The Probability Density Function Based Neuro-Fuzzy Wind Power Prediction With Global Convergence

Reliable and accurate wind power prediction is fundamental for efficient operation of wind energy system. However, the stochasticity and volatility cause a lot of colored noises, making wind power modeling very complicated. Many of the traditional prediction models lack the ability to cope with these noises, which is often not sufficient in practice for the accurate estimation of the wind power. To address this issue, this study develop a probability density function based neuro-fuzzy wind power prediction method with global convergence. The probability density function (PDF) of modeling error is introduced as a metric to evaluate the performance of the neuro-fuzzy wind power prediction model. More specifically, the parameter updating of neuro-fuzzy model is transformed into the shape control of the modeling error PDF. It means the modeling error PDF is controlled to close to an ideal PDF, so that wind power prediction model parameters are tuned dynamically. In addition, in order to decrease the fluctuation of raw sequence, an improved multivariate variational mode decomposition is applied to synchronously decompose wind speed and power into multiple sub-series, overcoming time-frequency mismatch among different wind data. Moreover, the existing literature widely focuses on combination of sub-models by optimization algorithms, which easily lead to local optimum and experiences slow convergence. It may not describe the real physical relations between sub-series. Thus in our research the wind power forecasting results of multiple sub-models are integrated by an updating method based on the Lyapunov theory, which can guarantee the global convergence of the model weights. Simulation experiments verify that the suggested method can effectively suppress the colored noise in the wind power series, guaranteeing the robustness and accuracy of the model.