Figure 4.8: Contingency ranking of IEEE-39 Bus System for P IV Q
Figure 4.9: Bus voltages of IEEE-39 bus system after outage of line between buses 15-16
voltages at buses 19, 22, 25, 26 and 36 are found more than their respective max- imum permissible limit. This fact advocates line 15-16 is critical as far as voltage point of view.
4.3 Proposed Methodology for Contingency Anal-
(ii) Feature selection for selecting important attributes from the data.
(iii) Design of different Neural networks employing different architectures.
4.3.1 Data Generation
The data set should ideally represent all possible operating conditions and should be reliable. Reliable data generation is the most important step for the success of a neural network. Therefore, data is generated using the methodology proposed in the flowchart shown in Figure 4.10. The steps followed for data generation for security assessment and contingency analysis are:
Step 1 A large number of load patterns have been generated by randomly distribut- ing the real and reactive loads on all the load buses. This exercise ensures that the data set is a representative of all possible operating conditions.
Step 2 During simulation, the system load has been changed from 1.0 (base case) per unit to 1.0±0.05 per unit of base case.
Step 3 A contingency set consisting of all credible contingencies is considered. For each operating condition, a contingency is simulated. Single line outage is the most common event in a power system and hence, only one line outage at a time is considered here forming a set of NL−1 contingency, where NL is the number of lines.
Step 4 Single line outage corresponding to each load pattern are simulated by AC load flow (Newton Raphson Power Flow) and the violation of operating limits of various components are checked. Keeping the load level constant, each contingency is simulated several times to obtain a wide range of operating scenarios. For this study each contingency is simulated 15 times for a selected load level.
Step 5 The post-contingency state of the system is stored for each contingency to calculate the performance indices, P IV Q and P IM V A. The obtained values are normalized between 0 and 1 for each contingent case.
Yes
Yes
Yes No
No
No
Figure 4.10: Flowchart of data generation for static security assessment and contingency analysis
Step 6 The system state, contingency type, and the corresponding security are noted for every operating point and for all the contingencies of a credible contin- gency set.
Step 7 The whole data set is suitably divided into training set and test set for performance evaluation purpose.
4.3.2 Data Normalization
The input/output training and testing set data are scaled in the range of 0 to 1 for each load pattern. In this thesis, each input or output parameter ofxis normalized asxn before being applied to the neural network as
xn = 0.8×(x−xmin) xmax−xmin
+ 0.1 (4.7)
Where xmax and xmin are the maximum and minimum values of data parameter x.
4.3.3 Feature Selection Using IGWO
Feature selection is a binary optimization problem. For the Intelligent Grey Wolf Optimizer (IGWO) based feature selection method, a binary version should be de- veloped. In this work, a solution is represented as a one dimensional vector, where the length of the vector is based on the number of attributes of the original dataset.
Each value in the vector (cell) is represented by ‘1’ or ‘0’. Value ‘1’ shows that the corresponding attribute is selected; otherwise the value is set to ‘0’.
Feature selection can be considered as a multi objective optimization problem where two contradictory objectives are to be achieved with minimal number of se- lected features and higher classification accuracy. The smaller is the number of features in the solution and the higher the classification accuracy. Each solution is evaluated according to the fitness function [247], which depends on the k-Nearest Neighbors (KNN) classifier [248] to get the classification accuracy of the solution and on the number of selected features in the solution. In order to balance between
the number of selected features in each solution (minimum) and the classification ac- curacy (maximum), the fitness function in Equation 4.8 is used in IGWO algorithm to evaluate search agents.
F itness=λ1γR(D) +λ2|R|
|N| (4.8)
Where γR(D) represents the classification error rate of a given classier (the K- nearest neighbor (KNN) classifier is used here). Furthermore, |R| is the cardinality of the selected subset and |N| is the total number of features in the dataset, λ1 and λ2 are two parameters corresponding to the importance of classification quality and subset length, λ1 ∈ [0,1] and λ2 = (1–λ1) adopted from [229]. Generally, the monetary gain is the supreme motivation behind any investment along with the improvement of system performance.
4.3.4 Training and Testing Pattern
Off-line power flow calculation results corresponding to each contingent case are used to construct the training patterns. The load patterns were generated by randomly changing the load at each bus. Single line outage contingencies are considered in this work for online ranking, as these occur more frequently. Input features for ANN consists of pre-contingent variable and the PI values will be used as targets to perform contingency analysis. For this study, data consists of a large number of patterns that is normalized, shuffled and divided in two groups; one for training and the other for testing.
4.3.5 Selection of Neural Network Architecture
The neural network architecture selected for online security evaluation are Feed For- ward Neural Network (FFNN) and Radial Basis Function Neural Network (RBFNN).
The applicability of these two ANN-based methods has been investigated under different operating and contingency condition with IGWO based feature selection method. Initially these neural networks have been trained for estimation of PI value.
Table 4.3: PI based contingency classification for screening and ranking
Class/Rank PI Range Static Security Status
I: Most Critical >0.8 Insecure (0)
II: Critical 0.4-0.8 Insecure (0)
III: Less Critical 0.2-0.4 Insecure (0)
IV: Non Critical <0.2 Secure (1)
The neural network which gives the best results will further employed for contin- gency classification.
4.3.6 Contingency Classification States
First, the values of PI are determined, after this classification of the contingency on the basis of their PI values. Neural Network classification task is completed by classifying the contingencies either as secure or insecure in accordance with Table 4.3, that finally gives the ranking for all the contingencies on the basis of their PI values.
Class-I contingencies indicate that they are never safe under any operating condi- tion and require immediate attention. Presence of these contingency causes major violations grenerator’s active and reactive power limits, voltage limits of buses and the thermal limits of transmission lines.
Class-II contingencies indicate that they are not safe since there is major violation of all or some operating constraints respectively depending upon the operating con- dition and these contingencies require proper preventive control measures such as generator rescheduling or load shedding.
Class-III contingencies indicate that they are less critical but there are minor vi- olations of some system constraints depending upon the operating conditions and these contingencies become safe with proper control measures.
Class-IV contingencies indicate non-critical contingencies that never drives the power system into insecure state. Here, all the critical contingencies (Class I to III) are indicated by ‘0’ and Non critical contingencies (Class-IV) are indicated by ‘1’.