Research And Application Of Spatiotemporal Sequence Prediction Algorithm Based On Three-dimensional Convolution And Recurrent Neural Networ

The spatiotemporal series prediction aims to predict the spatial changes in the region in the future based on the spatiotemporal series data of historical observation.Accurate spatiotemporal series prediction has great application value in the fields of intelligent transportation,short-term and imminent prediction,wind power generation,etc.With the increase of prediction step size,the prediction model based on recurrent neural network will lose the long-range dependence,resulting in the inaccurate prediction results of long series.In order to solve the above problems,this paper proposes a spatiotemporal series prediction model H-STNet(Hybrid-Spatiotemporal Network)integrating 3D Convolution Neural Network(3D CNN)and cyclic neural network,in which 3D CNN and cyclic neural network are responsible for capturing global and local spatiotemporal features respectively.The main research contents are as follows:(1)In view of the problem that traditional 3D CNN loses time information with the deepening of network layers,this paper designs a dual-branch 3D CNN that can simultaneously extract features from the time dimension and channel dimension of data.In order to extract the key spatiotemporal information,this paper also designed a spatiotemporal attention module suitable for the dual-branch 3D CNN,which effectively correlates the low-dimensional and high-dimensional spatiotemporal features by crossing.Aiming at the problem that the traditional model is difficult to deal with the spatiotemporal non-stationary in the data,which leads to the blurring and distortion of the predicted image,this paper constructs a non-stationary LSTM(NS-LSTM)based on Conv LSTM(Convolutional Long Short Term Memory)to extract the non-stationary features.NS-LSTM improves the fine-grained modeling ability of the model by extracting the higher-order features after the first-order difference of the input data.For the above two modules,this paper designs a gate control unit to selectively fuse global and local spatiotemporal information to avoid information loss and feature redundancy.A large number of experiments have been carried out on four public data sets.The experimental results show that the proposed H-STNet has better performance than the advanced model.(2)In this paper,H-STNet is applied to regional wind speed prediction to verify the generalization of the model.Aiming at the problem of uneven distribution of strong and weak winds,the optimization method of adaptive weight is introduced to increase the punishment of prediction error of strong wind samples to a certain extent.While balancing the differences between categories,the differences of samples within categories are also concerned.The experiment shows that the optimization method of adaptive weight improves the prediction accuracy of strong wind samples.The selection of superparameters has an important impact on the final effect of the model.In this paper,the improved sine-cosine optimization algorithm is applied to the superparameter optimization in the wind speed prediction task,and the population diversity and convergence speed of the algorithm are improved through reverse learning,Levy flight and chaotic mapping.The experimental results show that the hyperparametric optimization is helpful to improve the practical application effect of the model.