Ultra Short Term Power Forecast Of Photovoltic System Based On Transfer Learning

For decades,global warming and the severe depletion of fossil fuels have encouraged the use and development of renewable energy sources(RES).Renewable energy sources are not only considered as novel solutions to global warming and severe depletion of fossil fuels,but also reflect the future of energy development.In terms of replacing conventional energy,solar energy has become one of the most popular methods.Solar energy has been implemented in many countries in the world.But solar energy is naturally fluctuating and intermittent,and this characteristic is mainly affected by the geographical location of the power plant and weather conditions.The intermittent nature of solar energy resources can cause serious problems with the balance between power generation and load demand.Therefore,the implementation of large-scale grid-connected solar photovoltaic power plants has brought major problems to the power grid,such as system stability,reliability,electric power balance,reactive power compensation,and frequency response.Solar photovoltaic power generation forecasting has become the most common and effective method to solve these problems.Among the many photovoltaic power prediction methods,the method based on deep learning networks stands out.However,in order to obtain the ideal photovoltaic prediction model,sufficient historical data is a prerequisite.For new power plants,the existing data is not enough to train the model.Therefore,this paper makes the following researches on ultra short term photovoltaic power prediction under data shortage:First,by studying the working principles of neural networks and deep learning,we set out to build multi-layer perceptron neural networks,recurrent neural networks,and long short-term memory neural networks that are widely used in the field of time series prediction.Using the prepared data set for training and testing,ultra short term photovoltaic prediction models based on three neural networks were obtained,and the prediction accuracies of three photovoltaic power prediction models were compared.The result shows that the long short-term memory neural network has the best performance.Secondly,this paper studies a transfer learning method applied in the field of photovoltaic power prediction,that is,transforming the constructed multilayer perceptron neural network,specifically adding an input layer and an adaptive layer with metrics,and using the prepared training set containing the source domain data and the target domain data to jointly train the net,which will be tested on the target domain test set,under semi-supervised way.Finally,an ultra short term photovoltaic power prediction model that can be applied to the target domain is obtained.The result shows that the proposed transfer learning method greatly improves the prediction accuracy of the multilayer perceptron on the target domain.Finally,the recurrent neural network and the long short-term memory neural networks were transformed and trained according to the proposed method,and two other ultra short-term photovoltaic power prediction models based on transfer learning were obtained,and their prediction accuracies were compared.Then,using the most accurate LSTM-based transfer model,it was explored that the selection of the source domain,the window length used in making the training set and the amount of the target domain’s data will affect the transfer accuracy.The experimental results show that the performance is better when the target and source domain power plants are closer in climatic conditions and installed capacity;Besides,when the input length is controlled at 4-7,the effect is best;What’s more,within a certain limit of the amount of data in the target domain,the transfer effect improves as the amount of data in the target domain increases.