A novel convolutional neural network framework based solar irradiance prediction method

摘要

As an important part of solar power system, photovoltaic grid-connected system and solar thermal system, solar irradiance has the inherent characteristics of variability and uncertainty. Existing data analysis methods are difficult to demonstrate better generalization. Hence, resource planners must be adaptable to accommodate these uncertainties while conducting planning. To improve the accuracy of solar energy prediction and efficiently organize the utilization of solar energy, a novel convolutional neural networks framework has been constructed. Firstly, we have established a convolutional neural network framework for solar prediction based on meteorological data from surrounding sites and different sampling times. Secondly, the chaotic GA/PSO1 hybrid algorithm is applied to optimize the hyper parameters of the novel framework, which alleviates the imperfect performance caused by improper hyper parameters. At the meantime, the hybrid algorithm can reduce the manpower and resources of manual parameter adjustment. Excitability of the novel framework has been verified by benchmark tests. In the solar irradiance prediction studies, the annual average Mean Absolute Error of the proposed method is reduced by 0.1463 MJ.m(-2) compared with single CNN2 framework. The annual average Mean Absolute Error of the proposed method is reduced by 49.47%, 47.6%, 20.34%, respectively, compared with ANN(3), K-means-RBF4 and GBRT.(5) The superiority has been fully illustrated through all the simulation test results. Therefore, the proposed method provides a basis for accurate estimation of solar power, which can promote further development of the whole power system.