Application of Artificial Neural Networks to Power Systems

Power System Analysis

The authors [PS55] propose an ANN application to state estimation computation. In power systems state estimation computation takes an important role in security controls, and the weighted least squares method and the fast decoupled method are widely-used at present. State estimation computation using the existing Von-Neumann type computer is reaching a limit as far as the solution techniques are concerned, and it is very difficult to expect much faster methods. In order to solve the problem, the authors employ an ANN theory, the Hopfield network theory, which has an ultra parallel algorithm and is different from the existing calculating algorithms, for state estimation computation. A feasibility study using a 6 bus system is shown.

The ANN system employed in the study [PS56] has 3 layers structure which is composed of an input, hidden, and an output layers. As a simulation example the calculation of electric power flow was considered. The results in this study show an idea how to construct the ANN system for fast calculation or simulation of the electric power equations.

The authors [PS57] report the development of a fast load flow method based on a ANN which can be used for such real time applications. A feedforward model of the ANN based on BP algorithm has been used for the load flow problem and the method has been tested for two sample AC systems and one multiterminal AC/DC network. The results reveal that the proposed model also provides sufficiently accurate results.

The paper [PS58] proposes a method for nodal voltage diagnosis in power systems using a self-organization ANN. ART2 is utilized to classify power system conditions. A probability voltage security index is evaluated by the resulting classification. The proposed method is used for tracking the voltage profile continuously.

At the Department of Power Engineering FEI TU in Kosice has been proposed an ANN-based load flow method. Power system model used to test the ANN approach to power system analysis is shown on Fig.1.

Fig.1 Model Power System

Fig.2. Neural Network and Scheme of the ANN Approach

Reffering to Fig.2. the design of the ANN approach to PS analysis starts with the creation of training samples in Data Preparation Unit. After creation, the samples must be normalized in Normalization Unit. The learning algorithm was used to train the ANN and the weights and biases are obtained. The weights and biases are used to test the accuracy and speed of the proposed approach. In order to test behavior of ANN, output values of the trained network must be denormalized in Denormalization Unit. The Data Preparation, Normalization, and Denormalization units are writen in Borland Pascal language.

The creation of training samples (input-target pairs) data is accomplished through the following steps:

1. Input the Model Power System (Y matrix);
2. Create sufficient amount of sample input vectors (P, Q);
3. Calculate the load flow (Newton-Raphson method) for each sample vector. Complex bus voltages are the target vectors;
4. Output the input-target pairs to be input to the ANN.

The 5-bus test power system has been used to test the effectiveness of proposed method. The ANN has 8 inputs and 8 output neurons. During tests the number of hidden neurons has been varied from 5 to 20. The ANN's sum squared error goal has been set to 0.001 and the maximum number of learning epochs has been set to 10,000. Logistic-sigmoid, tangens-sigmoid and linear transfer functions have been tested.

The test results has shown that the most efficient is the linear transfer fuction in the hidden layer. The sufficient number of hidden neurons is from 10 to 15. The mean absolute error of ANN is less than 0.05 per cent.

The trained network operated adequately. The simulation results shows that proposed ANN approach can be used to make on-line load flow calculations more faster than traditional approaches by a factor of at least ten on small distribution systems and perhaps more than that on larger systems.

The learning parameters and test results are given in Tab.1.

Table 1. Learning and test results.

Previous | Home | Next | Table of Contents | Glossary | Reference

Please send Questions, Comments, Suggestions to bencr@hotmail.com

Copyright © 1997, 98 Rado Benc