LSTM-Based Estimation of Harmonic Source Equivalent Parameters in Power Systems

This paper proposes a deep learning-based method to estimate equivalent parameters of harmonic sources, thus improving the reliability of harmonic contribution assessment. Harmonic contribution assessment is a technique for quantitatively estimating how much individual harmonic sources contribute to harmonic distortion in power systems. In general, harmonic contribution assessment requires the equivalent parameters of individual harmonic sources for each harmonic order, which are estimated using measurement data in power system and numerical analysis methods. However, numerical analysis-based estimation methods deteriorate when measurement data contain some noise or when equivalent parameter conditions change, thereby degrading the accuracy of harmonic contribution assessment. Therefore, these methods require appropriate setting of algorithm factors for maintaining estimation performance, but such setting is very difficult. To overcome this issue, this paper proposes a long short-term memory (LSTM)-based method that estimates the equivalent parameters of harmonic sources using time-series voltage and current data measured in power systems. Because the proposed method is based on a pre-trained LSTM model, algorithm parameter settings are not required even under changes in equivalent parameter conditions. In addition, the method does not require algorithm initialization in changing the conditions of harmonic sources, as well as maintaining stable estimation performance even in the presence of some noise. In the case study, an LSTM-based learning model was developed for the modified IEEE 15-bus and IEEE 33-bus test systems using PSCAD/EMTDC, and its superior performance was verified by comparison with the previous methods under various operating conditions.