Application Of Meta-Learning In Multivariate Spatial-Temporal Sequence Prediction

In recent years,due to the increasing development of sensor networks,hardware storage space,and positioning systems,multi-temporal and spatial sequences have shown explosive growth in various fields.The rational use of the massive amount of spatiotemporal data can bring great convenience to people’s production and life.However,traditional time series forecasting methods are difficult to extract spatial characteristics in spatiotemporal series,and existing machine learning methods have limitations in capturing complex and dynamic relationships between multiple variables,especially in scenarios where data is scarce or the underlying process changes rapidly.In recent years,the existing deep learning methods have been widely used in spatiotemporal series prediction,but there are problems such as long training time and weak model generalization ability.All in all,the nonlinearity,spatiotemporal correlation and high latitude of multivariate spatiotemporal series bring great difficulties to prediction,and how to effectively construct a multivariate spatiotemporal series prediction model is a common problem to be solved at the technical level of spatiotemporal forecasting.As one of the hot research directions in the field of deep learning in recent years,meta-learning can quickly adapt the model to new tasks and environments by learning from different tasks,and enhance the generalization ability of the model.Based on this,in order to meet this challenge,this paper locates the specific tasks of multivariate spatiotemporal sequence prediction on the two problems of urban PM2.5 concentration prediction and groundwater level prediction,and skillfully integrates the strong generalization ability of meta-learning algorithms with spatiotemporal correlation and nonlinearity analysis methods of multivariate spatiotemporal sequences,and constructs a new deep learning model for meta-learning algorithms,attention mechanisms,variational autoencoders and hybrid modeling mechanisms for multivariate spatiotemporal sequence prediction.The main research results of this paper are summarized below:(1)Aiming at the core problem of urban PM2.5 concentration prediction and early warning,the urban area is divided into detailed zones for forecasting and analysis for the first time,and a new spatiotemporal element learning model for PM2.5 prediction is proposed.The basic construction of the model design includes multiple convolutional neural network(CNN)modules and Long Short-Term Memory networks(LSTM),the former can capture the spatial correlation characteristics in the multivariate spatiotemporal series,and the local trends in the time series data,and the latter is used to learn the long-term dependence and spatiotemporal correlation characteristics in the spatiotemporal series.The adaptive feature attention module is designed between CNN and LSTM,which can more effectively enhance the multivariate features in the space-time sequence for feature extraction.The design of CNN+Feature Attention+LSTM can adaptively extract spatiotemporal correlations and improve prediction capabilities.Based on the real urban air quality dataset,experimental analysis shows that the model can effectively predict urban PM2.5 concentration.(2)On the architecture based on hybrid deep learning,the network is divided into two,designed as a structure of inference network + decoder,and cleverly combines the network structure with the structure of variational autoencoder network.By designing efficient metatraining and meta-testing algorithms,the end-to-end prediction task of the model is effectively solved.For the first time,the relevant theory of Bayesian meta-learning is applied to the prediction problem of air pollutants,and the concept of variational inference is applied to the meta-learning algorithm to capture the spatiotemporal dependence and uncertainty of the dynamic change of air pollutant concentration.Extensive experiments were conducted on urban air quality datasets to evaluate our proposed method.The experimental structure verifies that the proposed model can predict the divided urban area and outperforms other benchmark models.(3)Aiming at the bottleneck faced by traditional groundwater level prediction methods,a hybrid deep network model based on meta-learning algorithm is proposed to improve the accuracy and effectiveness of groundwater level prediction.In this study,a hybrid CNN-LSTM network structure is designed,and CNNs are used to extract local multivariate trend features affecting groundwater level depth,and LSTM is used to treat the temporal dependence after feature fusion.The dynamic changes of groundwater level are simulated in space and time.(4)For deep learning models,large-scale training data is usually required as support,and the characteristics of realistic groundwater prediction problems usually lack sufficient samples.In this study,a meta-learning algorithm is added to the CNN-LSTM network structure,so that the model can train the meta-learning model with fewer training samples,so as to complete the groundwater level prediction task.Similarly,for a new task,the model can gain knowledge from the previous groundwater level prediction task without training from scratch.In this study,extensive experiments were carried out on the real groundwater level dataset of the middle and lower reaches of the Heihe River to verify the performance of the proposed model.Experimental results show that compared with other control models,the prediction effect of the proposed model is better,and it can maintain high prediction accuracy even under the condition of insufficient training samples.The proposed model can accurately predict groundwater levels.