Išlić, Marko

Centralizirana zaštita razgranatoga odvoda u distribucijskom elektroenergetskom sustavu korištenjem umjetnih neuronskih mreža : doktorski rad / mentor Ante Marušić - Zagreb : M. Išlić; Fakultet elektrotehnike i računarstva, 2021. - ix, 127 str. : ilustr. u bojama ; 30 cm. + CD-ROM

Bibliografija str. 120-124.

Ova doktorska disertacija bavi se prepoznavanjem kratkih spojeva na razgranatom odvodu u distribucijskoj mreži i to pomoću algoritma umjetne neuronske mreže koji dosad u dostupnoj literaturi nije opisan. Osim toga, za razliku od dosadašnjih rješenja zaštite ovaj sustav zaštite ima centraliziranu funkciju zaštite koja na jednom mjestu odlučuje o postojanju u mjestu kvara.
Cilj ovog algoritma je prepoznati kratki spoj i dati nalog za isključenje grane na kojoj je kvar nastao od ostatka zdrave mreže. Trenutno zastupljeni sustavi zaštite takve kvarove otklanjaju uz kratkotrajne beznaponske pauze. Ako bi se spomenuti problem išao rješavati strujnim stupnjevanjem izbjegla bi se potreba za beznaponskom pauzom, ali bi vrijeme za otklanjanje kvara bilo preveliko.
Za potrebe algoritma potreban je precizan i vjeran model odvoda iz kojeg se generiraju iznosi struja za potrebe učenja i ispitivanja umjetne neuronske mreže. Odabrani pokazni odvod je 34 čvorišni IEEE standardni odvod. Sve podatke o elementima modela i iznosi struja kratkog spoja za kvarove na sabirnicama su iz službenog izvještaja radne grupe IEEE PES. Iznosi struja kratkih spojeva na drugim mjestima i u drugim uvjetima koji mogu utjecati na iznose struja kratkog spoja izračunati su iz modela. U disertaciji su za modeliranje spomenutog modela korišteni MATLAB alati.
Rezultati su pokazali točno djelovanje sustava zaštite za dovoljno velik broj testiranih slučajeva.
U prvim poglavljima objašnjeni su pojmovi iz naslova disertacije, opis problema i dosad korištene sustave zaštite. U sljedećim poglavljima objašnjen je predloženi algoritam te promjene u arhitekturi sustava zaštite s razmatranjima svakog elementa tog novog sustava. U završnim poglavljima prikazani su i diskutirani uspješni rezultati.
Ključne riječi: umjetna neuronska mreža, zaštita; distribucijska mreža, klasifikacija, Backpropagation; IEEE; 34 čvorišta; razgranata mreža; nadstrujna zaštita; zaštita odvoda
The power distribution grid connects end customers and power transmission grid. Power transmission grid allows transmission of a huge amount of energy generated in power plants using high voltages, all that with negligible losses.
Power system protection is necessary for preventing damages in primary equipment such as lines, transformers, busbars, etc. that are caused by short circuits in a power grid. Protection devices should recognize such faults and instantly switch them off from the rest of the healthy grid. This is necessary to reduce the time of short circuit currents flowing within the grid i.e. reduce thermal and mechanical stress that can cause damage to a piece of primary equipment. Also, during the fault, it is very important that the healthy grid remains energized as much as possible.
Radial or meshed grids are built according to customer arrangement. In this thesis, only radial grids were taken into consideration. Opposite to meshed grids, radial grids are fed only from one bus bar (node). Losing this node leaves the rest of the grid without voltage.
A conventional protection scheme for radial feeders uses the coordination of a recloser and fuses, but it cannot locate and isolate the laterals where the fault occurred without a full feeder outage.
This scheme can be improved in the way that selectivity of the protection of such a grid can also be achieved by installing circuit breakers and relays instead of fuses, using the definite time current protection. The main problem of this approach is that the time required for fault locating can be very long (over three seconds, for a fault on the main feeder near the transformer). Such a long fault clearance time is not allowed by distribution system operators. In reality, eleven types of faults occur:
• short circuit between three phases ABC,
• short circuit between three phases and the ground ABCG,
• short circuit between two phases; AB or AC or BC,
• short circuit between two phases and the ground; ABG or ACG or BCG and
ground faults; AG or BG or CG.
The extensive development of computers and computational science has resulted in the application of computers in the area of electric power system protection in unprecedented ways.
Industrial computers have been continuously improved so that they can process increasingly demanding tasks. Without such improvements, the application of machine learning algorithms such as ANNs (artificial neural networks) would not be possible.
An ANN mimics the human nervous system and solves the most complex operations by doing many simple operations in parallel. If we assume that its development will be equally rapid in the near future, it will be possible to take a new, centralized protection scheme in the electric power distribution into consideration.
The proposed algorithm for decision-making from this thesis uses an ANN, which means that it is necessary to provide a large dataset for training and testing. Such a dataset is obtained from a simulation model made just for this purpose.
Every ANN has a specific:
• number of neurons in the output and input layers,
• number of hidden layers and
• number of neurons in the hidden layers.
A method for optimization of ANN parameters is also presented. The results showed that optimal ANN made for the IEEE 34-node radial feeder consists of one input layer with 51 neurons, one hidden layer with a minimum of 40 neurons and one output layer with 18 neurons.
In normal operation, three-phase current measurements processed in merging units are connected with the CIED (centralized intelligent electronic device). After the process of normalization and standardization is completed, these values are suitable for the input layer of the ANN.
Data normalization scales all short circuit values in an interval from minus one to one. This procedure speeds up the gradient descent (refers to backpropagation in neural networks) by avoiding many extra iterations that are required when one or more features take on much higher values than the rest.
Data standardization is performed to make the ANN more stable. This is done by dividing the difference between the average value of all values in the dataset by its standard deviation.
In the training stage, the current values obtained from the simulation model are used. In the hidden layer, a sigmoid activation function, typical of classification problems, is used. The outputs from the output layer issue a trip command to a specific circuit breaker through the communication infrastructure via IEC 61850 GOOSE messages.
The process of training the ANN, using the backpropagation algorithm, took place with 1,071,000 different simulated states, simulated especially for this purpose on the network model created in Matlab SIMULINK. Every faulty state generated by the network model had a different type of fault, fault impedance, ground impedance, fault location, system frequency and ambient temperature.
The simulation model consists of all feeder elements that have an impact on short circuit currents. It is possible to simulate a large number of various states that can affect the feeder (regular operation or faulty state) and have an impact on the current values. The current values obtained from the simulation are used as input vectors in the dataset. The outputs of the dataset are actions that affect circuit breakers (trip or no trip).
The comparison between the current measurements on the nodes during faults occurring in the simulation model and the study obtained from IEEE shows minor differences. The rest of the simulated faulty states that are not obtained by IEEE resulted in reasonable values.
Because of the novelty of the presented algorithm, it is impossible to compare the results from this study with any other studies, but 100% accuracy of the algorithm is shown. Total accuracy is necessary for the protection functions to provide full selectivity.
The centralized protection scheme consists of a CIED which collects and processes current measurement values obtained from current transformers on feeder laterals and makes a decision about which circuit breaker has to trip so that most of the feeder remains energized.
The described centralized protection system demands the new architecture of the protection system. The architecture presented in this thesis connects the CIED with current transformers installed on specific laterals of the feeder. So, it is necessary to install current transformers and circuit breakers in the right places and build effective communication infrastructure that provides fast and secure communication between them and CIED.
The installation of current transformers based on Rogowski coils is highly recommended because of their objective advantages.
CIED processes measurement values in real-time and, depending on the values, it causes the circuit breakers installed on the same feeder to trip. Permeable and reliable communication infrastructure for the interconnection between transformers, circuit breakers and the CIED was assumed.
This solution provides instantaneous tripping of the circuit breakers as well as the isolation of the part of the feeder where the fault occurred without causing a power outage in the rest of the healthy feeder.
Also, for the correct measurement sampling very precise time synchronization protocol must be used, for example, PTP (precision time protocol).
Communication infrastructure built for the purpose of such systems should be redundant and of high quality. A redundant scheme is provided with HSR (high seamless redundancy) and PRP (parallel redundancy protocol) approach. Constraints regarding the number of nodes of HSR to an accurate satisfactory level of quality are discussed. HSR communication is presented as an optimal scheme for described purposes because of economic criteria.
The complete method of the presented approach to the radial feeder protection consists of a simulation model for the dataset creation, the preparation of dataset values for the training and testing stage of the algorithm as well as an algorithm based on an ANN that makes a decision on whether the fault is present in the feeder and which circuit breaker must trip.
This thesis presents a new solution to the protection of a power distribution radial feeder supplied from a solidly grounded transformer. The novelty presented in this thesis is the new method for protecting a radial feeder in a power distribution system. The method consists of a protection system architecture and a new protection algorithm suitable for the described architecture.
The new protection system architecture consists of a CIED that has to be connected with current transformers and feeder circuit breakers.
This thesis confirms that there is a theoretical possibility to centralize the protection for radial feeders fed from a solidly grounded source with an accurate and fast algorithm.
Besides the fact that the deeper selectivity without losing voltage is achieved, by installing such a number of circuit breakers and current transformers, there are also some new possibilities for other systems such as control system and power quality monitor system.
In the introduction, basic terms from the title are defined and the problem and the proposed solution are presented. What artificial networks are and what they can achieve is discussed in the following chapters.
The creation and preparation of datasets from the simulation model with the complete proposed protection scheme is described. Finally, in the last chapters, results and conclusions are presented and discussed.
Keywords: artificial neural network; electric power distribution protection; classification; backpropagation; IEEE 34 node; radial; overcurrent protection; feeder protection

Središnja knjižnica Fakulteta elektrotehnike i računarstva, Unska 3, 10000 Zagreb
tel +385 1 6129 886 | fax +385 1 6129 888 | ferlib@fer.hr