一种基于多层前向神经网络的谐波检测方法

A Harmonic Measuring Approach Based on Multilayered Feed Forward Neural Network

谐波幅值和相位是有源滤波中的两项关键检测参数,两者均可由人工神经网络实现非线性映射。提出了一种用多层前向神经网络(MLFNN)来同时实现对两项参数进行测量的方法,并构造了一隐层采用正切函数,输出层采用线性函数的三层前向神经网络来进行仿真,以3次谐波为例阐述了该神经网络的训练方法和训练样本的组成。利用Matlab提供的工具箱,先用训练样本对神经网络进行训练,然后测量构造的未训练样本,其结果验证了方法的有效性。与传统FFT谐波检测方法的仿真比较表明:该方法在实时谐波检测中具有较高的精度和灵活性,且对采样数目没有严格限制,离线训练好的MLFNN可以适用于谐波源固定的场合。

Magnitude and phase angle of harmonics are two key measuring parameters in active power filter, both of them can be non-linearly mapped by ANN. A method is proposed to simultaneously measure the two parameters. A three layers feed forward neural network whose hidden layer adopts tangent function and the output layer adopts linear function is constructed to simulate the method. The training method and training sample composition in the neural network are presented using 3rd harmonic as an example. Training the neutral network with training samples first, then measuring untrained samples, the simulation is performed by using Matlab＇s toolbox and the results show the validity of the proposed harmonic measuring approach. Comparison with traditional FFT harmonic measuring method shows that the proposed method has higher precision and flexibility in real time harmonic measuring and the proposed method has no restrict limitation to the samples number. The off-line trained MLFNN may suit for the occasion where the harmonic source is constant.