Prediction Model Of Solar Irradiance Based On Variational Mode Decomposition And Relevance Vector Machine

At present,the global photovoltaic power industry is developing rapidly,and the annual installed capacity and cumulative installed capacity are increasing with high-speed,and the share of photoelectric in the power grid continue to increase.As the solar energy has a typical fluctuation and intermittent characteristics,resulting in the output of photovoltaic power generation is not stable,which has a great impact on grid connected photovoltaic power generation and power grid safe and stable operation,seriously hindered the large-scale photovoltaic grid.The accurate prediction of PV output power is beneficial for the power system department to schedule and plan ahead of schedule,on the premise of ensuring the safe and stable operation of power grid,connecting as much as possible photovoltaic grid.Irradiance is the most direct and significant factor affecting the output of photovoltaic power,therefore,accurate prediction of solar irradiance will improve the prediction accuracy of the output power of photovoltaic power generation system,so accurate irradiance prediction has important application value.The solar irradiance was affected by climate,weather and other factors,which had the characteristics of randomness and instability,so the prediction accuracy of the traditional solar irradiance point forecasting models was often not high.At the same time,the results of point forecast can not characterize the randomness of solar irradiance,and brought risk to decision-making.If the deterministic forecasting result of solar irradiance was given,at the same time the interval forecasting results of solar irradiance fluctuation was depicted,which would be beneficial to the risk analysis and more reasonable dispatch of the power sector.Therefore,based on the in-depth study of the optimal variational modal decomposition algorithm,chaotic grasshopper optimization algorithm,locust mixed kernel relevance vector machine,non parametric kernel density estimation method,the solar irradiance point prediction model and interval prediction model were studied,the main research work and innovation results were as follows:(1)Parameter optimization in solar irradiance prediction model directly affects the performance of the model,aiming at the parameter optimization problem,the chaotic grasshopper optimization algorithm(CGOA)was proposed based on the grasshopper optimization algorithm(GOA),which was a bionic intelligent algorithm to simulate thebehavior of grasshoppers in the foraging process.The biggest feature of GOA was that the next position of each particle is determined not only according to its current position and the current optimal position,but also according to all other particles in the iterative process.Traditional bionic intelligent algorithms often do not consider the location of other particles,so GOA was easy to jump out of local optimum and converge faster.The paper made three improvements on the basis of GOA : 1)The linear decreasing function was changed into the nonlinear inverse incomplete gamma function,which was beneficial for GOA to focus on global exploration in the early stage,and to focus on more detailed local development in the later period;2)Updating the cross-boundary particles to the endpoint of the position variable interval was instead by the inside of the interval,which was to prevent GOA from frequent optimization in the upper or lower boundaries and to influence the speed of convergence;3)The chaotic search algorithm was used to optimize the current optimal position to improve the accuracy of the solution of GOA.Through the test function,it was found that CGOA had faster convergence rate and higher precision accuracy than GOA.(2)In view of the instability of solar irradiance,a variational modal decomposition(VMD)was proposed to decompose the original solar irradiance sequence to obtain relatively stable modal components.VMD algorithm had three parameters: modal number K,penalty factor a and fidelity coefficient t,which were difficult to determine,so the optimal variational modal decomposition algorithm(OVMD)was proposed.For the modal number K,the correlation between the residual component and the original sequence after VMD decomposition was proposed to guarantee the accuracy of the decomposition and to determine the optimal value of K.For a and t,OVMD established the residual index according to the residual after VMD,which was as the fitness function of CGOA,and then searching for the optimal a andt through CGOA.The actual solar irradiance sequences were decomposed by the widely used ensemble empirical mode decomposition(EEMD)and OVMD,which were compared and analyzed by index of orthogonality.It was concluded that OVMD had better decomposition performance than EEMD.(3)The relevance vector machine(RVM)was used to established the prediction model for each OVMD component,which was optimized by CGOA,the prediction results of each component model was accumulated,and the point prediction model of solar irradiance(OVMD-CGOA-RVM)was obtained.The selection of kernel function in RVM and the setting of parameters in kernel function directly affected the prediction performance ofRVM.For kernel function,the hybrid kernel function was obtained by combining the Gauss kernel function with strong learning ability and the polynomial kernel function with strong extension ability,so the learning ability and extension ability were guaranteed;the parameters in the hybrid kernel function were optimized by CGOA,in order to establish the optimal RVM prediction model.The monitoring data of two monitoring platforms in different times and regions were divided into four sets of data according to the four seasons of spring,summer,autumn and winter,the hourly irradiance of the last two days in each set of data were predicted by using OVMD-CGOA-RVM model,CGOA-BP mode,ARIMA mode,CGOA-LSSVM,CGOA-RVM model and EEMD-CGOA-RVM model respectively.The prediction results were evaluated by the error evaluation index,which showing that the OVMD-CGOA-RVM model was better than the other five models in all the four seasons,and the prediction accuracy is higher,which had better prediction accuracy,and was more stable.(4)On the basis of the VMD-CGOA-RVM point prediction model,the nonparametric kernel density estimation(KDE)method was adopted to calculate the probability density of the solar irradiance prediction error,and then three times spline interpolation was used to fit the probability distribution curve of prediction error,so as to the prediction interval model of solar irradiance was obtained,which was called OVMD-CGOA-RVM-KDE.Aiming at the problem that the kernel width of the kernel function in KDE was difficult to determine,a comprehensive evaluation function constructed with average coverage and average width was proposed as the fitness function,and CGOA was used to achieve the optimal window width.Through the experiment of monitoring data from the above two monitoring platforms,it was found that the interval prediction model of OVMD-CGOA-RVM-KDE was more reliable and accurate than CGOA-RVM and CGOA-BP-KDE.The prediction interval of solar irradiance based on OVMD-CGOA-RVM-KDE model had a small interval width,which was helpful for power system dispatching department to carry out power grid planning,risk analysis and reliability evaluation.