Probabilistic Power Forecasting For Distributed Photovoltaic Based On Bayesian Neural Network And Its Application

Recently,distributed photovoltaic generation with its clean and flexible features has been rapidly deployed.However,the power of photovoltaic is greatly affected by the sunlight,which shows obvious intermittence and randomness to bring great challenges to the security of the power system.The traditional deterministic forecasting methods are difficult to forecast the short-term fluctuation of photovoltaic power,and can not provide reliable data for the dispatching system.In this thesis,the data-driven artificial intelligence technology is used as a starting point to study the probability forecasting model of photovoltaic output based on machine learning and its probability scenario generation method,and it is applied to the optimal dispatching of power grid with high proportion of distributed photovoltaic to improve the economy and accuracy of the dispatching strategy.Firstly,aiming at the problem that the traditional deterministic forecasting model is difficult to forecast the minute-level short-term fluctuation of photovoltaic,this thesis proposes a photovoltaic probabilistic forecasting method based on improved Bayesian neural network.At the input of the neural network,the full connected neural network and the one-dimensional convolutional neural network are introduced to deal with the weather data and historical data respectively.On this basis,an improved Bayesian neural network photovoltaic power prediction model is constructed.Then,the t-distribution neighborhood embedding algorithm is used to reduce the dimension of the input data in order to reduce the training complexity of the model and improve the sensitivity of data.The example shows that the proposed algorithm can effectively improve the forecasting accuracy of photovoltaic probabilistic forecasting model for short-term fluctuation of photovoltaic power,and the uncertainty forecasting ability of the model.Secondly,aiming at the problems of small sample data and poor model generalization of new photovoltaic units,this thesis proposes a photovoltaic forecasting method based on transfer learning to reduces the dependence of historical data on the model.Other photovoltaic units with sufficient data are trained offline to learn the nonlinear relationship between environmental characteristics and photovoltaic output.Then the models are transferred to the forecasting task of the new unit,and the online update with a low learning rate is carried out according to the accumulated new data of the unit so that the performance of the model is continuously optimize.The weight optimization method is used to integrate multiple migration models to further improve the prediction accuracy of the model.Furthermore,aiming at the problem that there is a big difference between the generated scenario distribution and the real photovoltaic output distribution in the traditional scenario generation algorithm,this thesis proposes a photovoltaic scenario generation method based on the improved generative adversarial network.The rolling forecasting of Bayesian neural network is used to replace the traditional generator of generative adversarial network so as to break through the limitation of the preset probability distribution of the scenario,and make the statistical characteristics of generated scenario closer to the characteristics of real photovoltaic scenario,and accurately give the probability of different scenarios according to the input weather conditions,so as to provide data support for the optimal dispatching.Finally,aiming at the dispatching of distribution network with high proportion of distributed photovoltaic,a stochastic optimal dispatching model and solution method based on scenario method are proposed.The scenario method is used to transform the optimization problem whose objective function is the expected value into a two-level deterministic optimization problem which is easier to be solved.Aiming at the problem of too much computation in optimal dispatch caused by scenario method,on the one hand,a scenario reduction algorithm based on K-means is used to extract typical scenarios;on the other hand,the particle swarm optimization algorithm is used to solve the optimization problem quickly.Simulation results show that,when the output of photovoltaic unit changes suddenly,the dispatching cost of this method is lower than traditional method which uses deterministic forecasting result.