Short-Term Wind Speed Forecasting Using Neural Network Models for Chanbin Offshore Area

  • 孫 億昇

Student thesis: Doctoral Thesis


In 2016 Taiwan’s government published a new energy policy mandating making Taiwan a nuclear-free island by 2025 Offshore wind energy has become an important project to fill the energy gap after the shutdown of nuclear power plants The biggest problem with wind power is wind instability A precise forecasting model not only smooths the operation of wind farms but also reduce damage to the power grid In this study wind data from 2017 to 2019 from the Taipower Meteorological Mast in Chanbin offshore is used in the analysis and model training A short-term multi-step wind forecasting model is established based on an artificial neural network model and is optimized by changing the model parameters The data analysis is divided into three parts: a trend analysis a statistical analysis and a spectrum analysis The trend analysis uses a time series plot and a wind rose plot to describe the trends in wind speed and direction over a three-year period The statistical analysis explores the similarities and differences between each year using statistical values Finally the spectrogram obtained by the Wavelet Transform explains the features of the different wind sources Based on the previous analyses the accuracy and a further error analysis of the forecasting results are discussed The LSTM model is the main model used in this study The parameter tuning focuses on the input data in an attempt to obtain better performance from the data rather than from the model At the end of this study we found the wind regime in spring autumn and winter to be similar every year and to be easily affected by the southwest monsoon and typhoons in summer The average wind speed is high in autumn and winter but low in spring and summer The Wavelet Transform results showed that the wind conditions can be roughly classified into northeast wind southwest wind and local wind The accuracy (R^2 value) of the multi-step forecasting model reached 0 991 0 981 and 0 970 respectively in the first three step prediction Although parameter tuning did not significantly improve the accuracy of the forecasting mode it greatly improved our understanding of the forecasting result
Date of Award2020
Original languageEnglish
SupervisorTa-Hui Lin (Supervisor)

Cite this