ETD system

Electronic theses and dissertations repository

 

Tesi etd-08292018-092234


Thesis type
Tesi di laurea magistrale
Author
CAMPINOTTI, DAVIDE
URN
etd-08292018-092234
Title
Multi-step ahead short-term wind power forecasting for a wind farm exploiting machine learning algorithms
Struttura
INGEGNERIA DELL'ENERGIA, DEI SISTEMI, DEL TERRITORIO E DELLE COSTRUZIONI
Corso di studi
INGEGNERIA ELETTRICA
Supervisors
relatore Tucci, Mauro
tutor Betti, Alessandro
Parole chiave
  • machine learning
  • neural network
  • svm
  • wind power nowcast
  • short term forecast
Data inizio appello
28/09/2018;
Consultabilità
Secretata d'ufficio
Data di rilascio
28/09/2088
Riassunto analitico
In the wake of the ever growing level of wind power penetration into the electric grid, many a challenge has been posed to the Transmission System Operators (TSOs) which aim is to guarantee both the system quality and reliability. The presence of a very variable and hardly predictable variable as wind is, may jeopardize the balance that should be kept between production and consumption lest it occur a frequency transient.

Aim of this work, is to predict with the due accuracy the electric power generated up to 6 hours ahead with a sampling rate of 10 minutes of a 18 MW wind farm located in the south of Italy.


To this end both historic and NWP data have been exploited and, albeit some researchers had considered the latter of no avail for such a short time horizon, this work has widely shown the opposite.
These data, especially historic data, have undergone an intensive pre-processing phase so as to get rid of not meaningful data and compute their availability.
The pre-processed data have been fed to a bunch of selected machine learning algorithms whose training strategies had been previously evaluated on a validation set in order to tune accurately their hyper-parameters and find the most performing ones.

A ensemble structure have thus been developed to use these algorithms and obtain a even better predictive performance taking advantage of the single algorithm best features and compensating its weaknesses.

Eventually, a post-process phase occurs whence, inter alia, a metric error has been evaluated to provide a way to quantitatively compare the performance of competing models amongst each others and with the benchmark.
File