Application of Copula Quantile Regression Method in Wind Power Ultra Short Term Output Prediction

In recent years, the proportion of renewable energy generation in China's power industry has been increasing, and the installed capacity has surpassed that of coal-fired power. However, wind power output has strong randomness and volatility. Compared with traditional prediction, quantile prediction methods can provide comprehensive probability distribution information of wind power and achieve more reliable wind power output prediction, which is of great significance for the safe and stable operation of the power grid system. A quantile regression method based on Copula (QCopula) is proposed to address this issue. The advantage of Copula function is that it can describe the correlation between the marginal distribution function of random variables and the joint distribution function between variables. Firstly, the optimal Copula function is determined using the AIC criterion. Based on the correlation between wind power and wind speed described by the Copula function, the conditional probability distribution function of power is expressed. Secondly, by taking different conditional probability values, wind power prediction results at different quantiles are obtained, and then interval prediction results with different confidence intervals are obtained. These results are compared with three traditional quantile regression methods (QR, QRF, QLSTM), And three indicators, PICP, PINAW, and CPIA, were used to evaluate the interval prediction results of the four quantile regression methods. Finally, the crossover of quantile curves for each method was analyzed. This article takes a wind power plant in Gansu Province as a case study, with wind speed and power data (in MW, with an interval of 15 minutes) from September 2022 to June 2023. There are a total of 29088 sample points, and the data is divided into training and testing sets in an 8:2 ratio. The training set is used to establish models using various quantile regression methods, and the testing set is used to verify the accuracy of each model. The results showed that under different confidence intervals, the accuracy range of QCopula's modified prediction interval was between 0.701 and 0.773, with an average of 15%, 9%, and 13% higher than QR, QRF, and QLSTM, respectively, and better than the other three quantile prediction methods. In quantile cross validation, QCopula did not exhibit quantile cross validation, and the predicted power values for each sample point monotonically increased with probability. However, QR, QRF, and QLSTM all exhibited varying degrees of quantile cross validation. In summary, QCopula can characterize smaller interval widths and higher interval coverage, and the quantile curve does not cross, resulting in higher reliability.