Prepoznava dogodkov v elektroenergetskem omrežju s pomočjo strojnega učenja na podlagi Fourierove vrste

VILIĆ, MAJ (avtor), Rudež, Urban (mentor)

Sodobne PMU (ang. Phasor Measurement Unit) naprave so zmožne spremljati dogajanje v EEO v realnem času. To dopušča možnost za boljši nadzor in odzivnost na dogodke v EES. RTDS (ang. Real Time Digital Simulator) nam omogoča simulacije, ki so zelo dober približek realnim razmeram v EE omrežjih, kjer uporabljamo naprave PMU. Z uporabo RTDS so bile pridobljene meritve 7 tipov dogodkov, do katerih lahko pride v EE omrežjih. Pridobljene meritve na simulacijskem RTDS modelu so bile analizirane v sklopu programskega jezika Python. V nalogi je opisan način detekcije dogodkov v EES in njihova prepoznava. Na motnje v EES se je prilegalo Foruierovo vrsto v kombinaciji z metodo najmanjših kvadratov. Na ta način so bile pridobljene matematične funkcije, katere so bile uporabljene kot pripomoček za pridobivanje značilk. Slednje so omogočale podlago za izvajanje prepoznave dogodkov v EES. Fourierova vrsta je bila uporabljena na dva načina. Pri prvem se je pridobivala informacija o trendu obravnavanega signala motnje v EES, pri drugem pa informacija o pogostosti krožnih frekvenc v prilegani funkciji na motnjo v EES. Za izbiro najboljšega prileganja Forurierove vrste je bila v prvem primeru uporabljena statistika prilagojen R2. Magistrska naloga vključuje uporabo dveh algoritmov strojnega učenja, in sicer i) algoritem metode podpornih vektorjev in ii) algoritem K-najbližjih sosedov. Za izpopolnitev slednjega sta v nalogo vključeni dve metodi za zmanjšanje dimenzij; analiza glavnih komponent in analiza sosednjih komponent. Vse uporabljene metode strojnega učenja so bile večinsko opravljene s pomočjo knjižnice Scikit-learn. Skladno z uporabo dveh različnih velikosti učne množice je opisan celotni postopek, od učenja in validacije do izbire in nastavitve parametrov algoritmov strojnega učenja. Naloga podaja primerjavo učinkovitosti obeh verzij algoritmov za pridobivanje značilk, metod strojnega učenja in načinov za zmanjšanje dimenzij. Rezultati primerjav so razdeljeni na manjšo in večjo učno množico, kjer so opravljeni tako testni rezultati kot tudi rezultati napovedi na novih meritvah raznih dogodkov v omrežju.

• Jezik: Slovenski jezik
• Ključne besede: RTDS simulacije, strojno učenje, dogodki v EES, podatkovna znanost, Fourierova vrsta
• Vrsta gradiva: Magistrsko delo/naloga
• Organizacija: FE - Fakulteta za elektrotehniko
• Leto izida: 2021
• PID: 20.500.12556/RUL-126282
• Datum objave v RUL: 14.04.2021

Sekundarni jezik

• Jezik: Angleški jezik
• Naslov: Electric-power system event recognition based on machine learning with Fourier series

Izvleček

Modern Phasor Measurement Unit (PMU) devices are capable of monitoring real time events in the electrical-power networks. This allows better control and faster responses when network faults occur. The Real Time Digital Simulator (RTDS) device allows one to perform digital simulations, which represent good approximation of the real signals in power grids when using PMU devices. Using RTDS, 7 types of faults occurring in energy networks were obtained. The data obtained from RTDS was analysed in the Python programming language. The research contains descriptions of electrical-power networks faults recordings in energy networks and procedures on how to identify them. The Fourier series in combination with the least squares method was fitted to the power grid fault signals. By doing this, mathematical functions were obtained, which were then used as a tool for obtaining features. Those functions provided groundwork for energy grid fault signal recognition. The Fourier series was used in two ways: i) Fourier series was used to gather information regarding signal trend. ii) Fourier series was used so we could acquire represented angular frequencies of a fitted function on faulted signal in electric power system. In the first option adjusted R2 statistics was used to select the best Fourier series fit. The thesis includes usage of two machine-learning algorithms, namely the Support vector machines and K-nearest neighbours. For further improvement two dimension reduction algorithms were applied on K-nearest neighbours method; Principal Component Analysis (PCA) and Neighbourhood Components Analysis (NCA). All the machine learning methods used were mostly done with the help of the Scikit-learn library. The whole procedure from testing and validation to parameter tuning and parameter selection is described in accordance with two used training sets of a different size. The research provides comparisons of the performance for both versions of feature acquisition algorithms, machine-learning methods, and of ways to reduce dimensions. The results of all comparisons are split into smaller and larger training sets, where both test results and prediction results on new unseen energy networks signal faults are evaluated on.

• Ključne besede: RTDS simulations, machine learning, energy system faults, data science, Fourier series