tvnovellas.info History NEURAL NETWORK DESIGN ELECTRICAL ENGINEERING PDF

NEURAL NETWORK DESIGN ELECTRICAL ENGINEERING PDF

Friday, July 19, 2019 admin Comments(0)

Ebook Neural Network Design Electrical Engineering currently available at tvnovellas.info for review only, if you need complete ebook Neural Network Design. Neural network design electrical engineering martin t hagan howard b demuth mark beale on sitecom free shipping on qualifying offers the well known. Request PDF on ResearchGate | Neural Networks and Fuzzy Logic in Electrical These exercises include modelling, analysis and design of a control system. Article in International Journal of Electrical Engineering Education 40(1).


Author:TOBIE SCHNIBBE
Language:English, Spanish, Portuguese
Country:Kyrgyzstan
Genre:Health & Fitness
Pages:544
Published (Last):21.06.2015
ISBN:884-5-36426-332-4
ePub File Size:24.36 MB
PDF File Size:8.72 MB
Distribution:Free* [*Sign up for free]
Downloads:41492
Uploaded by: TARI

Neural Network Design Electrical Engineering Preparing the books to read repair manual Free nbsp;neural network design has 29 to. Neural. Network. Design. 2nd Edition. Hagan. Demuth. Beale. De Jesús .. masters (in Microsoft. Powerpoint format or PDF) for each chapter are available on the web at Electrical and Electronic Engineering Department at the University of. International Journal of Electrical Engineering Education 40/1. Neural networks These exercises include modelling, analysis and design of a control system. Key Keywords fuzzy logic; intelligent control; modelling; neural networks . systems , adequate performance is often only achieved through manual and heuris-.

Some features of this site may not work without it. Application of neural networks to temperature control in thermal power plants Cui, Xianzhong; Shin, Kang G. Cui, Xianzhong; Shin, Kang G. To overcome these problems, an adaptive controller has been designed using neural networks, and tested extensively via simulations. One of the key problems in designing such a controller is to develop an efficient training algorithm. Neural networks are usually trained using the output errors of the network, instead of using the output errors of the controlled plant. However, when a neural network is used to control a plant directly, the output errors of the network are unknown, since the desired control actions are unknown.

Thus, a hybrid approach is desirable, in which emerging memory devices provide the dynamics for neuromorphic functions and transistors supply the signal amplification to enable larger and multi-layer networks 32 , Moreover, such hybrid devices could exploit alternative mechanisms, such as capacitive rather than resistive coupling, to interact with its synapses, featuring a low static power dissipation to trigger active neuron operations 34 , 35 , We recently exploited the time-dependent electrical conductance resulting from the Ag mass migration due to the combined electrochemical 37 , 38 , 39 , 40 and diffusion processes 41 in a dielectric medium between two electrodes responding to an applied voltage to emulate the dynamics of ion channels in neurons 31 , Input signals are integrated by the new element that mimics dendritic spatial and temporal summation.

Results Temporal signal summation of dynamic pseudo-memcapacitors An ideal memcapacitor 36 , 42 , 43 , if one existed, would possess a bias-history-dependent capacitance, similar to a memristor with resistance determined by the past inputs 44 , Different realizations of switching capacitors have been proposed 46 , 47 , 48 , 49 , 50 , 51 , 52 , 53 , 54 , 55 , For instance, electrochemical capacitors 56 , bias-dependent polarization 46 , 48 , and nanobattery effect 39 , 57 may all have the potential to implement memcapacitive systems.

We developed a unique dynamic pseudo-memcapacitor DPM by integrating memristor and capacitor structures, as shown in Fig. Hysteresis loops of the logarithm of the charge vs. The physical origin of the diffusive memristor threshold switching has been investigated and explained by its innate electro-thermal and ion migration dynamics 37 , 41 , A non-volatile memristor can replace the diffusive memristor in the structure shown in Fig.

Compared to memristors, DPM features low power consumption. As DPMs store energy in the electrostatic field rather than converting electricity to heat, the DPM-based capacitive networks feature low energy operation compared to networks built on other emerging devices e. The electrostatic energy is proportional to the capacitances of the elements.

It shall also be noted, charging DPMs or capacitive elements takes energy away from the signal sources while discharging DPMs or capacitive elements returns the electrostatic energy back to the signal sources, as they essentially provide temporary energy storage. Temporal signal summation is one of the fundamental functions performed by a single neuron, powered by the electrochemical gradients 39 , 57 , which has been mathematically modeled by the Hodgkin—Huxley model The summation of signal in time is associated with the switching of the voltage-gated sodium and potassium ion channels, which integrate the post-synaptic potentials and initiate the subsequent action potential.

As shown in Fig. Once the membrane potential exceeds the threshold at time t2, the fast inward-flow of sodium ions results in a significant further rise of the membrane potential. This positive feedback raises the potential rapidly until all available sodium ion channels are open, leading to the observed large upswing of the membrane potential in Fig.

IEEE Xplore - Conference Table of Contents

Once reaching the maximum, the membrane experiences repolarization at time t3 due to the inactivation of the sodium ion channels and the opening of the potassium ion channels. By virtue of the biomimetic Ag dynamics 41 , the response of a membrane potential to an input pulse train was replicated by the DPM. The electrostatic energy is proportional to the capacitances of the elements.

It shall also be noted, charging DPMs or capacitive elements takes energy away from the signal sources while discharging DPMs or capacitive elements returns the electrostatic energy back to the signal sources, as they essentially provide temporary energy storage.

Temporal signal summation is one of the fundamental functions performed by a single neuron, powered by the electrochemical gradients 39 , 57 , which has been mathematically modeled by the Hodgkin—Huxley model The summation of signal in time is associated with the switching of the voltage-gated sodium and potassium ion channels, which integrate the post-synaptic potentials and initiate the subsequent action potential.

As shown in Fig. Once the membrane potential exceeds the threshold at time t2, the fast inward-flow of sodium ions results in a significant further rise of the membrane potential. This positive feedback raises the potential rapidly until all available sodium ion channels are open, leading to the observed large upswing of the membrane potential in Fig.

Once reaching the maximum, the membrane experiences repolarization at time t3 due to the inactivation of the sodium ion channels and the opening of the potassium ion channels. By virtue of the biomimetic Ag dynamics 41 , the response of a membrane potential to an input pulse train was replicated by the DPM.

As the charge over the series capacitor increased, the voltage across the series capacitor rose with each subsequent pulse, closely reflecting the expected behavior illustrated for a neuron at t1 in Fig. The next pre-synaptic input spike would switch OFF the diffusive memristor, and start a new cycle of integrate-and-fire process. The volatility of the diffusive memristor essentially provides the repolarization and self-inactivation features of ion channels in neurons, which differs from non-volatile memristive neurons that require RESET pulses 8.

Compared to volatile memristive neurons with parallel capacitance 7 , 9 , 31 , the non-polar diffusive memristor of the DPM plays the role of sodium ion channels in the phase of integration and the role of potassium channels to drain the stored charge and recover the potential in the phase of repolarization. Compared to single memristive neurons with intrinsic analog switching 8 , 11 , 31 , DPMs provide physical embodiments of the cell membrane and ion channel, leading to the observed fidelity of neural function emulation.

The hysteresis loops reveal the volatile switching between the intrinsic parallel capacitor CP of the memristor and the series capacitor CS.

At time t1, the membrane potential did not reach the threshold so few sodium channels were open upon the arrival of the signal. The memristor was first switched OFF and then switched ON again with opposite bias, which quickly discharged the capacitor and brought the membrane potential back to its resting value, similar to the repolarization caused by the opening of potassium ion channels.

The gate voltage mimics the membrane potential at the soma and the drain current replicates the electric charge spike or action potential flowing down the axon of a neuron see Fig. A train of input pulses is integrated on the neuro-transistor gate to create a voltage spike as shown in Fig.

When the interval between input voltage spikes was increased, the volatile filament growth of the diffusive memristor meant that more spikes were required to build up the charge on the transistor gate to fire an action potential. Spatial signal summation and propagation of neuro-transistors The dendritic tree of each biological neuron interfaces with adjacent branched axon projections via synapses to propagate the electrochemical stimulation from neighboring neurons to its own soma The morphology of dendrites varies, providing neurons with different functionalities Samples of each of the phase current are compared with the samples of the same phase current taken half cycle and one cycle before.

Fault selector module outputs are made based on the combination of the current samples using Equations 3 — 5.

[PDF] Neural Network Design (Electrical Engineering) Download Full Ebook

If the fault classifier output signals exceed a pick-up threshold for three consecutive samples, a fault on the relevant phase is detected. The level above which the classifier operates could be set to prevent operation in the case of system load changes.

The outputs of the three fault classifiers are ORed according to Equation 6 to make the fault detection signal. A further characteristic is also used to detect fault type changes. One cycle of data of each of the phase currents prior to the fault incidence is stored in the memory. Fault current samples are compared to the prior fault current samples based on the same principle used in Equations 3 — 6 to identify the fault type changes.

Networks inputs selection A number of protection applications using ANNs have been suggested in the literature. In some applications, voltage and current input information is processed first by a conventional algorithm, for example a DFT-based distance measurement relaying algorithm, to calculate the appropriate phasor. The output of this algorithm then is processed by suitable neural networks [9,10].

It makes the total algorithm more complex, and higher processing and computational capabilities are required to process the whole algorithm. In this application, samples of current and voltage signals are directly processed by neural networks. Consecutive samples of three-phase voltages and currents as well as neutral current are chosen as the inputs.

The appropriate data window length is a major factor, which should be considered. Thus, each phase voltage and current and neutral current was represented by its five consecutive samples. Networks structure Multilayer feedforward networks were chosen to process the sampled input data. Number of the networks to estimate the fault distance was decided by experimentation, which involved training and testing different number of networks.

Design electrical engineering pdf network neural

As shown in Figure 2, four different networks were finally selected to process different fault types, that is phase to ground faults, phase to phase to ground faults, phase to phase faults, and three-phase faults. Based on the fault type, appropriate network is selected and activated by the fault classifier. Appropriate inputs to each of the networks are also selected by the fault classification module. For the faults involving ground, the neutral current samples are also selected as the network inputs.

For example, in case of phase A to ground fault AG, samples of the phase A voltage and current signals are selected as the inputs of the first network, the phase to ground network. This network therefore, should have 15 inputs five consecutive samples of each of the voltage and current signals of phase A as well as the neutral current. In the case of a phase B to ground fault, samples of the phase B voltage and current signals as well as the neutral current are selected as the first network inputs.

Structure of the selected neural networks. Network type Structure Final error training Final error test Single phase to ground 0. Samples of voltages and currents of both phases A and B as well as the neutral current are selected as the inputs of the appropriate network. Therefore, the phase to phase to ground neural network has 25 inputs. Similarly, the phase to phase and three-phase networks should have 20 and 30 inputs, respectively.

As these faults do not involve ground, the neutral current samples are not considered as the network inputs. For these types of faults, fault resistance is quite small and distance function measurements are not usually affected by fault resistance. A single neuron output layer was chosen for each of the networks.

If a fault is located in the tripping region, the network target output should be 1, otherwise it should be 0. For each of the four above-mentioned neural networks, a few different network structures with appropriate number of neurons in their hidden layers were considered.

Different networks with one and two hidden layers were considered and trained. On the other hand, networks with two hidden layers provided better results without having to have high number of neurons in their only hidden layers. The suitable networks, which showed satisfactory results and were finally selected, are shown in Table V. For all the networks, Hyperbolic tangent function is used as the activation function of the hidden layer neurons.

Saturated linear function is used as the output layer activation function [14]. ML algorithm is a nonlinear least square algorithm applied to the batch learning of multilayer perceptrons. It was found that the networks trained with the ML algorithm provide better results compared with the results of the networks trained with the BP algorithm. Therefore, it was decided to use the ML training algorithm for this application.

Once trained, the networks performance was tested using test patterns that were different from the training patterns. Some of the simulation results are presented in the next section. The validation set fault patterns were different than the fault patterns used to train the network.

Pdf engineering network neural electrical design

For different faults of the validation set, fault type, fault location, fault inception time, and pre-fault power flow direction were changed to investigate effects of these factors on the performance of the proposed algorithm. Extreme cases like faults near to the protection zone boundary including fault resistance were also included in the validation dataset.

For the studies performed in this section, the first protection zone was set to 80 km. Distance relay operation time test results. In this study, three consecutive outputs of the appropriate network are averaged. Averaged outputs which are greater than 0. A DFT-based impedance measurement distance relay was also simulated. Performance of the conventional distance relay was compared with that of the proposed ANN-based relay.

The proposed and conventional distance relays operation times for a few faults with different power system conditions are presented in Table VI. As an example, test results for a single phase to ground, AG fault at 20 km from the relay location is presented in the first row of the Table VI.

For this fault, the relative angle of the sending-end source with respect to the angle of the receiving-end source was 10 degrees. Also the fault inception angle with respect to phase A voltage zero crossing was 30 degrees. Relay operation time for three different amounts of fault resistance is shown in the last three columns of the table.

For the faults which involve ground, the relay operation time for 0, 5, and 10 V fault resistance is investigated.

For the faults which do not involve ground, fault resistance is not a critical factor. Therefore, only relay performance without fault resistance is investigated.

As shown in Table VI, the relay performs quite fast and reliably. The proposed relay operated faster and is more reliable than the conventional distance relay simulated by DFT algorithm. As shown the proposed relay has operated for faults inside the protection zone and close to the zone boundary.

However, the conventional relay has not been able to operate appropriately for some of these faults. For most of the faults inside the first protection zone, the relay is able to respond in just about 6 ms after the fault inception. For the faults including 10 V fault resistance, for which the conventional relays response is slow in nature, the proposed distance relay operates on average in just about 8 ms. For some of these faults, the conventional relay operates unnecessarily.

For the faults close to the boundary zone, performance of the extended ANN-based algorithm was checked for both positive and negative power angles.

Design pdf electrical engineering network neural

The relay does not encounter with the problem of improper action even for the case when fault resistance is 10 V. For the faults close to the boundary zone, the conventional distance relay performance is not reliable as shown in Table VI. However, the ANN-based network described in the previous work which used less number of inputs was not able to distinguish some of these faults correctly.

The proposed relay output for a few other faults with different power system conditions is shown in Figures 5—8. For these cases, a fault is applied to the system at 20 ms and the relay output is shown for the first 50 ms after the fault inception, which is of utmost interest. The relay output for a single phase to ground fault BG, is shown in Figure 5. Fault location was 78 km from the relay location, with 5 V fault resistance. The relay averaged output is shown in Figure 5.

As shown in this figure, the relay is able to respond to the fault correctly in a timely fashion. Figure 6 shows that the relay performs fast and correctly. Figure 7 shows the output of the relay for a fault at 82 km. It shows that the relay output remains stable for this fault, which is beyond the protection zone.

To examine capabilities of the proposed relay against variation of type fault, an evolving fault case is simulated.

Login using

An ABG fault is applied at 77 km from the relay location and the relay output is shown in Figure 8. For this test, type fault was changed to ABC 30 ms later. As shown in this figure, the relay performed correctly and quite fast even for this complicated case which included fault type change. The relay output for a few faults with different power system conditions is presented in this section. In general, the network performs better and faster for more usual fault cases.

Figure 5. Similar results are obtained using the proposed relay in the new power system. If the training set of a neural network is selected appropriately covering a wide range of conditions, the trained network would be able to distinguish system faults for a new power system as well.

This fact is also shown in some other research studies using neural network based protective relays [8,16]. Relay output for AB fault at 82 km. Hardware configuration of the implemented distance relay is shown in Figure 9. Architecture of the minimum system card is depicted in Figure Hardware configuration of the implemented distance relay. Configuration of the minimum system card. This card controls the data-acquisition process. It also processes the acquired signals and implements the proposed technique.

Electrical neural pdf engineering design network

Appropriate hardware and software set-up for accomplishing data acquisition were designed and fabricated. The prepared digital data is then introduced to the minimum system card.

The sampling frequency fs used in this work is 1 kHz. To attenuate high frequency components, an array of low-pass filters is used.

Other features of the selected components are provided in Table VII.