Date of Award
Doctor of Philosophy
Electrical and Computer Engineering
James D. Mccalley
Power System security assessment and the associated planning studies are becoming more and more complex with ever increasing uncertainties in all time horizons. An effective means of performing operational and investment planning studies of network limitations associated with static or dynamic post-disturbance performance problems has been to take a Monte Carlo simulation based approach. The approach harnesses computing power to develop a database of post-contingency response over a wide range of different operating conditions, and then apply statistical or machine learning methods to extract useful planning and operational information from the database.
Key to the machine learning based planning approach is the manner in which the different operating conditions are sampled to generate a training database. This work develops an efficient sampling procedure that maximizes information content in the training database while minimizing computing requirements to generate it, by finding the most influential region in the sampling state space and sampling operating conditions from it according to their relative likelihood. The Monte-Carlo variance-reduction methods are used to construct the proposed sampling approach, which is envisioned to allow market-oriented industries to operate the system according to economic rule.
The dissertation also develops a comprehensive methodology to perform decision tree based security assessment for multiple contingencies. The system security limits and associated operating rules depend on the set of contingencies considered for planning. Considering the probabilistic nature of the power system, this work develops a risk based contingency ranking method that helps in screening the most critical contingencies from a contingency list. The developed contingency risk estimation method gives realistic risk indices since it takes into account the non-parametric nature of operating condition distribution, and it also saves tremendous computational cost since it uses linear sensitivities to estimate the risk. Finally, a contingency grouping method is proposed that guides in generating common operating rules for every group that performs well for all the contingencies in that respective group, thereby providing system operators the benefit of dealing with lesser number of rules. The contingency grouping is based on newly devised metric called progressive entropy that helps in finding similarities among contingencies based on their consequences on the operating conditions along all the load ranges, and not just their proximity in the grid.
The proposed methods are implemented in the west France, Brittany region of RTE-France's test system to derive decision rules for multiple contingencies against voltage stability problems.
Venkat Kumar Krishnan
Krishnan, Venkat Kumar, "Efficient processing of system scenarios in statistical and machine learning studies for power system operational and investment planning" (2010). Graduate Theses and Dissertations. 11315.