Short-term photovoltaic power forecasting for photovoltaic power station based on EWT-KMPMR

As one of the renewable energy sources, photovoltaic generation technology has gradually become an important power generation method ranking only the second to the wind power generation technology, however, because of the uncontrollable influence factors that come from the day and night alternation and unstable meteorology condition, the output of photovoltaic power has intermittent and strong nonlinear characteristics unavoidably. At present, the short-term photovoltaic power forecast models in different areas are not same, and the physical forecasting method was used in the photovoltaic power station that is close to downtown and the area with rich meteorological observatory resources, however, due to its complex modeling and poor adaptivity, it is difficult to predict the photovoltaic station power output accurately when the weather changes suddenly. SVM (support vector machine) and other various computational intelligence methods have been used widely in the short-term photovoltaic power forecast, whose essence is to simulate operation law of the historical data to implement the photovoltaic power station output prediction, so it is still difficult to achieve a higher prediction precision when the data change severely or under abnormal weather conditions by the single forecasting method based on neural network. Based on empirical wavelet transform (EWT)-kernel minimax probability machine regression (KMPMR), a kind of combined forecasting method is proposed to improve the short-term photovoltaic power forecasting accuracy, and the photovoltaic output power on sunny days, cloudy days and rainy days is forecasted and analyzed respectively. EWT inherits the advantages of empirical mode decomposition (EMD) and wavelet transform, and takes the advantages of strong theorization, small amount of computation and fewer decomposed modes. The KMPMR method achieves nonlinear data classification in the high-dimensional space with the help of kernel functions mapping, and minimizes the maximum probability of the classifier which was misclassified. Based on the advantages of EWT and KMPMR, at the same time, in view of the effect of the selection of training sample on effectiveness of the predicted results, the Corrcoef function is used to obtain the training samples whose photovoltaic power output and change characteristics are parallel to the data of the forecast day, and then the photovoltaic power sequence is decomposed into different AM-FM components with different characteristics by using EWT. Finally, the different KMPMR model is used to forecast each AM-FM component according to their respective characteristics, and the predictive value of each component is superimposed to obtain the final prediction result. The experimental results show that the proposed method can improve the prediction accuracy, with the reduction of MAE (mean absolute error) and RMSE (root mean square error) of 56.19% and 55.19%, 54.15% and 53.36%, and 76.33% and 78.43% compared with the SVM method on sunny days, cloudy days and rainy days. Compared with the EMD-KMPMR method, the MAE and RMSE can be reduced by 9.42% and 9.59%, 38.74% and 33.96%, and 64.52% and 65.70% respectively. In the end, the experimental results show that the proposed method can obtain a higher prediction in 3 kinds of weather by using the actual operation data of photovoltaic power station in Aksu area. In addition to this, through comparing the results of the 3 experiments, the improved prediction accuracy proportion of EWT-KMPMR method in the experiment of cloudy and rainy days is larger than that of sunny days. Therefore, the EWT-KMPMR method has a good application value for the photovoltaic power output prediction under non-conventional weather, which can effectively reduce the influence of randomness on photovoltaic power for the power grid safety and reliable operation. © 2017, Editorial Department of the Transactions of the Chinese Society of Agricultural Engineering. All right reserved.