This work studies a novel artificial intelligence (AI) method for the assessment of power system security and the initiation of preemptive protection procedures to ensure transient stability. Power system security is one of the fundamental concerns of utility operators as it is directly related to public consumers and industries. Therefore, the goals of this research include developing an AI based transient stability simulator that may provide faster solutions to the power system settings that would overcome transient faults. Such a simulator could be paired with algorithms to operate power systems real-time, in an optimized but best performing mode in terms of economy and security.

For research purposes, a small power system containing two generators was simulated. The output from this simulation was used to train a feedforward neural network using the Levenberg-Marquardt algorithm (LMA), with the intent that the trained neural network could then determine the stability of the power system model much faster. Training data were tabulated by simulating the power system under a wide range of conditions, and then analyzing the resulting load flow and rotor angle stability. For example, artificial faults were created, the load was varied in size and type, and controller parameters were changed. The controllers that are most commonly used in utility-scale power generation and transmission systems were studied, including the Static VAR Compensator (SVC) and the Power System Stabilizer (PSS).

After training, the neural network was tested to determine the accuracy of its stability predictions for a set of different input conditions. The network performed well, showing promise for the use of this approach in power system analysis. Recommendations were made to further improve and evaluate the approach, including the use of additional training data and more sophisticated neural networks.

A simple post-processing adaptive method that would use the neural network to determine a set of stable and desirable power system parameters was then successfully created and tested. A more sophisticated method could be created by subsequent researchers to, for example, automatically reconfigure a power system to the most cost-effective stable condition. Future research could also investigate the use of neural networks in fault diagnosis, security assessment, load forecasting, economic dispatch, and harmonic analysis.