TGN: A Temporal Graph Network for Physics Prediction

The generation of synthetic time series is important in contemporary water sciences for their wide applicability and ability to model environmental uncertainty. Hydroclimatic variables often exhibit highly skewed distributions, intermittency (that is, alternating dry and wet intervals), and spatial and temporal dependencies that pose a particular challenge to their study. Vine copula models offer an appealing approach to generate synthetic time series because of their ability to preserve any marginal distribution while modeling a variety of probabilistic dependence structures. In this work, we focus on the stochastic modeling of hydroclimatic processes using vine copula models. We provide an approach to model intermittency by coupling Markov chains with vine copula models. Our approach preserves first-order auto- and cross-dependencies (correlation). Moreover, we present a novel framework that is able to model multiple processes simultaneously. This method is based on the coupling of temporal and spatial dependence models through repetitive sampling. The result is a parsimonious and flexible method that can adequately account for temporal and spatial dependencies. Our method is illustrated within the context of a recent reliability assessment of a historical hydraulic structure in central Mexico. Our results show that by ignoring important characteristics of probabilistic dependence that are well captured by our approach, the reliability of the structure could be severely underestimated.

PSTFormer: A novel parallel spatial-temporal transformer for remaining useful life prediction of aeroengine

Traffic forecasting is a challenging task due to complex spatial and temporal dependencies across sensor locations and time. Interest in solving this task has increased, but current research focuses on manually constructing neural network architectures without the aid of neural architecture search (NAS). In our work, we explore evolutionary neural architecture search (ENAS) by deploying a genetic algorithm (GA) to find optimal neural network architectures for predicting traffic conditions. The search space for the GA consists of arbitrary combinations of dilated convolutions and graph convolutions for modelling temporal and spatial dependencies respectively, limited in complexity only by technical constraints. Experimental results show that model architectures obtained via GA are able to match the current state-of-the-art on traffic prediction benchmarks.

Understanding crowd behavior is a pivotal step toward urban scene analysis. This is considered a very challenging task and has rarely been addressed to date due to the complexity of motion dynamics co-occurring across a given scene, which involves spatial and temporal dependencies. Unlike the mainstream research, which usually treats the temporal and spatial dependencies in a crowd separately, this paper presents a deep end-to-end approach that jointly considers the spatiotemporal information, leading to a rich understanding of crowd behavior. We first extract the displacement information describing crowd motion patterns from tracklets/trajectories. This information is subsequently fed into a convolutional layer to learn the underlying motion patterns and create high-level representations. The derived representations are used as inputs to a long short-term memory-based architecture to learn the underlying spatiotemporal cues in a single operation for the entire crowd in a given scene. We evaluate our approach on a widely used, large-scale benchmark datasets for three critical applications: pedestrian path forecasting, destination estimation, and holistic crowd behavior classification. The results show a drastic improvement compared to recently trending works.

A latent variable representation of count data models to accommodate spatial and temporal dependence: Application to predicting crash frequency at intersections

Dynamic multi-scale spatial–temporal graph convolutional network for traffic flow prediction

Traditional resource (demand or supply) forecasting models mainly focus on modeling temporal dependency. However, spatio-temporal data include complex non-linear relational and spatial dependencies. In addition, dynamic contextual information also impacts resources. Methods that consider context assume that the impact of context on resources is fixed, which is not realistic. For example, in a bicycle-sharing system, bike supply in stations is affected by the weather, and that effect changes over time. We propose a novel graph-based context integrated relational model, Context Integrated Graph Neural Network (CIGNN), which models temporal, relational, spatial, and dynamic contextual dependencies for multi-step ahead resource forecasting. We define a resource graph, where nodes represent locations with associated resource time-series, and context graphs (one for each type of context), where nodes represent locations with associated contextual time-series. Assuming that various contexts have dynamic impact on resources, our proposed CIGNN model employs a novel fusion mechanism that jointly learns from multiple contextual time-series. To the best of our knowledge, CIGNN is the first approach that integrates dynamic contextual information using graph neural networks for resource forecasting. Empirical results on two real-world datasets demonstrate that CIGNN consistently outperforms state-of-the-art approaches.

Traffic forecasting has emerged as an important task for developing intelligent transportation systems. Recent works focus on representing traffic as graph operation and using graph neural networks for spatial–temporal prediction. Most of the approaches assume a predefined graph structure based on node distances. However, spatial dependencies change over time in many scenarios of traffic flow. In this regard, this study takes an investigation capturing the spatial and temporal dependencies with no prior knowledge structure of traffic road networks. Specifically, we propose a multi-step prediction model named Dynamic Spatial Transformer WaveNet Network (DSTWN) to capture the dynamic conditions and directions of traffic flow in which a temporal convolution layer is adopted for the long time sequence and a spatial transformer layer is proposed to capture the dynamic spatial dependencies. Furthermore, we introduce a new traffic dataset, which is collected from the vehicle detection system in an urban area (UVDS). In particular, compared with existing benchmark traffic data, UVDS contains more complicated spatial information, which is similar to many real-world scenarios of traffic flow. Experiments on both benchmark traffic datasets indicate the promising results of DSTWN compared with state-of-the-art models in this research field.

Landslide displacement prediction has garnered significant recognition as a pivotal component in realizing successful early warnings and implementing effective control measures. This task remains challenging as landslide deformation involves not only temporal dependency within time series data but also spatial dependence across various regions within landslides. The present study proposes a landslide spatiotemporal displacement forecasting model by introducing attention-based deep learning algorithms based on spatiotemporal analysis. The Maximal Information Coefficient (MIC) approach is employed to quantify the spatial and temporal correlations within the daily data of Global Navigation Satellite System (GNSS) observations. Based on the quantitative spatiotemporal analysis, the proposed prediction model combines a convolutional neural network (CNN) and long short-term memory (LSTM) network to capture spatial and temporal dependencies individually. Spatial–temporal attention mechanisms are implemented to optimize the model. Additionally, we develop a single-point prediction model using LSTM and a multiple-point prediction model using the CNN-LSTM without an attention mechanism to compare the forecasting capabilities of the attention-based CNN-LSTM model. The Outang landslide in the Three Gorges Reservoir Area (TGRA), characterized by a large and active landslide equipped with an advanced monitoring system, is taken as a studied case. The temporal MIC results shed light on the response times of monitored daily displacement to external factors, showing a lagging duration of between 10 and 50 days. The spatial MIC results indicate mutual influence among different locations within the landslide, particularly in the case of nearby sites experiencing significant deformation. The attention-based CNN-LSTM model demonstrates an impressive predictive performance across six monitoring stations within the Outang landslide area. Notably, it achieves a remarkable maximum coefficient of determination (R2) value of 0.9989, accompanied by minimum values for root mean squared error (RMSE), absolute mean error (MAE), and mean absolute percentage error (MAPE), specifically, 1.18 mm, 0.99 mm, and 0.33%, respectively. The proposed model excels in predicting displacements at all six monitoring points, whereas other models demonstrate strong performance at specific individual stations but lack consistent performance across all stations. This study, involving quantitative deformation characteristics analysis and spatiotemporal displacement prediction, holds promising potential for a more profound understanding of landslide evolution and a significant contribution to reducing landslide risk.

Statistical methods to bias correct global or regional climate model output are now common to get data closer to observations in distribution. However, most bias correction (BC) methods work for one variable and one location at a time and basically reproduce the temporal structure of the models. The intervariable, spatial, and temporal dependencies of the corrected data are usually poor compared to observations. Here, the authors propose a novel method for multivariate BC. The empirical copula–bias correction (EC–BC) combines a one-dimensional BC with a shuffling technique that restores an empirical multidimensional copula. Several BC methods are investigated and compared to high-resolution reference data over the French Mediterranean basin: notably, (i) a 1D BC method applied independently to precipitation and temperature fields, (ii) a recent conditional correction approach developed for producing correct two-dimensional intervariable structures, and (iii) the EC–BC method. Assessments are realized in terms of intervariable, spatial, and temporal dependencies, and an objective evaluation using the integrated quadratic distance (IQD) is presented. As expected, the 1D methods cannot produce correct multidimensional properties. The conditional technique appears efficient for intervariable properties but not for spatial and temporal dependencies. EC–BC provides realistic dependencies in all respects: intervariable, spatial, and temporal. The IQD results are clearly in favor of EC–BC. As many BC methods, EC–BC relies on a stationarity assumption and is only able to reproduce patterns inherited from historical data. However, because of its ease of coding, its speed of application, and the quality of its results, the EC–BC method is a very good candidate for all needs in multivariate bias correction.

Traffic forecasting is of great importance to vehicle routing, traffic signal control and urban planning. However, traffic forecasting task is challenging due to several factors, such as complex spatial topological structure and dynamic changing of traffic status. Most existing methods have limited ability to capture both spatial and temporal dependence of traffic data. In this paper, we propose a novel end-to-end deep learning model, Attention based Graph Bi-LSTM networks (AGBN) to perform the traffic forecasting task. It uses graph convolutional network (GCN) to extract spatial features and bidirectional long short-term memory networks (Bi-LSTM) to capture the temporal dependence. The attention mechanism is used to select relevant features at all time steps. Experiments show that our model could extract both spatial and temporal dependence well and outperforms other baselines on real-world traffic datasets.

Traffic prediction provides vital support for urban traffic management and has received extensive research interest. By virtue of the ability to effectively learn spatial and temporal dependencies from a global view, Transformers have achieved superior performance in long-term traffic prediction. However, existing methods usually underrate the complex spatio-temporal entanglement in long-range sequences. Compared with purely temporal entanglement, spatio-temporal data emphasizes the entangled dynamics under the restrictions of traffic networks, which brings additional difficulties. Moreover, the computational costs of spatio-temporal Transformers scale quadratically as the sequence length grows, limiting their applications on long-range and large-scale scenarios. To address these problems, we propose a decomposed spatio-temporal Mamba (DST-Mamba) for traffic prediction. We aim to apply temporal decomposition to the entangled sequences and obtain the seasonal and trend parts. Shifting from the temporal view to the spatial view, we leverage Mamba, a state space model with near-linear complexity, to capture seasonal variations in a node-centric manner. Meanwhile, multi-scale trend information is extracted and aggregated by simple linear layers. Such combination equips DST-Mamba with superior capability to model long-range spatio-temporal dependencies while remaining efficient compared with Transformers. Experimental results across five real-world datasets demonstrate that DST-Mamba can capture both local fluctuations and global trends within traffic patterns, achieving state-of-the-art performance with favorable efficiency.

KSOF: Leveraging kinematics and spatio-temporal optimal fusion for human motion prediction

IML-UNet: A brain-inspired spatiotemporal collaborative encoding method for medical image sequence registration.

Spatio-temporal dynamic change mechanism analysis of traffic conflict risk based on trajectory data