The sunlight meter based on the photoelectric principle will be soon popularized and applied in the whole country. The inversion of solar radiation data based on the observation data of light intensity can effectively make up for the insufficient number of solar radiation observation stations. Aiming at the shortcomings of the existing solar radiation estimation methods, a composite model combining with the Principal Component Analysis (PCA), the Mind Evolutionary Algorithm (MEA) and the BP neural network is proposed, using the observation minutes data of solar intensity, solar altitude angle, temperature, and humidity to invert the solar irradiance. Based on the clearness index and the probabilistic neural network (PNN) classification method, the weather types are divided into three categories: sunny, cloudy, and overcast. The classification accuracy is 966948%. Then using the four influencing factors after PCA dimensionality reduction, the solar irradiance is obtained by the BP, GABP and MEABP methods for the three types of weather, and compared with the measured data of the standard radiation meters. The results show that the determination coefficient of the MEABP model is up to 09958 in sunny, cloudy and overcast weather. Compared with the single BP model, RMSE decreases by 49%, 3245% and 1064%, respectively. Compared with the GABP Model error, MAPE decreases by 4254%. The generalization capability of the MEABP composite model proposed in this paper has been effectively improved.