Application of machine learning for real time security assessment and security control of power systems

Abstract

In this thesis, two specific areas of research work are proposed under power system security enhancement. These are the use of an inductive inference based decision tree classification technique for real time power system security assessment and security control. The global objective is to use the learning capabilities of decision trees to capture the complex power system security assessment and security control problem and the use of that knowledge to enhance the speed and efficiency of the process with minimal human intervention. The security assessment problem is approached as a classification problem in the thesis. It is found that the trainable classifier could be used to capture common underlying characteristics of a pre-contingent power system state to predict the post-contingent security state with reasonably •good accuracy. The performance of the proposed production rule classifier for security assessment is tested with two other established classification methods with significantly better classification results. The classification accuracy as high as 97-98 % is obtained from the proposed classification method, when the problem domain is partitioned into a number of sub-problems by topology, contingency and the type of violations wherever appropriate. The proposed production rule classification technique not only classifies a power system state as secure or insecure, but also provides a simple and very effective approach to power systems security control. Both real time steady state security control and transient stability preventive control schemes are successfully implemented with many useful features. The proposed security control scheme does not require any prior knowledge of control action in the event of a contingency, keeping the system operator intervention to a minimum. This is crucial in complex operating conditions where the operator requires considerable expertise in order to disclose a safe control strategy. In the proposed technique, suitable control actions are automatically proposed in an optimal way to be used in the real time environment. The operator's wisdom and the need for exhaustive operation planning stages in devising control strategies are no longer required. The proposed scheme can also be used in any contingency situation with any modelling sophistication to suggest preferred security control actions economically. The key concept of the proposed scheme is the identification of secure regions, which provide potentially new operating states that may be the objective of security control actions undertaken when the current state is found to be insecure. The thesis exploits this simplicity of identifying secure regions in two different perspectives for security control and accomplishes the task of moving the current insecure operating point into one of those identified secure regions in an optimal way. In real time security control, it is very important to have a feasible security control action for any insecure state. The proposed security control scheme can be implemented to realise a feasible solution to any insecure power system state with no false alarms or false dismissals in the new state. These concepts are validated in the thesis through simulations on standard test power systems.