Congestion Management Using an Optimized Deep Convolution Neural Network in Deregulated Environment

. The technical issue of congestion, which is predominantly found in deregulated power systems, is caused by the failure of transmission networks to satisfy load power demands. This failure is primarily caused due to an increase in loads or loss of transmission lines or generators in modern restructured power networks. This work introduces a CM approach using Deep Convolution Neural Network (DCNN) for minimizing congestion and supporting Independent System Operators (ISOs). The purpose of the work is to generate enhanced prediction outputs for congestion management with reduced error values. These objectives were achieved through the actual power rescheduling of generators. The proposed work adopts DCNN which is optimized using an Improved Lion Algorithm (LA) and aids in providing significant outcomes for congestion management with reduced error. By implementing customized IEEE 57-bus, IEEE 30-bus, and IEEE 118-bus test systems, the suggested approach has been successfully verified for its performance on test systems of varied sizes. This analysis incorporates restrictions such as line loads, bus voltage influence, generator, line limits, etc. The most important results for the test system indicating convergence profile, congestion cost, and change in real-power and voltage magnitude are obtained by the simulation in MATLAB, and on the basis of the obtained simulation outcomes, it is evident that the proposed Improved Lion Algorithm optimized Deep Convolution Neural Network displays phenomenal computation performance in minimizing congestion losses at minimum congestion costs. When compared to several contemporary optimization techniques, the suggested technique performs better in terms of congestion cost and losses by generating improved prediction outputs with reduced errors.


INTRODUCTION
Electric energy is the driving force behind the functioning of the modern world and its rise in prominence is mainly because of industrialization, urbanization and enhanced life style. Consequently, the overdependence and ever-increasing demand for electric energy has led to several rapid advancements in the power sector. Previously, vertically integrated utilities were used to operate the power grids, and the government mostly controls this regulated power system. Thus, both incurred expenditure and the resultant revenue of the power system are both handled by the government [1][2][3]. However, the excessive demand for power in recent times effectuated the deregulation and privatisation of electric power system. This in turn has contributed to the restructuring of the power system with the inclusion of numerous smaller generation plants, comprising of sustainable power sources to meet the booming number of loads [4]. As a result of excessive power requirements, transmission systems are operating beyond their thermal and stability limits, placing strain on the current power system architecture. Moreover, in a deregulated environment, the DISCOs, GENCOs and TRANCOs are not controlled by a common institution, instead different organizations manage these companies and the establishment of coordination between these companies is left to an ISO. The transactions made by the DISCOs and GENCOs are unpredictable, abrupt and ahead of time, resulting in transmission line congestion [5,6]. The issue of transmission line congestion mainly occurs due to rise in load demand, generation outages and equipment failure. The vital task of relieving this congestion and ensuring a safe and secured working of power system, is entrusted to ISO. The major techniques followed by ISO to relieve congestion are cost free and not cost-free methods [7]. The former involves Flexible AC Transmission (FACTS) devices, transformer taps, network reconfiguration, phase shifters or congestion lines out-ageing. The latter entails approach like curtailment of loads, generation prioritization and generation rescheduling. In certain situations, ISO informs the consumers about the specific line congestion and facilitate load adjustment inside the limits of system constraints. In severe cases, the CM is carried out by physically restricting the transaction, irrespective of the inconvenience to consumers [8,9]. FACTS devices are regarded as technology that, lowers transmission congestion and improves grid infrastructure use. The usage of FACTS controllers has some drawbacks, including challenges with placement, size, cost, and modelling that are ideal. In order to manage congestion in reorganized electricity markets, this article discusses the application and ideal position of the FACTS device series [10,11]. Through the creation of an algorithm to improve working measure of contingency analysis as well as positioning and control of Thyristor-Controlled Series Compensator (TCSC) [12], and operation of TCSC for transmission line optimization and congestion is explored. The best location for TCSC [13] in terms of increasing power transmission efficiency, limiting steady-state instability, and preserving power system voltage stability. TCSC is used in power systems to enhance transient response and congestion control. The explanation of the objective models for minimising expense and load shedding involved optimising welfare of society, limiting load shedding, as well as increasing load served. Two generators and bus sensitivity factors were presented along with Particle Swarm Optimization (PSO) technique. However, PSO exhibits demerits including sensitivity to parameters, lack of diversity and premature convergence leading to inaccurate outputs [14]. In [15], Genetic algorithm is engaged for finding best generation schedule for CM in an unregulated power system but shows challenges in the improvement of congestion management performance. The Grey Wolf Algorithm (GWO) is employed for congestion management due to its ability of enhanced convergence speed yet gets trapped on local optimum value [16]. Firefly Algorithm (FA) is another metaheuristic algorithm employed for handling congestion management but exhibits inability in handling optimization problems with constraints [17]. The line overload problem during congestion management is eliminated in power system by grasshopper algorithm (GA), however, the inappropriate selection of parameters may lead to premature convergence of the algorithm [18]. The differential algorithm is adopted the hourly congestion management but demands increased consumption of resources leading to resource shortage [19]. Bat algorithm is also deployed for the congestion management in power systems but faces issues related to computational complexity [20]. Several studies introduce deep neural networks together with metaheuristic algorithms for congestion control in response to these problems. In [21], glow worm swarm optimization is adopted for the optimization of DCNN which in turn adjusts the weight initialization. Anyway, with the increase in data size, slight fluctuations occur in memory usage of the algorithm. In [22], atrous convolution algorithm is used for the optimizing of DCNN but the accuracy results attained are not high. In [23] swarm intelligent based algorithms are adopted for the optimizing of DCNN. However, these algorithms face issues related to convergence and accuracy. Considering these shortcomings, the novelty of the work engages a DCNN network with Improved Lion Optimization, which is a recent optimization strategy showing remarkable performance towards congestion management. Contributions of the study are,  An Improved LA optimized DCNN is proposed for relieving congestion in a deregulated environment.  The presented CM approach is tested for its effectiveness in IEEE 118-bus, 57-bus, and 30-bus systems.  The proposed methodology is effective in minimizing congestion cost and losses.

A. Problem Formulation
The primary goal is to lower the systems z cost, which is taken into account.

B. Bus Sensitivity Factor (BSF)
BSF is defined as ratio of incremental changes occurring in th m power of the bus to an incremental change in real power flowing through bus " " which is linked to buses " " and " , " as shown below. On the basis of greatest negative sensitive indexes, BSF offers the best location for pumped hydro storage unit deployment.
From the expression above, the degree to which the amount of real power changes in accordance with amount of real power injected at bus m in a transmission line is represented by indicates the incremental changes in real power that flows in bus which is connected between and buses, pm  represents an incremental change in th m power of the bus.
Equation (12) is used to derive BSF, as shown below Here, The Improved LA-optimized DCNN is employed for congestion management in this work and the presented approach is shown in Figure 1.

C. Optimized DCNN with Improved Lion Algorithm (LA) for classification
In this work, DCNN is adopted in which the automatic optimization of hyperparameters is carried out by improved LA. In Figure 2, the general flow diagram of DCNN with optimization is indicated. Here, back propagation is used for the learning process. The obtained prediction output from the fully connected layer is compared with an actual value and subsequently, the loss function calculates the error value. The Stochastic Gradient Update (SGD) function is used in the training procedure of DCNN. Consider, the n samples of the training dataset and assume ( ) as the loss function in which denotes the index and denotes the parameter vector. The objective function is given by,

Fig.1. Congestion management using Improved LA-optimized DCNN
The following expression denotes the objective function gradient at .
The computational cost for each independent variable iteration, if gradient descent is used, is given by () is uniformly sampled at each iteration of SGD for updating by computing () Here, indicates the learning rate. The DCNN structure used in the proposed work is AlexNet which is an updated architecture generating improved accuracy with less computational time. Table 1 represents the Alexnet DCNN layer architecture used in the proposed work.
The convolutional layer extracts the features from the data and is normalized by ReLU. Subsequently, the pooling layer of size 3×3 reduces the number of sizes thereby minimizing the complexity. In this DCNN, categorical crossentropy is adopted as the loss function DCNNs are trainable architectures with biological inspiration that acquire on invariant aspects.
A commonly used smoothing filter is the discrete Gaussian filter The output feature map retains a dimension by utilising discrete convolution at specific locations on input feature maps and is expressed as The convolutional layer with its membership function including multilayer perceptron is expressed as s  feature map and ℎ unit in . The weights are updated with the backpropagate of error in the network after the calculation of network error. The optimization algorithm updates the weights till the minimized value of error is obtained and the error does not get reduced further. For better prediction outcomes, it is preferable to make the values ideal rather than generating some random evaluation values. However, the automatic finding of hyperparameters of DCNN is crucial and requires the involvement of metaheuristic algorithms. The tuning in this work is done using the optimization idea, specifically, a novel tuning approach is presented.  The Improved LA model is used in the work is presented to optimise a weight of DCNN. Here, the current LA method is enhanced so that it is capable of handling the difficult optimization problems. Self-improvement has generally been shown to be promising in conventional optimization techniques. The live nature of lion species served as the basis for LA model. It consists of four stages, including "mating, pridegenerating, improved territorial takeover and territorial defence". The proposed Improved LA adopts the improved territorial takeover phase in which the lions are updated based on the maximum age of cubs. In contrast, conventional LA do not have specific updating process. The solution vector of Improved LA is referred as

Solution Encoding and Objective Function
The random constraints are specified by variables 12 , rrand , which are produced and lie between [ Len , respectively. Also, female update process is indicated as  .

Matching
Gender-based clustering occurs as a result of the crossover and mutation processes that occur during mating. Cubs are generated by mutation and crossover process and are referred as, cubs s which are produced by cross over process and new s by mutation process. Thus, a lioness gives birth to four cubs when it is pregnant, and another four cubs are created through the crossover process. These four cubs are used to carry out mutation procedure in order to create four further cubs. (50) Figure 4 shows the flow chart for the suggested Improved LA model. From the expression above count of generation is indicated by , which is set to zero at initial and further increased to 1, during the territorial takeover. max it and th er stands for maximum generation and error threshold, respectively. The list of hyperparameters for the evaluation selected with the help of improved LA is mentioned in Table 2.

RESULTS AND DISCUSSION
In this study, an Improved LA optimized DCNN is used for resolving congestion issue in unregulated environment. The optimized Deep CNN facilitates the active power rescheduling of generators with reduced congestion cost. Around 500 loading scenarios are being generated among which 78% of patters are adopted for training and 22% of patterns are adopted for testing. Out of the 390 loading scenarios of training set, the number of congested scenarios identified is 378 while the non-congested cases is 12. Among 110 loading scenarios of testing set, the number of congested loading scenarios is identified as 100 whereas the number of non-congested loading scenarios is identified as 10. An apparent power load, active power load and reactive power load are applied as inputs to DCNN in which the dimension of the input layer is given by 3×21×1.
In order to apply power load, active power load and reactive power load as inputs, data requires pre-processing. Initially, the data has to be collected at regular intervals and further preprocessed which involves removal of outliers and conversion of data into time-series format. The data could then be formatted into a tensor or array, where each row represents a time step and each column represents a feature, such as active power load. Finally, the formatted data can be fed as input to the DCNN.The proposed work is verified by implementing in MATLAB and is tested under variety of networks including IEEE 30-bus, IEEE 57-bus and IEEE 118-bus. An upper voltage of the load bus is 1.1 p.u, while the lower voltage of the load bus is 0.9 p.u. Table 3 lists the test systems considered for evaluating the performance of Improved LA optimized DCNN for CM, while the congestion line details are presented in Table 4. Table 3 Test System details

IEEE 30-Bus Test System
For comprehending the potential of proposed DCNN based CM approach, a revised version of IEEE 30bus system that comprises of 24 load buses, 6 generator buses and 41 transmission lines is considered. The two different cases considered here are: Case 1Apower outage causes congestion between lines 1-7 and 7-8; Case 1Bload rises to 50% at every bus and the lines 1-2, 2-8 and 2-9 are congested. Table 3 gives the details about the obtained results from which it is noted that the proposed work generates improved outputs of 18.707 for case1A and 161.14 for case 1B. Table 5 Test system results TRRG-Total Real power Rescheduling Generator, TC-Total The simulation results for case 1A are provided in Figure 5. On analyzing the figure, it is detected that a congestion cost is minimum for the proposed CM approach using Improved LA optimized DCNN. [24] 719. voltage magnitude is also maintained within a reasonable range (0.9 to 1.1) after CM

Fig. 6. Case 1B simulation outcomes (a) Convergence profile (b) Congestion cost (c) change in realpower and (d) Voltage magnitude.
From Figure 6, which gives the simulation outcomes for case-1B, deduces that the congestion cost is comparatively lower for the proposed DCNN based CM approach. In this case, increase in load along with the outage of line between 1 and 7 results in overloading Moreover, the system losses are also reduced to 14.59 MW from 37.24 MW after CM using Improved LA optimized DCNN.

IEEE 57-Bus Test System
Next, a revised topology of IEEE 57-bus test system considered for CM is made up of80 transmission lines, 50 load buses and7 generator buses. Its reactive and real power values, 336 MVAR and 1250.8 MW respectively. Moreover, the details and results of the two cases coming under this test system is provided in Table 6.  With the occurrence of congestion, there is an overloading between lines 6-12 and 5-6. After CM using the proposed methodology in case 2A, the system loss is significantly reduced to 24.558 MW from 69.64 MW. whereas simulation outcomes for case 2B are illustrated in Figure 8. In case 2B, line overloading is created by reducing line limit to 20 MW from 85 MW between lines 2-3. From analyzing Table 4, it is observed that the proposed approach delivers comparatively better performance in Case 2A also. In this case the system losses are greatly reduced to 28.22 MW from a primary value of 78.23 before CM. On the whole, the violation of overloading lines is alleviated by the optimized real-power rescheduling.

IEEE 118-Bus Test System
The proposed DCNN based CM is also evaluated for its effectiveness in a larger test system by deploying it in a revised topology of 118-bus test system, made up of 54 generator buses, 64 load buses and one 186transmission lines. In this case, the lines between 5 and 8 are disconnected, while the loads between lines 20 and 11 are increased 1.57 times. Figure 8 gives the simulation resultsforCase3 In this case, the total system loss becomes 230.505 MW after CM using DCNN. The value of system loss before CM is 277.301 MW. Thus, it is significantly apparent that the proposed DCNN methodology is effective at minimizing congestion in any test system, regardless of its size. Figure 10 represents the comparison of convergence in terms of cost and iteration. From the curve it is clear that the Improved LA exhibits rapid convergence rate when compared to conventional LA.      Table 6 represents the comparison of the proposed ILA-DCNN for generator rescheduling with existing works. The listed values indicate that the proposed work outperforms other ones with enhanced prediction outputs indicating reduced error percentage.

CONCLUSION.
This study suggests a novel robust methodology for CM in an unregulated open access electricity environment. In order to satisfy several electrical constraints, problem was developed as multipleobjective function, with losses and congestion costs as vital factors. Conventionally, FACTS devices or nature-inspired algorisms were prominently employed for CM in many works. Meanwhile, in this work, DCNN is chosen for congestion minimization in an unregulated environment for solving the tasks of issues in congestion management due to uncertainties in load. The working of the DCNN is enhanced further by using Improved LA optimization. The proposed DCNN-based generator rescheduling approach is put to test for its performance in three different test systems of varied sizes. Moreover, its performance is evaluated by analogizing with other existing methodologies in these test systems. The proposed work is simulated in MATLAB and on the basis of the obtained simulation outcomes, it is evident that the proposed Improved LA optimized DCNN displays phenomenal performance in minimizing congestion losses at minimum congestion costs. Moreover, it also outperforms other techniques in terms of its superior performance in managing congestion. The future extension of this work can include the adoption of hybrid optimization algorithms for the enhancement of neural network parameters. Moreover, the effects of optimization over multiobjective functions have to be analyzed in a detailed manner. Conflict of interest. The authors declare that they have no conflicts of interest.