PRESENTED BY:
M.SADEES
[
attachment=13531]
TRANSIENT STABILITY ASSESSMENT
ANN Method
The training of ANN is executed using important selected features as inputs and critical fault clearing time (CCT) at pre-selected set of critical contingencies as desired target.
Multilayer feed forward neural network trained with back-propagation algorithm is used to provide the CCT.
Levenberg-Marquardt optimization based for
weight and bias updating algorithm is selected because it
provides a fast convergence and a better performance.
ANN training for TSA monitoring
Network Architecture
A single hidden layer feed-forward structure which is based on pattern matching technique with the back-propagation training network is implemented to relate the input/output pattern.
The number of input neurons depends on the total number of selected features.
The number of output neurons is equal to one for CCT estimation.
The number of hidden neurons depends on the complexity of the problem under implementation and the quality of available data.
Training and Performance Evaluation
The input/output pattern is merged, shuffled and normalized to avoid saturation and memorization problems during the training process.
The normalized data then divided into training, testing and validation subsets beside a set for testing as unforeseen data.
Among all back-propagation training algorithms, Levenberg-Marquardt optimization based for weight and bias updating algorithm is selected because it provides a fast convergence and a better performance.
SELECTED FEATURES IN THE MULTI-CONTINGENCY CASE
PERFORMANCE OF ANN IN MULTI-CONTINGENCY CASE
The estimated CCT and the target CCT during testing process
The plot regression between targets CCT and to estimated CCT
CONCLUSION
ANN is a very fast tool for TSA compared to traditional methods but should be trained carefully over a wide hyperspace in order to avoid over fitting.
The ANN based TSA can then be used to initiate the online remedial action.
The results show that the most sensitive parameters were found to be the post fault generators terminal voltage drop which represent the system state at post fault so that the experimental feature selection plays a crucial role in the ability of ANN to achieve a better performance in TSA.
Analysis of Large-Scale Power Systems
Analytical techniques alone do not provide all functionalities that control canter operators would like to have current operating point qualitative evaluation, stability margins, visualization of security regions, available transfer capability, preventive and/or corrective controls and "optimum” load shedding.
A possible solution to overcome this drawback is the application of the pattern recognition approach.
A major drawback for using MLPs as it requires extensive training process.
SUPPORT VECTOR MACHINES
SVMs are nonlinear models based on theoretical results from the Statistical Learning Theory.
SVMs minimizes the number of training errors.
They rely on Support Vectors (SVs) to identify the decision boundaries between different classes.
SVMs can map complex nonlinear input output relationships, and they are considered to be very well suited for TSA because the learning focus is on the security border.
Multilayer Perceptron Models
The MLPs have been trained by the Stuttgart Neural Networks Simulator [19], which is a free software developed in C.
The back-propagation training with adaptive learning rate, momentum constant and cross-validation has been used.
Support Vector Machine Models
The SVM classifier is based on training patterns, called support vectors, located at the separation region.
The SVsdefine the largest margin of separation between the two classes.
The optimization process emphasizes the minimization of the unstable patterns training errors.
MLP training and testing on TSA data set
TEST RESULTS OF MLPS TRAINED WITH 224 INPUT VARIABLES
TEST RESULTS OF SVMS TRAINED WITH 224 INPUT VARIABLES
TEST RESULTS OF MLPs TRAINED WITH REDUCED SETS OF INPUT VARIABLES
TEST RESULTS OF SVMs TRAINED WITH REDUCED SETS OF INPUT VARIABLES
Polynomial SVM ROC curves for different input sets
MLPS AND SVMS PERFORMANCE INCLUDING “HIGH RISK” CLASS
CONCLUSIONS
The SVM learning technique allows a deep understanding of its practical implications, which can be used to identify the important parameters for the classifier.
This work have successfully achieved reductions of approximately 50% on the number of input variables for MLPs.
It would be also possible to use larger training sets for SVMs to improve their performance, whereas the training time would not be considerably increased.