Application Of Kernel Adaptive Filtering In Photovoltaic Power Forecasting

The traditional way of generating electricity from fossil fuels has caused serious problems such as environmental pollution and waste of resources.Increasing the share of photovoltaic power generation in the structure of the electricity industry is one of the effective ways to solve such problems.However,due to the inherent randomness and intermittent nature of solar energy,photovoltaic power generation can have an impact on the operational homeostasis of the grid-connected PV system,making the scheduling of the grid more difficult.The use of photovoltaic power forecasting technology to make more accurate predictions of the PV power level in the future can provide data support and guidance reference for the subsequent operation and dispatch of the grid-connected PV system to ensure the stable operation of the power system.Kernel Adaptive Filtering(KAF)is a class of kernel learning methods with online learning characteristics,which not only allows fast online processing of large sample sizes but also real-time adaptive adjustment of modelling parameters.In view of the successful application of kernel learning methods such as Support Vector Machines(SVM)and Kernel Extreme Learning Machines(KELM)in photovoltaic power forecasting modelling,the KAF method is extended to the field of PV power prediction in order to further improve the real-time and adaptive nature of PV power prediction models.The contents of all the research covered in this paper are shown below:(1)The PV power prediction technology is investigated.Mainly covers the development status of photovoltaic power forecasting techniques in recent years and the main photovoltaic power forecasting modelling approaches.(2)Basic KAF algorithms are studied,including the kernel recursive least squares algorithm(KRLS),the extended kernel recursive least squares algorithm(Ex-KRLS),the ALD-KRLS algorithm incorporating the approximate linear correlation(ALD)criterion,and the FB-KRLS algorithm incorporating the fixed budget(FB)criterion.The basic KAF algorithms are applied to the Mackey-Glass(MG)time series forecasting and short-term solar irradiance forecasting modelling,and compared with the SVM model and KELM model.The experimental results show that the basic KAF models are significantly better than the SVM model and KELM model,with the root mean square error index(RMSE)decreasing by an average of 26.91% and 13.63% in the two examples,respectively.(3)The basic KAF algorithms are improved by proposing a class of extended kernel recursive least squares(NKF-KRLS)algorithms based on nonlinear Kalman filter algorithms,including Ex-KRLS algorithm based on extended Kalman filter algorithm(EKF-KRLS),Ex-KRLS algorithm based on unscented Kalman filter algorithm(UKF-KRLS),and Ex-KRLS algorithm based on square-root cubature Kalman filter algorithm(SCKF-KRLS),and the proposed algorithms are optimised based on ALD,FB sparsification criterion respectively.The NKF-KRLS algorithms are applied to MG time series forecasting and short-term solar irradiance forecasting modelling to compare with the basic KAF forecasting models.The experimental results show that the NKF-KRLS models can forecast significantly better than the basic KAF prediction models.The average decrease in RMSE in the two examples is 36.94% and 28.58%,respectively.The SCKF-FBKRLS model has the best forecasting effect in both examples.(4)NKF-KRLS algorithms are applied to several PV power prediction models and compared with the basic KAF prediction model.Experimental results show that the NKF-KRLS model outperforms the basic KAF model for photovoltaic power forecasting The average decrease in RMSE for the NKF-KRLS models is 17.01% in spring,39.10% in summer,26.55% in autumn and 7.94% in winter for the different seasons of forecasting modelling respectively.The average decrease in RMSE is 24.25%,23.39% and 23.34% for the three forecasting scales of 5min ahead,10 min ahead and 15 min ahead respectively.The average decrease in RMSE is 19.70% when climate variable inputs were taken into account.Of all the NKF-KRLS models,the forecasting model based on the SCKF-FBKRLS algorithm has the best forecasting results in all photovoltaic power forecasting examples.