Research On Power Data Prediction Method Based On Spatiotemporal Neural Networks

Accurate prediction of power data has important and far-reaching significance for improving the safety and stability of grid system operation.For multi-type power data,this paper studies the power data from low spatial dimension and high spatial dimension separately(the spatial dimension mentioned in this paper are attribute space dimension),and proposes specific prediction algorithms for power data with different characteristics.The proposed algorithms is mainly used to predict the input and output data of the energy system,which provide accurate data support for power generation and power planning of the power grid system,and play important roles in improving the reliability and stability of the power grid system.Firstly,for low-dimensional wind power data,this paper creatively proposes a bi-temporal neural network(BTNN)based on its temporal features,which can fully extracts local and global temporal features to obtain accurate wind power prediction.Considering that the original wind power data contains noise components and the trend of data is extremely abrupt in short terms,this paper uses the singular spectrum analysis algorithm to denoise and smooth the coarse data to obtain relatively clean data.Then,for power data with high spatial dimensions,in order to fully exploit its spatiotemporal features,this paper proposes a crossed two-dimensional convolutional network(C2D)to perform cross-convolution in both temporal and spatial dimensions.Among which,horizontal convolution is used to adaptively learn the importance of each attribute to the prediction results,and the vertical convolution achieves accurate prediction by extracting the temporal correlation in the time dimension.Meanwhile,in order to solve the problem of small amount of data in the target dataset,this paper introduces a migration learning mechanism based on the similarity measurement of time series dataset,which can effectively prevent over-fitting while improving the learning ability of the model.The validity and feasibility of the proposed methods are verified by sufficient comparison experiments on the wind power dataset CZWind Data,photovoltaic power dataset German Solar Farm and user power dataset Appliances Energy.