Comparison of adaptive filter and neural network approaches to canceling current harmonics

The work done in this dissertation involves comparing two approaches for canceling unwanted harmonics. The first approach uses a neural network which has been trained with current harmonic data collected from several personal computers in a lab so that it could reproduce an accurate representation of the 60 Hz fundamental current. The second approach involves using an adaptive filter based on LMS algorithm, which adaptively tracks the input signal which is representative of the load current, including harmonics, and re-creates a replica of the 60 Hz fundamental. Two personal computers equipped with a/d and d/a cards were used to test both methods. One computer was used to re-create a voltage signal representing the sampled load current based on the statistical data collected from a bank of personal computers. The second computer was used to implement both the neural network and the adaptive filter in order to compare how the methods fair in canceling unwanted harmonics. Results show that the neural network approach gave excellent results in that it produced a very accurate recreation of the fundamental reference current, and therefore gave very good results for canceling unwanted harmonics. The adaptive filter approach has the advantage of requiring no prior training as it adapts itself in real-time to re-create the fundamental 60 Hz reference. The disadvantage is that there is a slight delay between the re-created reference, and the actual 60 Hz component of the load current. This is due to the delay associated with the d/a and a/d sampling involved. The neural network was trained to produce an output two time steps ahead (prediction), which effectively eliminated any delay in the re-created current reference as compared to the original fundamental load current. Conclusions are that while the adaptive filter is much easier to implement, the neural network gives no phase delay in the re created reference, and can compensate for harmonic changes, and changes in the fundamental amplitude in far less than one cycle time. The adaptive filter however, took as long as eight to ten cycle times to adapt to changes in the load current.