Ensemble time series forecasting with applications in renewable energy
Date of Issue2016-04-07
School of Electrical and Electronic Engineering
Renewable/sustainable energy sources draw increasing attention of researchers due to the shortage of fossil fuel. The fossil fuel also has adverse impact on our environment. However, because of the intermittent nature of the renewable energy sources, it is difficult to be integrated into the power grid. Accurate forecasting is required for improving the reliability of power dispatch, unit commitment, reducing the cost of energy storage devices, optimizing the operation and maintenance schedule. Besides time series forecasting, wind power ramp classification and detection is also covered in the thesis due to its importance for wind farm operation and power integration. The first part of the thesis focuses on the development of ensemble forecasting methods related to renewable energy: wind speed and wind power and solar irradiance. Firstly, state-of-the-art ensemble time series forecasting methods are reviewed and categorized into competitive ensemble forecasting and cooperative ensemble forecasting. The advantages and limitations of each category are discussed. Secondly, empirical mode decomposition (EMD) based time series forecasting methods are proposed and discussed. A support vector regression (SVR) takes the decomposed and re-constructed feature sets from EMD for short-term wind speed forecasting and the proposed model has outstanding performance. Next, the ensemble versions of EMD are investigated. An AdaBoost based EMD-artificial neural network (ABEMD-ANN) is proposed and the performance on wind speed forecasting outperformed the EMD-ANN. Another ensemble approach to EMD is noise assistance and among the various realizations. Complete ensemble EMD with adaptive noise (CEEMDAN) based SVR has the best performance among the ensemble EMD based forecasting methods on wind speed forecasting. The performance of ensemble EMD based SVR methods on solar irradiance forecasting are also superior to other state-of-the-art forecasting methods. The second part of the thesis focuses on the analysis of random vector functional link (RVFL) neural network and its applications in wind speed and solar irradiance forecasting and wind power ramp forecasting. First the importance of direct input output connections in the RVFL network is studied based on the comparison of eight different RVFL configurations and two hierarchical ensemble forecasting methods on wind speed forecasting. Then the RVFL network is applied to classify the significant wind power ramp in the next 6 or 12 hours. The performance of RVFL network on power ramp classification outperformed other popularly used machine learning methods. Finally, some future research directions are pointed out.
DRNTU::Engineering::Mechanical engineering::Alternative, renewable energy sources