Deep Learning Based Harmonics and Interharmonics Pre-Detection Designed for Compensating Significantly Time-varying EAF Currents

In this research work, time-and frequency-domain Deep Learning (DL) based methods have been developed to predetect harmonic and interharmonic components of a current waveform of an Electric Arc Furnace (EAF) application. In the time-domain DL based approach, a DL-based algorithm predicts future samples of an EAF current waveform, which is then used in a multiple reference frame (MSRF) analysis together with exponential smoothing (ES) to detect harmonics and interharmonics. Comparably, in the frequency-domain DL based approach, sliding Discrete Fourier Transform (DFT) followed by DL is employed to predict the future samples of harmonics and interharmonics. Moreover, to obtain the most accurate and robust prediction system, grid search has been employed for parameter optimization of the DL structure. Due to the high computational complexity of the DL training phase, an NVIDIA TITAN XP Graphics Processing Unit (GPU) is employed which utilizes an efficient multi-core parallel processing infrastructure which was critical for making this work feasible. Testing on recorded field data resulted in outstanding prediction of all harmonics up-to 50th order and interharmonics with 5Hz resolution. In addition, the effectiveness of our proposed system for Active Power Filters (APFs) for harmonics and interharmonics has been evaluated in a simulation environment using field data and has shown to provide successful results. Since the proposed method due to its predictive nature can reduce the response and reaction time of APFs to zero while maintaining high compensation accuracy. The developed method can be considered to be a feasible candidate solution for generating reference signals to the controllers of a new generation of compensation devices which we refer to as predictive active power filters (pAPF).