diplomsko delo
Abstract
V diplomskem delu predstavimo napovedovanje multivariatnih časovnih vrst z uporabo povratnih nevronskih mrež, ter primernost pristopa k napovedovanju preizkusimo na področju energetike. Za pametno krmiljenje električnih naprav je namreč nujno potrebno poznavanje posledic, ki jih imajo naše akcije na stanje naprav in njihove okolice. Stanje naprav definira več spremenljivk, zato spreminjanje stanja skozi čas opisuje multivariatna časovna vrsta. Za električno napravo grelnik vode pripravimo napovedni model, ki temelji na povratni nevronski mreži arhitekture LSTM. Ker pa se lastnosti naprave in s tem opisujoče časovne vrste lahko s časom spreminjajo, moramo za ohranjanje natančnostosti napovednega modela le-tega sproti prilagajati. V diplomskem delu predstavimo različne strategije sprotnega učenja modela in primerjamo njihovo učinkovitost na napovednem modelu za grelnik vode.
Keywords
multivariatne časovne vrste;napovedovanje časovnih vrst;povratne nevronske mreže;katastrofalno pozabljanje;diplomske naloge;
Data
Language: |
Slovenian |
Year of publishing: |
2020 |
Typology: |
2.11 - Undergraduate Thesis |
Organization: |
UM FERI - Faculty of Electrical Engineering and Computer Science |
Publisher: |
[N. Uremović] |
UDC: |
004.032.26:004.8(043.2) |
COBISS: |
42183427
|
Views: |
641 |
Downloads: |
76 |
Average score: |
0 (0 votes) |
Metadata: |
|
Other data
Secondary language: |
English |
Secondary title: |
Time series forecasting by using recurrent neural networks |
Secondary abstract: |
In this thesis we present the forecast of multivariate time series, by using recurrent neural networks, and test the adequacy of such approach to forecast on the field of energetics. For smart management of electric devices, it is crucial to understand the impact of our actions on the machine and its surroundings. The state of the machine is defined by multiple variables, which means the change of the state through time is defined by a multivariate time series. We prepare a forecast model for a machine water heater, basing on a recurrent neural network of the LSTM architecture. Because the characteristics of the machine, and thus the time series describing it, can change over time, it is necessary to simultaneously adapt the forecast model, to assure the accuracy of the model is preserved. In this thesis, we present different strategies for model adaptation and compare their effectiveness on the forecast model for a water heater. |
Secondary keywords: |
multivariate time series;time series forecasting;recurrent neural networks;catastrophic forgetting; |
Type (COBISS): |
Bachelor thesis/paper |
Thesis comment: |
Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Računalništvo in informacijske tehnologije |
Pages: |
IV, 28 f. |
ID: |
11952903 |