Short-Term Solar Power Forecasting Based On Fuzzy C-Means And Gaussian Process Regression

Photovoltaic(PV)power has become an important renewable energy generation after wind power generation.However,due to the strong volatility,randomness and intermittent nature of PV power,the integration efficiency of PV power into electricity grid is limited due to the reason of safety and stability.Therefore,PV power forecasting is an important measure to solve this problem.Because it not only can provide a reliable reference for grid power generation planning and optimization of dispatching decisions,but also provide data support for coordinated control of multi-energy complementary systems.It is a key technology to improve the integration efficiency of PV power into electricity grid.Based on the existing prediction algorithms,this paper proposes a prediction model based on Feature Weighted Fuzzy C-Means(FWFCM)and Weighted Gaussian process regression(WGPR)optimized by crisscross optimization algorithm(CSO).The main works are as follows:PV power is closely related to the weather type,so the FWFCM is used to classify the weather type in the stage of data preprocessing,and then the historical data of the same weather type as the forecast day is selected as the model training samples.In particular,due to complex weather effects and measurement errors,it is unavoidable outliers in the measured meteorological data.In order to reduce the effects of outliers on the prediction results,an innovative method employing the WGPR approach is proposed in this paper,such that data samples with higher degree of abnormality will be set to a lower weight by the Isolation Forest algorithm(iForest),thus helping the model to better fit the mapping between input variables and PV power.In the model optimization stage,CSO algorithm is used to optimize the hyper-parameter selection process of WGPR.Finally,Customer can get the prediction results and the confidence interval of the PV power by input the test set into the model proposed in this paper.In the part of experimental Analysis in this paper,the data collected at the Alice Springs PV power station in Australia are used.To compare the forecasting performance between CSO-WGPR model and other traditional prediction models,we select the PV power data in the weather type of sunny,cloudy and rainy.The simulation results show that:(1)FWFCM can divide the cluster center accurately according to the daily features,it is effectively to obtain the historical data of the same weather type as the forecast day as the model training samples.(2)The WGPR model proposed in this paper can reduce the influence of outliers in complicated weather,and has better prediction performance than the traditional GPR model.(3)The CSO algorithm can be effectively applied to the hyper-parameter selection process of GPR-based model,and the optimization effect is better than the conjugate gradient descent method in the traditional GPR model.