms are simulated and proved by MATLAB. The applications in harmonic measurement for three neural networks: adaptable neural network. BP (Back Propagation) neural network, RBF(Radial Basis Function) neural network are mainly researched.Firstly, for adaptable network, a single neuron based network structure is adopted in this paper. The study results show that for the power system with unknown harmonic sources, this method is not only with simple circuit structure, but also with relative good real-time and high precision. However, for the power system with heavy sudden harmonics loads, its real-time is not high enough needing 0.04~0.08s to track. In order to improve the real-time and precision, the methods with multi-layer feed-forward neural networks are investigated more.The measurement methods based on two multi-layer feed-forward neural networks: BP network and RBF network are analyzed and studied. In order to improve the measurement precision, the system using the software method to calculate the harmonic phase and usingthe neural network to measure the harmonic amplitude. The practical examples are simulated which prove the networks" measurement precision and real-time. Finally, the network performances between these two networks are compared by simulation. The simulation results show that RBF network are better than BP network on training speed, approximation capability and the generalization property. Under the condition of the same network scale (with 20 hidden layer cells), for RBF network, the iterative error is less than 2.2E-5 after 20 iterative calculations, and for BP the iterative error is about 1E-4 after 3000 iterative calculations. The simulation process also shows that the train time for BP network is much longer than RBF network. On the generalization property, the measurements for the untrained samples show that the most error for RBF network is 1%. while for BP network is 5.77%. For the system with certain power electronics equipment, the harmonic measurement method with multi-layer feed-forward neural network can get good performance on both real-time and precision.