神经网络法与直接计算法在APF谐波电流检测中的应用比较

Comparison between nerve network method and direct computation method in harmonic current detection for APF

神经网络法和直接计算法是检测原理相同的2种有源电力滤波器谐波电流检测方法。首先,对这2种方法的复杂性、计算量、检测精度和实时性进行了比较,结果表明直接计算法在这些方面都明显地优于神经网络法。然后,对负载电流突然增加、线性增加、按指数规律减小和含有尖峰干扰时的4种情况进行了仿真比较,比较表明:当神经网络法的误差非常小时,2种方法计算出的基波有功电流幅值的仿真波形完全相同(即2种方法的静态与动态特性完全相同),这是因为它们的检测原理完全相同,神经网络法的自学习和自适应功能无从体现。对于神经网络法,负载电流的幅值和负载电流在1个周期内的采样个数都不能取得很大,否则指数函数的计算结果将发生溢出,而直接计算法则不会出现这种情况;神经网络法需要设定初值,而直接计算法不需要设定初值。

The NNM （Nerve Network Method） and DCM（Direct Computation Method） of harmonic current detection for active power filter have same principle. Comparison is carried out between them in complexity,computation load,precision and real-time performance. It is found that,DCM is obviously better than NNM in all these aspects. The simulative comparison in step change and linear increase of load current,exponential decrease of load current and sharp disturbance shows that,if the error of NNM is set very small,the calculated amplitude of fundamental active currents are completely same,i.e. ,their static and dynamic performances are completely same because of the same detection principle,and the adaptive function of NNM could not be brought into play. Different from DCM,the amplitude of load current and its sampling frequency of NNM should not be proper to avoid calculation overflow and the initial values are necessary.