Research On Very-short-term Online Photovoltaic Power Forecasting Model And Algorithm Based On Dynamic Neural Networks

PV power generation has been widely used due to its cleanness,safety and abundant reserves.However,as the proportion of photovoltaic(PV)connected to the power grid continues to increase,its randomness and intermittence have a more severe impact on the balance between supply and demand of the power and the stability of the power grid,which also directly affects the power quality and the effective consumption of photovoltaic energy.By predicting the output of photovoltaic power,the power supply can be scheduled in advance to ensure the power balance between the supply side and the load side,and likewise diminish the impact of volatility on the power system.At present,neural networks are still the most commonly utilized prediction methods.The conventional neural network forecasting model is completely dependent on offline data.After the offline training,no further structural adjustments will be made during the forecasting process,which does not cater for the dynamic PV power supply.Focus on the online photovoltaic power forecasting model and algorithm based on dynamic neural networks,this thesis designs a two-stage PV power prediction model based on the resource allocation network(RAN)to achieve online learning of the model.Considering the noise in the data,a secondary dynamic adjustment strategy is further proposed to improve the accuracy of data feature selection during the online learning phase.Finally,a parallel multi-subnet prediction model is proposed to improve the efficiency and accuracy of the prediction process.The main research contents include:(1)A very short-term online forecasting model for photovoltaic power based on two-stage resource allocation networkThe offline learning method which traditional neural network mainly utilized relies solely on the historical data,and the un-modeled pattern encountered in the practical application cannot be learned by the forecasting model.A photovoltaic power prediction method based on two-stage resource allocation network is proposed in the thesis.The initial prediction model is trained offline according to the RAN learning rules in the first stage,and the prediction data is further screened and learned online in the second stage.The proposed forecasting model realizes the online adjustment of the un-modeled pattern and improves the adaptability to the dynamic variation of photovoltaic power.(2)A PV power prediction algorithm based on RAN with secondary dynamic adjustmentThe working conditions of a photovoltaic system are complicated,and the power output is dynamically changed under the interference of various external factors.Therefore,various abnormal noisy data is inevitably contained in the online data.The noisy data can influence the accuracy of online adjustment,leading to the learning of occasional abnormal data pattern.Aiming at the influence of abnormal noisy data on the online learning process,a secondary dynamic adjustment strategy is developed,and a RAN photovoltaic power prediction model based on the secondary dynamic adjustment strategy(RAN?SA)is proposed in the thesis.By constructing a data feature buffer and redesigning the judgment conditions for online learning,the accuracy of filtering data is improved,which makes the typical un-modeled patterns be learned accurately.The proposed secondary adjustment strategy also improves the adaptability of the prediction model to dynamically changing data.(3)A parallel multi-subnet prediction model for photovoltaic power and its comprehensive algorithmWhen a single RAN?SA model is applied to a scenario with a large data size or a long time range of training data,a great deal of nodes will be allocated.However,a single large-scaled prediction model can bring heavy calculation burden,resulting in low efficiency and poor prediction results.The multi-subnet model structure can effectively reduce the calculation cost of the prediction and improve the operating efficiency.Meanwhile,only a single sub-model structure will be changed during the subsequent adjustments,which diminishes the impact of online adjustment on the overall accuracy.A multi-subnet prediction model with parallel structure is proposed in the thesis.By dividing the data set with the fuzzy c-means clustering algorithm,the amount of data in the subset is effectively reduced,while the data features in each cluster are more concentrated,which makes the training and learning more accurate.The parallel multi-subnet structure developed in this thesis not only improves the overall efficiency of the prediction,but also further improves the prediction accuracy in actual PV applications.The simulation results of actual photovoltaic data also verify the effectiveness of the designed method.