Power Quality State Estimation on base of Neural Networks and Fast Fourier Transform

Abstract:
The high penetration of converter-based electricity generation and consumption increases the need for monitoring tasks within highly complex distribution grids. For example, photovoltaic power plants and electric vehicle charging stations typically operate at high rated power and may cause harmonics and transient events. Costly power quality meters may not cover sufficient nodes to employ conventional state estimation algorithms, which are typically only applicable to simpler network topologies such as transmission grids. However, power quality state estimation based on neural networks offers measurement augmentation to address these challenges in close-to-real-time. The proposed model estimates fundamental frequency components, harmonics, and tracks transient events. A case study demonstrates the model’s capabilities in a heterogeneous distribution network. The training data is generated through simulations of harmonic flows and transient faults in the frequency-domain. A dense neural network is trained on mixed simulation data with spectra, which can contain frequencies from DC up to 20 kHz. This is a typical bandwidth for power quality analyzers. The trained model is tested on previously unseen spectra and system states. Accurate estimation results are achieved for the 44-busbar CIGRE LV reference grid with fictitious power quality meters placed at three nodes only.