Research On Distributed Photovoltaic Power Prediction Method

In recent years,distributed photovoltaic,as a new way of utilizing solar energy,has developed rapidly.In power systems,the penetration rate of distributed PV has increased year by year,and the total installed capacity has increased rapidly.However,distributed photovoltaic systems are significantly random and intermittent due to factors such as weather changes,cloud drift,and the rotation of the Earth,which pose significant challenges to power systems.Accurate distributed photovoltaic power prediction can help reduce the uncertainty of photovoltaic output and provide assistance for safe and economic grid scheduling.At the same time,distributed photovoltaic has the characteristics of large quantity,small capacity,wide distribution,and generally does not have meteorological measurement devices,lacking data information.At the same time,some distributed photovoltaic power stations do not have power recording data.These factors greatly increase the difficulty of photovoltaic output prediction.In response to the above issues,this article conducted the following research:Firstly,the data situation of distributed photovoltaic is introduced in detail,and a method of reconstructing distributed photovoltaic power data based on spatial correlation is proposed.For four types of distributed photovoltaic data loss,neural networks,transfer learning,and other methods are used to conduct research on distributed photovoltaic data power reconstruction.For the first type of small amount of data loss power stations,a spatial mapping model is established using long short term memory(LSTM)neural networks to achieve power compensation;for the second type of large amounts of missing data or newly built power stations,a deep neural network model is established,and the transfer learning idea is used to achieve power reconstruction;a model migration method is proposed for the third type of power stations with no power data and similar power stations;for the fourth type of power stations with no power data and no similar power stations,an energy conversion algorithm is proposed.An empirical simulation is conducted with an actual power station,and the power reconstruction results are obtained and compared with other methods.Then,for distributed photovoltaic ultra-short-term power prediction,a distributed photovoltaic ultra-short-term power prediction method based on multi-location NWP and combination neural network is proposed.Obtaining numerical weather prediction(NWP)information from multiple nearby centralized photovoltaic power stations as input to a 1D convolutional neural network(1DCNN)to extract spatially correlated meteorological prediction information;in combination with 1DCNN’s data space reconstruction ability,information extraction ability,and LSTM’s deep mining ability for power series data,the fluctuation and change information of time series is extracted,and time series prediction is performed;The two prediction information is learned through a fully connected neural network pain to obtain the final prediction results,achieving distributed photovoltaic power ultra-short term prediction.The prediction is validated with the actual measured data of photovoltaic power stations,and compared with various methods.Finally,a distributed photovoltaic cluster short-term power prediction method based on multi-source prediction dynamic clustering was proposed.Firstly,the regional public weather forecast is statistically encoded,combining the public forecast code with NWP meteorological forecast data,at the same time,an improved autoencoder is designed to prevent data flooding and perform feature extraction.Then,a self organizing map(SOM)network is used to predict daily dynamic clustering and cluster division.The cluster is used as the prediction object,and the LSTM network is used for prediction,and the cluster power data is accumulated,Realize regional distributed photovoltaic short-term prediction,and conduct simulation experiments with actual power data,and compare it with various prediction methods.