Title Vėjo jėgainių generacijos laiko eilučių analizė ir prognozavimas /
Translation of Title Wind turbines generation time-series analysis and forecasting.
Authors Slivikas, Aivaras
Full Text Download
Pages 78
Keywords [eng] wind ; generation ; forecast ; time ; series
Abstract [eng] The aim of this master thesis is to reduce electricity system imbalance prices with more accurate wind power plants generation forecasts. The main objectives were: 1) to develop wind turbines generation forecast models which are eligible to forecast up to 38 hours horizon ahead. 2) To evaluate accuracy of different forecast models. 3) To compare developed model errors with two systems forecasts output. 4) To combine systems forecasts with exogenous variables and achieve the better result than any of those systems separately. 5) To calculate financial benefit, using the most accurate models results. This master thesis represents wind power turbine generation time series analysis and forecasting. 17 different models have been tested. 17 univariate models, with input of forecasted time series historical data have been developed. Selection of exogenous variables, which have strongest impact to time series values, has been performed. 4 multivariate with exogenous variables models have been tested. Multivariate forecast modelling has been performed in two following ways: 1) with constant training set and 2) with periodically growing training set. Three different ensemble models have been created. The forecasts have been produced for thirteen different wind turbine plants and sum generation time series. Each model has been evaluated with its performance error. Financial benefit has been introduced using the best performing models results. The research shows that multivariate models are more advanced than univariate. Model ensembles give positive results and ensembles most of the times provides with more accurate result than any of the models individually. Train set re-evaluation gives slightly more positive result than no re-evaluation, although calculation resources are way higher. The most accurate of the univariate models are ARIMA and NNETAR. The most accurate of the multivariate models are ARIMAX and LOESS regression. Two system results have been outperformed only when their forecasts have been included in the models.
Dissertation Institution Kauno technologijos universitetas.
Type Master thesis
Language Lithuanian
Publication date 2019