Study Of Generalization Ability Of Solar Irradiance Forecast On The Basis Of Neural Network

Energy conservation of air conditioning systems has become a very serious problem, as the ratio of architectural energy consumption to the total energy consumption increases rapidly, especially, since the status of "energy crisis" was realized by people decade ago. One of the key measures in energy conservation of an air conditioning system is accurately to determine the load of the system, and accurate load forecast is the base of optimal operation of the system. A solar irradiance forecasting model with good generalization is a key precondition of the accurate forecast. This paper aims to a deep research into the theories and methods of improving the generalization ability of solar irradiance forecast on the basis of neural network.Solar irradiance is of non-linear characteristics. Wavelet neural networks are capable of handling non-linear problems. Because of the neural networks' capability of non-linear function approximation and self-study, self-adapt, it becomes the basis of solar irradiance forecasting.The intelligence algorithm enlightened by biological system is more and more emphasized by people. The immune algorithm, neural network, genetic algorithm is named the "Three Big Bionic Algorithm". Nowadays, the combination of artificial neural networks with genetic algorithm and immune system can make use of the general search ability of genetic algorithm and immune algorithm, and utilize wavelet neural networks' capability of parallel handling ability in large scale, non-linear function approximation and self-study, self-adapt as well. All contribute to a great improvement of generalization ability of solar irradiance forecast on the basis of wavelet neural network.Due to the relatively poor generalization ability of most existed solar irradiance forecast models on the basis of neural network, and the generalization ability is influenced by many factors, this paper has done some research on the aspects of network architecture, samples, training precision, network initialization. It has been demonstrated that the generalization ability of neural network depends on the suitable size of the training dataset, i.e., a well selected training dataset can not only guarantee the forecast precision, but also reduce the duration time of training. Meanwhile, only when the structure complexity of network matches the size of training dataset can we ensure the expected generalization ability. Network training is divided into three parts. First, the generalization ability is weakened as the duration time increases. Second, the generalization ability is then weakened after it is strengthened first when duration time increasing. it is in a period of dynamic changing. At last, the generalization ability is strengthened again as the duration time increases. For the given network structure, there is a lowest point of forecasting error before the lowest point of training error, i.e., there is a optimization time for training stop, which gives a guide for qualitative analysis of training network.According to the analysis above, the achievements of this paper are a complete set of models of forecasting solar irradiance, including daily total solar irradiance forecast model, daily diffuse solar irradiance forecast model. This paper uses the meteorological data in Macao from 1991 to 2000 as the investigation object of the forecast models. The performance parameters, relative errors and absolute errors of tracing forecast and emulating forecast were analyzed according to the example. In order to further validate the improvement of the generalization ability of solar irradiance forecast on the basis of wavelet neural network, comparisons have been completed between the cases with and without initialization of various intelligent methods. The results demonstrate that the root mean square errors, mean absolute errors, mean relative errors are improved greatly, and the precision of the neural network after intelligent initialization is more accurate, the generalization ability is improved.Besides, this paper has done many experiments to improve the generalization ability of the wavelet neural network from the aspects of network structure, samples, and training precision. The summarized principles and corresponding conclusions can not only guide the improvement of the generalization ability of solar irradiance forecast on the basis of wavelet neural network, but also can guide the training neural network applied in other fields.