Study On Some Photovoltaic Ultra Short-term Output Power Probability Forecasting

With the increasing attention on energy shortage and environmental issues,solar energy is regarded as one of the most promising renewable clean energy sources and has been rapidly deployed for photovoltaic(PV)power generation.However,the nature of PV is chaos,with the increasing of the proportion of PV power generation in power system,it has become a big challenge to power system safety and reliable operation.To deal with these issues,the accurate and reliable ultra short term PV power forecasting becomes very important to optimize the operation cost and reduce uncertainties in power system.The shortages of forecasting based on traditional physical model method include low accuracy,poor versatility and complex parameter selection,and so on.As the representative of data-driven modeling technologies,machine learning methods make great progress in recent years,which have excellent nonlinear approximation ability and portability.The output power of PV power generation highly depends on the chaos weather conditions which made the PV power has strong uncertainties.The traditional deterministic forecasting methods are hard to meet the demand of the power system,because the error of forecasting is inevitable and the only limited information can be supplied to decision-makers.To solve these issues,our research works are focused on probabilistic prediction method.The main work of this paper can be summarized as follows:1)The extreme learning machine(ELM)has excellent generalization performance and high computational efficiency.The ELM based lower–upper bound estimation method(ELUBE)is proposed to construct the prediction intervals for ultra short term PV power forecasting.Two indicators are introduced to evaluate the PIs performance and the comprehensive evaluation function for PIs are developed.To obtain the higher quality prediction intervals(PIs),the improved particle swarm optimization(PSO)method is used to optimize ELM parameters.The effects of several commonly used activated functions and the number of hidden nodes for ELM regression performance are investigated.Experiments are organized based on measured PV data from rooftop DC Micro-grid.the effectiveness of PIs approach for ultra-short term PV power forecasting is verified through the experiments.2)The performance of the proposed ELUBE method drops in the presence of large process as well as strong data uncertainties and several parameters local minima.An ensemble ELUBE approach is proposed for PV ultra-short term power generation forecasting.To obtain the diversity and high quality of ensemble members,the ELUBE models are given three different activation functions and different hidden nodes,respectively and eachELUBE is applied to develop many PIs.A group of the best models is selected for combining PIs.To optimizing the weighted for each sub PIs model,the improved DE is used based objective function of optimal combined PIs.The effectiveness of the proposed combined PIs approach for ultra-short term PV power forecasting is verified through the experiments.3.To further improve the PIs quality,and reduce the impact of PV data the uncertainties on the regression model,the nonlinear relationship between PV power output and meteorological data is investigated based on cosine nonlinear correlation measure results,and meteorological parameters with more nonlinear correlation are select the input of K-Means classification to classify the data samples.avoid the difficulty of the parameter selection,when the multi-objective problem converts to single-objective problem in chapter two and chapter three,a novel PIs method is proposed for PV ultra-short term power generation forecasting by using Pareto Optimality and Non-dominated Sorting Genetic Algorithm II(NSGA-II)is applied to optimal the ELM parameters.the effectiveness of the proposed methods is verified by using Measured PV data.4.We study the nature of uncertainties when the machine learning methods are applied for ultra short term PV power forecasting.A new method is proposed,which involves the model uncertainties and noise uncertainties,and PIs are constructed with a two-step formulation.In the first step,the variance of model uncertainties is obtained by using extreme learning machine to make deterministic forecasts of PV power.In the second stage,three approach are developed to quantify the uncertainties of data noise.The performance of the proposed approach is examined by using the PV power and meteorological data measured from rooftop DC micro-grid system.The validity of the proposed method is verified by comparing the experimental analysis with other benchmarking methods,and the results exhibit a superior performance.