Space-Vector Neural Networks: Efficient Learning from Three-Phase Electrical Waveforms

During the past decades, significant progress has been made in the field of artificial neural networks to process image, text or speech information including Recurrent Neural Networks, Convolutional Neural Networks or Transformers. However, the processing of electrical three-phase waveforms using these general network architectures ignore important signal characteristics and therefore lack of computational efficiency. This can lead to performance problems and can limit the application of neural networks for a fast and efficient analysis of electrical three-phase signals. To address this issue, this paper presents space-vector neural networks as new model architecture to process electrical current or voltage waveforms. A novel space-vector encoding and decoding layer is introduced to learn meaningful representations for different downstream tasks. The performance of space-vector neural networks is compared with other neural network and statistical methods for the reconstruction of unbalanced three-phase voltage signals and the detection of cyber-attacks in digital substations using a benchmark dataset. In both case studies, the results show a superior computational efficiency of space-vector neural networks in terms of the required number of parameters as well as the execution times with comparable reconstruction accuracies and detection capabilities. This demonstrates the high application potential of space-vector neural networks in the substation automation.