Enhancing the response of thyristor‐controlled reactor using neural network

Abstract

Summary In this paper, a neural network controller is proposed to retrieve the voltage balancing conditions in three‐phase power systems. The neural network is suggested to calculate the required set of firing angles for the thyristor‐controlled reactor accurately to balance the three‐load voltages quickly. The proposed controller is fed by different parameters within different feeding techniques, namely, root mean square (RMS) values of the three load voltages, RMS value of the space vector signal calculated from the three load voltages, and the RMS values of both the three load voltages and their associated space vector. The intentions of the proposed techniques are to combine between reducing the number of measured parameters and providing the controller with qualitative data about system status. The influence of the measured parameters on the neural network performance is examined by calculating the regression coefficients through several test cases. Accordingly, only the effective parameters are utilized to reduce the neural network complexity and to enhance the controller response time. Additionally, the calculations of the controller input parameters are made in terms of space vector cycle, which is half of system sinusoidal cycle. Consequently, the calculation time is reduced significantly. The Aqaba‐Qatrana‐South Amman power system is considered and modeled as a real case study. In addition, several test cases have been conducted to test and validate the ability of the proposed neural network controller in retrieving the voltage balance conditions precisely and quickly. The results have revealed the ability of the proposed neural network controller to calculate the firing angles quickly within 10 milliseconds and achieve very low voltage unbalance factor.

Hybrid PSO–ANN algorithm to control TCR for voltage balancing

“…The number of samples used to train the proposed ANN is moderate in comparison with the other techniques. Moreover, the number of samples used in this work is higher than the ones in [22]. It is important to emphasise that the range of VUF included by the proposed hybrid PSO–ANN algorithm is large compared with the other techniques.…”

“…In [21], Kulkarni and Udupi proposed a hybrid GA‐ANN algorithm to control SVC to mitigate the voltage unbalance problem. In [22], Ragab et al proposed a trial‐and‐error‐based algorithm to train ANN that controls TCR for voltage balancing of long line EPS.…”

“…On the basis of literature review, the following points address the main drawbacks in some algorithms used for controlling SVCs for voltage balancing: The number of ANNs used in [19, 21] is large (three ANNs), which leads to a complex controller structure. In [19, 21], each ANN has a large number of hidden layers (three hidden layers), which increases the computational time. A large number of ANN inputs (more than three) makes the computational process more complex [19, 21, 22]. The contribution of this work for solving the voltage unbalance problem could be summarised by the following points: Applying PSO algorithm with the derived general mathematical model of the long transmission line to determine the optimal set of TCR firing angles for voltage balancing accurately and easily. Proposing a simple structure feed‐forward ANN with a single hidden layer, which requires only three inputs to determine the set of TCR firing angles, in online mode. Training the proposed simple ANN with a large dataset obtained using PSO, which makes the ANN able to mitigate the voltage unbalance problem for a large range of VUF at low average response time. Building a laboratory prototype for AQSA EPS including its long transmission line to validate the ability of the proposed hybrid PSO–ANN technique in mitigating the voltage unbalance problem at the load side. The flowchart in Fig.…”

“…A large number of ANN inputs (more than three) makes the computational process more complex [19, 21, 22].…”

“…It can be seen from Table 5 that the proposed ANN has the simplest structure because it consists of a single hidden layer and it requires only three inputs, whereas the other ANNs have three hidden layers as in [19, 21] or their inputs are more than three as in [22]. Moreover, the proposed ANN is trained using an optimal dataset obtained using PSO while the ANN in [22] is trained using a trial‐and‐error technique which is a time‐consuming method and does not guarantee the convergence to the optimal values of the firing angles. The number of samples used to train the proposed ANN is moderate in comparison with the other techniques.…”