(572e) Renewable Wind Power Prediction and Optimization Using Gaussian Mixture Copula Model and Bayesian Inference Based Local-Global Gaussian Process Regression Method

Renewable energy has attracted considerable attention recently because of the deterioration of the environment and the depletion of conventional energy resources. As a non-pollution renewable energy source for electricity generation, wind energy is increasingly attractive and has been well recognized as environmentally friendly, socially beneficial and economically competitive. The efficiency of wind turbines can be significantly improved if the wind farm operation is controlled and optimized through accurate wind speed forecasting for load requirements and maintenance scheduling. However, the intermittency of wind power generation and the stochastic, non-stationary and unpredictable nature of wind speed pose significant challenges as to the most efficient utilization of the wind energy source.

To overcome the limitations of conventional wind speed prediction methods, a Gaussian mixture copula model (GMCM) based approach is first applied to the wind speed data to identify the multi-seasonality and non-Gaussian component of diverse atmospheric conditions that influence wind speed. Then a Bayesian inference based local-global Gaussian process regression (GPR) approach is developed for wind speed prediction. Further, Gaussian process regression models within localized copula components are aggregated using the Bayesian inference based posterior probabilities into a global model for time series prediction of wind speed. The integrated GMCM-GPR approach is able to model the stochastic and non-stationary wind speed with appropriate handling of system uncertainty and multi-seasonality.

The proposed approach is demonstrated using real wind speed data from various locations in the US and compared against the auto-regressive integrated moving average (ARIMA) and support vector regression (SVR) methods. The comparison indicates that the GMCM-GPR approach provides more accurate and reliable predictions on wind speed than ARIMA and SVR methods. The intermittency and random fluctuations in wind speed that are difficult to characterize using regular dynamic modeling techniques are well captured by the presented approach.