Research On Multiple Temporal Scale Solar Photovoltaic Power Forecasting Method

As an important method for solar energy utilization,photovoltaic(PV)power generation developed rapidly due to the advantage of no fuel consumption,no pollutant emission,flexible application form,unlimited capacity scale,safety and reliability,simple maintenance,and so on.However,the intermittency and stochastic volatility of PV power output bring a significant challenge to the balance of power generation,transmission and consumption,which restrict the utilization of PV power.PV power forecasting is one of the economical and effective ways to solve the above problem.The manifestation of periodicity and volatility of PV power varies with the temporal scale.Meanwhile,the scheduling of power grid also includes multiple time range,which needs PV power forecasting support at different temporal scale.Therefore,to promote the PV power consumption,ensuring the safe and stable operation of the grid,the research on multiple temporal scale solar photovoltaic power forecasting method is essential.Short-term PV power forecasting should provide the PV power data in the following 1?3 days.The weather condition will influence the PV power output characteristic.Therefore,the weather status pattern recognition model is needed to forecast the PV power output according to weather status.In this paper,the influence of sample data and classifier characteristic on the performance of weather status pattern recognition model is studied.Then based on the weather status classification and recognition method,a day ahead weather status prediction based short-term PV power forecasting method is proposed.Multiple solar irradiance forecasting models and a voting process are utilized to predict weather status,then a Long Short-Term Memory and Recurrent Neural Network based PV power forecasting model are used to forecast the PV power.The forecasting results are modified according to time correlation model.Ultra-short-term PV power forecasting should provide PV power data in the following 4 hours with a higher demand for forecasting ability of PV power fluctuation in cloudy days.Facing the above demand,multiple algorithms are applied to realize the multiple parallel forecasting of PV power,and then establish the data fusion model to fuse the multiple forecasting results into one final result.Considering the randomness of PV power,the performance of different models in the multiple parallel forecasting may vary with the time,and a single fusion model is unable to adapt to these different circumstances.Therefore,an ensemble model for ultra-short-term PV power forecasting consists of multiple parallel forecasting models,fusion pattern recognition model,data fusion models is proposed.An adaptive time-section fusion technique is developed to further improve the accuracy of the ensemble model.The surface solar irradiance may show great and rapid fluctuation in minute level temporal scale due to cloud motion on cloudy days.The corresponding minute level power fluctuation can impact on the safe and stable operation of the power grid.However,traditional weather and power data-driven based machine learning and time series models can not forecast this kind of minute level power fluctuation.Therefore,the ground-based sky image based minute level temporal scale PV power forecasting method is proposed.Firstly,the direct observation of cloud is achieved based on ground-based observation equipment.Secondly,the Otsu and k-means clustering method are applied to recognize the cloud pixels in sky image.Thirdly,a phase correlation based cloud displacement vector calculation method is proposed,then the proposed method is improved according to image phase invariance characteristic.Finally,a surface irradiance mapping model based on sky image is established.