Solar Irradiance Prediction Of PV Power Station Based On Artificial Neural Network And Time-Periodicity

Photovoltaic power generation plays a prominent role in reducing environmental pollution, improving energy structure and easing the energy crisis, which has become an important direction for renewable energy development in the world. Solar radiation is the main factor that affects the PV power, PV power renders volatility and intermittent due to its own randomness and cyclical changes, which will bring great challenges on power system dispatching management and reliable operation of the PV power plant. In short, solar irradiance forecast is of great significance to photovoltaic power prediction and the planning, operation and control of photovoltaic power.At home and abroad the short-term prediction study around solar irradiance includes regression analysis, time series and neural network, compared to other methods, neural network forecasting model has better predict results. This article makes researchs on how to improve the capability of the irradiance neural network prediction model and three aspects, which respectly are forecasting performance in the input variable selection, structure optimization, the revised study of the forecasting model output. First, it makes analysis on the input and output variables of the existing irradiance neural network prediction model, the input variables of the existing model mostly are irradiance historical data sequences, on the categories, it does not directly take other factors related to the irradiance into account, and in number, most are24dimensions and even more, which may have the prediction accuracy and generalization ability reduced because of too much redundant information, for this problem, introduce the factors as the input of the predictive model,which are directly related to irradiation, such as environment temperature, wind speed, pressure and volume, and have the multidimensional history irradiance sequences simplified to the mean and variance of the irradiance, remove the redundant information, comparing different combinations of input variables of prediction neural network model, to determine the reasonable combination of input variables. Second, after the model input variables fixed, the structure of neural network model becomes a key factor in determining its forecast performance, to this end, using cross-validation method for optimizations of the model structure and parameters, according to the error index, determining the number of hidden layers, the number of hidden neurons, transfer functions, training functions and so on in the neural network model, then, it has completed its network structure optimization. Once again, after the model input variables and network infrastructure fixed, considering the periodicity of irradiance, use the historical irradiance data by means of a weighted summation to correct the output variables of the short-term forcasting model, which has determined the forecast model output and their reasonable weight of the revised data, and given the type coefficients of the revised data according to weather types, to differentiate the differences in terms of time periodic similarity among different types of weather, eventual to get the result which is the integration of neural network prediction model outputs and time cyclical correction. Finally, using the measured data of a PV power plant, to train the neural network model, through the comparative analysis of prediction results among a variety of neural network models, this article has verified the validity and rationality of this method. Measured data forecast results showed that the short-term forecasting model based on ANN and time periodicity has a high accuracy, which can be used for short-term forecasting of solar irradiance, laying a foundation for accurate prediction of photovoltaic power.