Integrating Machine Learning into supervisory control of power systems for minimizing line losses after abnormal state detection

A novel framework for integrating Machine Learning (ML) algorithms into the supervisory control of multi-machine power systems is proposed. In particular, ML is used to identify the operating state of the power system and minimize the transmission line losses based on bus angle data collected from Phasor Measurement Units (PMUs). The ML model is initially trained using power flow analysis to find the optimal power allocation among the Generation Units (GUs) for specific abnormal operation states of the power network. The power allocation for each investigated state defines the optimal power sharing ratio among the GUs, which is applied into a distributed secondary control scheme. Using Principal Component Analysis (PCA), the vector that contains all the bus angles is projected to the 3-dimensional space, reducing the size of the training data. Then, using the k-Nearest Neighbor algorithm (k-NN), the measured data are assigned to one of the classes that were defined in the training phase. The above ML-based method is incorporated into the complete hierarchical control structure of the system, thus establishing for the first time a detailed dynamic cyber-physical multi-machine power system. The proposed framework is validated on a 6-bus 3-machine power system resulting in reduced power losses compared to conventional techniques under several abnormal scenarios (line loss, load disconnection, high load variation).