A Markov Regime Switching Model for Ultra-Short-Term Wind Power Prediction Based on Toeplitz Inverse Covariance Clustering

Abstract

Read online

The rapid development of wind energy has brought a lot of uncertainty to the power system. The accurate ultra-short-term wind power prediction is the key issue to ensure the stable and economical operation of the power system. It is also the foundation of the intraday and real-time electricity market. However, most researches use one prediction model for all the scenarios which cannot take the time-variant and non-stationary property of wind power time series into consideration. In this paper, a Markov regime switching method is proposed to predict the ultra-short-term wind power of multiple wind farms. In the regime switching model, the time series is divided into several regimes that represent different hidden patterns and one specific prediction model can be designed for each regime. The Toeplitz inverse covariance clustering (TICC) is utilized to divide the wind power time series into several hidden regimes and each regime describes one special spatiotemporal relationship among wind farms. To represent the operation state of the wind farms, a graph autoencoder neural network is designed to transform the high-dimensional measurement variable into a low-dimensional space which is more appropriate for the TICC method. The spatiotemporal pattern evolution of wind power time series can be described in the regime switching process. Markov chain Monte Carlo (MCMC) is used to generate the time series of several possible regime numbers. The Kullback-Leibler (KL) divergence criterion is used to determine the optimal number. Then, the spatiotemporal graph convolutional network is adopted to predict the wind power for each regime. Finally, our Markov regime switching method based on TICC is compared with the classical one-state prediction model and other Markov regime switching models. Tests on wind farms located in Northeast China verified the effectiveness of the proposed method.

Keywords

• wind power prediction
• clustering
• pattern division
• markov regime switching model
• machine learning