深度学习模型在多元时间序列预测、智能驾驶、图像识别等多个领域广泛应用,其中多元时间序列预测是学者们关注的重点之一,多元时间序列预测是典型的回归任务,旨在通过海量的历史数据构建模型以预测未来状态,被广泛运用于交通、电力、金融等领域.多元时间序列数据具有复杂的时空依赖性,现有模型大多仅能捕获序列数据中的时间特征,难以捕获空间特征,而图神经网络解决了这一问题.图神经网络能够自然地建模实体间的复杂关系,可以很好地处理拓扑数据,而多元时序数据大多可以构造为拓扑图,因此图神经网络可以很好地学习多元时序数据中的空间特征.基于图神经网络的多元时间序列预测模型受到广泛关注并取得了一定的成果,但现有基于图神经网络的模型仍存在诸多不足.首先,现有方法大多分别捕获和建模多元时间序列数据中的空间特性和时间特性,未充分考虑多元时间序列的时空统一性,导致模型的次优建模;其次,现有方法主要基于静态预定义图或动态自适应图,其中静态预定义图通常根据监测节点之间的空间相关性进行构造且不会随着时间而改变,基于预定义图的研究忽略了时间序列数据中的时间特征,即忽略了数据模式随时间发生的改变;而自适应图通常由模型自主学习并不包含监测节点间的固有属性,基于自适应图的研究忽略了大量有效的领域知识,如道路的连通性和道路间的属性.为了解决上述问题,提出基于动态自适应时空图的多元时序预测模型MTP-Graph(Multivariate Time series Predic-tion model based on dynamic adaptive spatio-temporal Graph),利用时空融合模块将时空信息进行统一处理,避免了分开捕获时间特性与空间特性而导致的次优建模问题,提出图结合模块将静态预定义图和动态自适应图进行动态融合,获取时空信息的同时充分考虑领域知识,使模型可以更好地学习多元时间序列中的时空特性.在PeMSD3、PeMSD7和PeMSD8数据集上的大量实验结果表明,MTP-Graph预测性能优于其他基准方法,验证了MTP-Graph的可用性和有效性.
A Multivariate Time Series Forecasting Model Based on Dynamic Adaptive Spatio-Temporal Graphs
Deep learning models have been extensively used in many fields such as multivariate time series prediction,intelligent driving,image recognition,and so on.Among them,multivariate time series forecasting is one of the focuses from scholars.Multivariate time series forecasting is a typical regression task,which aims to construct a model to predict the future states through a huge number of historical data,and it is widely used in the fields of transportation,electricity and finance.Multivariate time series data have complex spatio-temporal dependencies,the existing models can only capture temporal features,but it is difficult to capture spatial features.The graph neural network emerges to solve this problem.Graph neural network can naturally model the complex relationship between entities and is very good at dealing with topological data,and most of the multivariate temporal data can be constructed as a topological graph,so the graph neural network can better learn the spatial features from multivariate temporal data.Multivariate time series prediction models based on graph neural networks have received widespread attention and achieved certain success.However,there are some drawbacks in the existing methods based on graph neural networks.Firstly,the existing methods only capture and model the spatial features or temporal features from multivariate time series data separately,without considering the spatio-temporal unity of multivariate time series data,leading to the suboptimal modeling problem.Secondly,existing methods are mainly based on static predefined graphs or dynamic adaptive graphs,the static predefined graphs are usually constructed based on the spatial correlations among the monitored nodes and do not change over time,and the research based on static predefined graph often ignore temporal features of time series data,which ignores the changes in data patterns over time.Whereas,adaptive graphs are usually learnt autonomously by the model and do not contain intrinsic properties among the monitored nodes,so the research based on adaptive graphs ignore a large number of effective domain knowledge,such as road connectivity and properties between roads.In order to solve the above problems,a Multivariate Time series Prediction model based on dynamic adaptive spatio-temporal Graph called MTP-Graph is proposed,which unifies the spatio-temporal information by using spatio-temporal fusion module,avoiding sub-optimal modelling problem caused by capturing temporal and spatial features separately,and the proposed graph combination module dynamically integrates combines the predefined graphs with adaptive graphs,and takes into full consideration domain knowledge while obtaining spatio-temporal information from data,in order to better learn the spatio-temporal features from multivariate time series data.Extensive experiments are conducted on the PeMSD3,PeMSD7 and PeMSD8 datasets,and the results show that MTP-Graph outperforms other benchmark methods in prediction performance,which shows the usability and effectiveness of MTP-Graph.
multivariate time series predictionspatio-temporal databasegraph neural networksattention mechanismmachine learning
万方数据
维普
