Distributed photovoltaic short-term power output forecasting based on extreme learning machine with kernel

Abstract: Along with the rapid development of rural power network in our country, the power demand increases and creates the environment for distributed energy access to the agricultural power network. Taking integration application program with the distributed photovoltaic power output systems through the rural power network and other application forms is a realization and effective method for the new clean energy can be absorbed at local, reducing carbon emissions and environmental pollution. In our country, the national government had developed and introduced relevant policies, and making plans toward to the rapid and healthy development of distributed photovoltaic power output systems for a period of time in the near future. In addition to the national planning, the State Grid Corporation of China also proposed their policies about macro planning and access technologies, that making convenience to the legal proceedings for the personal property distributed photovoltaic access, and offering the related technical support for the distributed photovoltaic power output system access. In order to make sure the operation stability and economy of micro-grid and distribution network with the distributed photovoltaic power output systems access at rural power network, the distributed photovoltaic power output forecasting technology need to be deeply researched combining with the forecasting system application environment characteristics of the rural power network. For the distributed photovoltaic power output system at user side, a short-term PV power output forecasting method based on the algorithm that extreme learning machine with kernel (ELM_k) is proposed. This method considers the operation cost constraints of short-term power output forecasting system, such as does not depending on the high cost of numerical weather prediction. This forecasting system also considers the application characteristics of rural power network, such as the lack of professional maintenance for the distributed photovoltaic panels and related inverter equipment. For the distributed photovoltaic systems with different capacities, the PV power output short-term forecasting model was built based on ELM_k algorithm. Taking the training samples filtration based on attributes weight to improve the computational efficiency of the PV power output forecasting model. The parameters of the forecasting model that relevant with ELM_k algorithm and samples filtration method is optimized through the particle swarm optimization algorithm. The forecasting model uses the low-cost samples with non-numerical weather prediction. For the distributed photovoltaic power output systems at tens of kilowatt, the mean absolute relative error was only 16-18%, and can complete a single power output forecasting within 10 milliseconds, when it implemented on the lower power processor, furthermore, this forecasting method can basically maintain the original accuracy when the low weight attributes were simplified. At the same time, under the distribution photovoltaic operation circumstance of random dust overlying and inverter partial failure, the prediction error remains largely unchanged, which proves the high adaptability of this forecasting model.