Riding the Wave: Multi-Scale Spatial-Temporal Graph Learning for Highway Traffic Flow Prediction Under Overload Scenarios

1 Introduction

Highway traffic overload is a condition where the volume of vehicles on a highway exceeds its designed or operational capacity, leading to congestion, reduced speeds, and inefficiencies in traffic flow [Jin et al., 2023b; Cui et al., 2020]. According to TomTom's 2024 statistics, the time spent by users traveling through Indianapolis highway sections increased by 27% due to traffic overload [Martichoux, 2025]. Predicting highway traffic flow under such situations can help alleviate bottlenecks by enabling traffic diversion or recommending optimal entry and exit times, reducing negative impacts on society and the economy.

∗Corresponding author

Highway overload scenarios differ from normal scenarios in both traffic dynamics and spatial interactions. We illustrate these differences using a case study of Highway G56 in Hangzhou, China, as depicted in Figure 1. In overload scenarios, the highway traffic flow consistently surpasses the peak threshold observed under normal conditions (Figure 1(c)), often causing severe traffic congestion. Such high traffic volume usually compels drivers to divert to parallel local roads to bypass congestion (Figure 1(b)), leading to spatial interactions between highways and local roads. However, most existing methods [Jin et al., 2023a; Wu et al., 2019; Han et al., 2024] primarily focus on normal scenarios and overlook the distinct characteristics of overload scenarios.

We study the problem of HIghway traffic flow Prediction under Overload scenarios, abbreviated as HIPO. This problem is more challenging than ordinary traffic prediction tasks, which is reflected in the following crucial difficulties: 1) Irregular traffic patterns. Traffic patterns during overload periods exhibit significant changes compared to normal periods. As shown in Figure 1 (c), traffic volume often exceeds highway capacity, causing prolonged congestion. Incidents and interweaving behaviors add sudden fluctuations, further destabilizing traffic flow. These unpredictable disruptions break regular traffic patterns, making traditional methods designed for consistent dynamics ineffective in capturing and predicting irregular patterns in overload scenarios. 2) Complex traffic behaviors. In Figure 1 (a) and (b), the traffic flow on highway segment s0 can enter local segment p through the interweaving area s1 when vehicles either reach their destinations or seek to avoid congestion. Conversely, segment p can also serve as the entry point for the highway, merging into segment s2 through area s1. These complex traffic behaviors obviously affect highway segments' flows. However, previous data-driven prediction approaches assume that vehicles remain solely on highway segments without accounting for diversion to local roads. 3) Insufficient contextual data. Highway overload scenarios, such as holidays, are concentrated in a few days, inherently leading to data sparsity. Other approaches that leverage point-of-interests or similar urban information for traffic prediction become ineffective on highway networks. The limited data is insufficient to support training a model with strong generalization ability.

To overcome these challenges, we propose a Heterogeneous Spatial-Temporal graph network With Adaptive contrastiVE learning (HST-WAVE) to address the HIPO problem. Specifically, we first construct a heterogeneous traffic graph (HTG) to depict various transfer interactions on highway networks. We then develop a multi-scale weaving Transformer network to adapt to irregular traffic patterns. A coupled heterogeneous graph attention network performed on HTG is delivered to learn the complex traffic behaviors. Both types of networks learn alternately to form a heterogeneous spatial-temporal module (HSTM) as the primary learner. To further improve the model's generality when facing irregular traffic patterns while alleviating the data sparsity, we develop four temporal augmentation strategies (flip, mask, replace, and noise) in a mini-batch data to implement contrastive learning. Moreover, we incorporate an adaptive learnable temporal pattern with a fixed length, concatenated with the original pattern, to retain common temporal knowledge during training. Our contributions are:

• We investigate a significant yet often neglected problem: highway traffic flow prediction under overload scenarios (HIPO). This problem is notably different from existing traffic prediction tasks due to the unique challenges caused by irregular flow patterns, complex traffic behaviors, and limited contextual data.
• We propose the HST-WAVE framework to address the HIPO problem. Our approach introduces a multi-scale temporal weaving module and a heterogeneous spatial interaction module, which effectively captures irregular temporal patterns and complex traffic behavior dynamics. Additionally, we propose an adaptive temporal enhancement contrastive learning strategy to enhance the model's generalization capability and mitigate data sparsity in overloaded scenarios.
• Extensive experiments on two real-world highway datasets demonstrate that our model achieves state-ofthe-art performance, confirming its effectiveness in addressing the HIPO problem.

2 ProblemFormulation for HIPO

We study the HIPO problem on a highway road that runs parallel to a local road. Since the highway comprises two separate, non-interfering bidirectional lanes, this study focuses solely on the single-directional highway with the highest traffic flow for simplicity.

Highway Network. A highway network consists of a highway road S and a parallel local road P, denoted by V = {S,P}. The highway road S = {s1, s2, ..., sn} is composed of directed highway segments. The interweaving area S P ⊂ S is a set of highway segments connecting the local road P. The local road P = {p1, p2, ..., pm} is made up of bidirectional local segments, where each segment pi ↔ pi+1. Within the network, highway segments are sequentially connected in one direction (si → si+1) while the interweaving area s ∈ S P serves as an interaction point linked with the corresponding local segment p ∈ P.

Overload Scenario. An overload scenario is a condition of the highway network where the traffic demand significantly exceeds the highway's capacity. Overload scenarios typically arise during holidays, road maintenance, or special events, performing irregular temporal patterns and increased interactions between S P and P.

Problem Formulation. Given a highway network {S,P}, we assign a traffic flow placeholder matrix into each type road, i.e., X = {XS ,XP }. We intercept a time interval with T +H steps during overload periods. Then we record the traffic flow in historical T-steps on all road segments at time step t, i.e., XS t-T:t = {xs1 t-T :t , xs2 t-T :t , ..., xsn t-T :t }, XP t-T:t = {xp1 t-T :t , xp2 t-T :t , ..., xpm t-T:t }, and Xt-T:t = {XS t-T :t , XP t-T:t }. The HIPO problem is viewed as a time series auto-regressive task, given the historically observed traffic flow matrix Xt-T:t, we aim to predict the traffic flow XS t:t+H within future H steps on highway segment set S. The HIPO problem can be formulated as:

where f(·) is the traffic prediction function.

3 HST-WAVE Framework

This section presents the HST-WAVE framework for solving the HIPO problem, as shown in Figure 2. The framework consists of two components: 1) Heterogeneous SpatialTemporal Module (HSTM) blocks integrating a Multi-Scale Weaving Transformer (MSWT) and a Coupled Heterogeneous Graph Attention Network (CHGAN); 2) an Adaptive Temporal Enhancement Contrastive Learning (ATECL) strategy integrating temporal enhanced contrastive learning and learnable temporal patterns.

3.1 Traffic Modeling with HSTM

Most current traffic flow prediction methods [Yu et al., 2018; Kong et al., 2024; Zhou et al., 2024; Jin et al., 2023a] de- Section 3.2

Figure 2: The overview of HST-WAVE's architecture. (1) The main work process for HST-WAVE, the lower part is a temporal enhancement contrastive learning module. (2) A multi-scale weaving Transformer comprises multi-scale 1D-gated CNNs, a Transformer encoder, and interscale mutual attention. (3) A coupled heterogeneous graph attention network includes a bidirectional heterogeneous attention mechanism and a hop decay aggregation.

pend on homogeneous graph modeling and simplistic temporal learning mechanisms, which are inadequate for highway traffic modeling under overload conditions. To address this limitation, we design a novel heterogeneous spatial-temporal block, termed HSTM, to effectively learn irregular traffic patterns and complex traffic behaviors.

The HSTM block is built upon a heterogeneous traffic graph (HTG) that captures the diverse types of road segments and their transfer interactions, as illustrated in Figure 1 (b). In this representation, each road segment-whether on a highway or a local road-is modeled as a node, while the transfer interactions between segments are represented as edges. An HTG is formally defined as G = {V, E, T ,R}, where V = {S,P} is the set of nodes, E is the set of edges, T = {highway, local} is the set of node types, R = {in, out, fwd, bwd} is the set of edge types, and N = |V| is the number of nodes. We define τ(v) ∈ T and ϕ(e) ∈ R as type mapping functions for nodes and edges, respectively, where v ∈ V and e ∈ E. Specifically:

where s∗ ∈ S, s′∗ ∈ S P , and p∗ ∈ P are nodes in V. The direction of an edge is specified by i < j.

Each node v ∈ V is associated with temporal features hv ∈ RT×d , representing the traffic flow on the corresponding highway or local road segment over the past T time steps, where d is the dimension of the temporal feature, and H ∈ RN×T×d denotes the temporal features of all nodes. Given the traffic flow matrix Xt-T:T , we construct the initial feature H0 by incorporating the traffic flow value, time of day, and day of the week for each time step.

We then input H0 into the HSTM blocks, which are formed by alternately stacking MSWT and CHGAN:

where Hl ∈ RN×T×d is the output of all nodes, and l denotes the number of layers. MSWT and CHGAN learn the complex temporal and spatial dynamic features of highway traffic data.

Learning Irregular Traffic Patterns with MSWT

Traffic flow in overload scenarios exhibits sustained high volumes and frequent fluctuations. Existing methods in temporal learning typically rely on time-step-based models [Yu et al., 2018; Jin et al., 2023a], which struggle to capture the dependencies with mixed irregular temporal patterns. We propose a multi-scale weaving Transformer (MSWT), incorporating single- and multi-window weaving attention to handle such irregular traffic patterns. MSWT module inputs Hl-1 that contains all nodes' temporal features and transforms hvl-1 of each node v into the output ˆhvl-1 . We omit the subscripts v and l - 1 for convenience.

We first improve the gated CNNs [Yu et al., 2018] by using multiple temporal scales to decompose the irregular temporal patterns into multi-scale window pattern representations. Given the temporal feature h for node v, we employ 1-D CNNs with c scale kernels combined with the gated linear unit (GLU) along the temporal dimension T. For the kernel size wi, the window pattern is sliding on h with padding wi - 1 zeros, denoted as Padwi (h). The convolution kernel Fwi ∈ Rwi×d×2d implements on it and following the GLU operation:

where P, Q ∈ RT×d are two parts with the d dimension in the output, which are computed by P ||Q = Fwi ∗Padwi (h). The function σ(Q) is a sigmoid gate determining the relevance of input P from the current states for uncovering compositional structure and dynamic variations in each window.

We then input the window pattern embedding hwi ∈ RT×d of each granular into a base Transformer encoder to achieve the interaction modeling between single-scale windows:

In the multi-scale window design, small-scale windows capture fine-grained, localized traffic variations, while largescale windows focus on broader temporal trends and context. Furthermore, we propose mutual attention pooling on the stacked output ˆhw ∈ Rc·T×d of multiple scale window Transformers to integrate local and global insights:

where WindowSum(·) is a summation function that applies on the window scale. The temporal outputs ˆh ∈ RT×d are concatenated into the representation ˆHl-1 of all nodes (in Equation 2). This attention allows the model to adaptively prioritize relevant temporal scales based on inputs, ensuring flexibility in handling mixed irregular temporal patterns.

Learning Complex Traffic Behaviors with CHGAN

The frequent transitions of traffic flow along highway segments, local segments, and interweaving areas under overload scenarios result in complex traffic behaviors. We develop a coupled heterogeneous graph attention network (HTG) to model these interactions in a Transformer-based manner [Mao et al., 2023]. CHGAN maps the MSWT's output ˆhvl-1 ∈ ˆHl-1 to the spatial output ˆh∗l ∈ Hl on G, which is formulated as a message passing mechanism liking general GNNs, and the subscript l - 1 is omitted:

where Agg(·), Attn(·), Msg(·), and Update(·) are the aggregation, attention, message, and update functions.

In CHGAN, the attention function Attn(·) generates the attention scores between two nodes by considering node types to learn important traffic behaviors across different segments. Specifically, for a target node v ∈ V, we select its all k-hop neighbors set N(v). We then define a group-specific Query and Key projected function for each node type. The target node v and source node u ∈ N(v) are mapped into Query and Key vectors to calculate dot product as attention:

Furthermore, we add the relative edge type bias to adjust the attention scores to enhance heterogeneous traffic interaction modeling. We assign a one-hot vector hϕ(ei) ∈ R|R| to each edge type, and the edge type feature is indexed by a learnable edge type matrix HR ∈ R|R|×d , i.e., ˆhϕ(ei) = hϕ(ei)HR. For node pair (u, v), we compute the average of edge type features along the path (e1, e2, ..., ek) from u to v, and a linear projection is employed to generate bias item:

the final attention score is calculated by:

The message function Msg(·) extracts the source node's feature of target node v by a Value projection:

We eventually aggregate all source nodes' messages into the target node by Agg(·). Considering the neighbor proximity, traffic messages from source road segments at the highhop pass into the target road segment will consume more than at the low-hop. Therefore, we design a simple decay function without trainable parameters in the aggregation stage:

where λ is the decay factor, and kd is the hop of a source node. Let the first output of Agg(·) be denoted by fwd(hN(v)) which incorporates the traffic flow features from all other nodes with different types and relations flowing to the target node. We flip the edge direction in HTG and put it into the above computation process to obtain the output bwd(hN(v)), which empowers the model with the ability to perceive the traffic flow features from all nodes reached by the target node's outgoing flow.

We use an activation linear projection combining the residual connection to update the target node's information from bidirectional sources:

where h∗ ∈ RT×d is the output of Update(·) in target node v, and we denote the spatial representation of all nodes as Hl (in Equation 2). CHGAN improves the heterogeneous graph Transformer [Mao et al., 2023] by introducing relative traffic behavior type learning with a hop decay function and bidirectional traffic flow-aware heterogeneous attention, effectively capturing complex traffic behaviors.

3.2 Traffic Data Augmentation with ATECL

The rarity of overload events leads to insufficient contextual data for training, requiring models to rely on normal traffic data. However, the temporal patterns and behaviors observed in normal traffic differ substantially from those during overload scenarios. Therefore, we devise an adaptive temporal enhancement contrastive learning (ATECL) strategy based on HSTM in this section, which enhances model generalization by bridging the gap between divergent temporal patterns and alleviates data sparsity by augmenting overload scenario data.

To enhance the generality of the model, we apply four augmentation operations A = {Flip,Mask,Replace,Noise} at the temporal level, where each operation serves to simulate different traffic conditions. Flip: we randomly flip T · γ positions by max ± x (using + if x ≤ 0, and - otherwise) to simulate temporal shifts in traffic patterns; Mask: we randomly mask T · γ positions to 0 to force the model to learn temporal contextual dependencies; Replace: we randomly replace T · γ positions to noise to simulate fluctuating traffic conditions; Noise: we add the Gaussian noise to all positions to mimic unpredictable changes in traffic behavior. Here, γ represents the modification rate, x denotes a value at an arbitrary position within a temporal pattern, and max indicates the maximum flow.

Given a batch of Nb samples, we sample two augmentation operations oi, oj ∈ A for each temporal pattern to obtain Hoi 0 ,Hoj 0 and put them into the HSTM model to obtain the 2Nb outputs. Following the general contrastive learning task [Chen et al., 2020], we enhance the similarity between positive pairs (Hi,Hj) from the same temporal pattern and decrease the similarity between negative pairs (Hi, Hk) from 2Nb - 2 different temporal patterns:

where I = {0, 1} is a binary function outputting 1 if k ̸= i and η is a temperature parameter.

To further enhance the learning process and preserve common temporal knowledge across different augmentations, we introduce an adaptive temporal pattern lv ∈ RN×T ′×d , which is concatenated with the augmented sequence:

where the pattern is initialized to a normal distribution and updated during model training, acting as a memory to retain key temporal cues regardless of the specific augmentation applied. It helps the model avoid over-fitting to noise while ensuring generalization to real-world overload scenarios.

3.3 Loss Function

We employ a full connection layer to map the output Hl into a prediction traffic flow matrix ˆXS t:t+H , and the ground truth matrix is XS t:t+H . For the HIPO problem, we use the mean absolute error as the training objective loss defined by:

The total training loss is calculated by the weighted sum of regression loss and contrastive loss:

where µ is a control factor.

4 Experiments

We conduct extensive experiments on two real-world datasets to evaluate our HST-WAVE's performance and answer the following research questions: Q1) How much improvement does the HST-WAVE achieve on the HIPO task compared to methods designed for normal traffic flow prediction? Q2) How much contribution does each component of the HSTWAVE make to the accuracy of addressing the HIPO task? Q3) How robust is the HST-WAVE on the segment with the highest traffic load compared to existing methods? Q4) How sensitive is the HST-WAVE to the different settings for hyperparameters?

4.1 Experiment Settings

Datasets. We collect the traffic flow datasets on two highway networks during the normal and overload scenarios, including highways No. G56 and G60 are located in Hangzhou, China, and we consider the holiday periods to be overload scenarios. The traffic flow is obtained by recording the vehicle track points within the kilometer piles at a sampling frequency of 5 minutes, and the specific information is described in Table 1.

For both the G56 and G60 datasets, the time range for normal scenarios spans from June 10 (or June 1) to September 27, 2023, while the time range for overload scenarios is from September 28 to October 3, 2023 (the Mid-Autumn Festival and National Day in China). The highway network of G56 consists of 110 segment nodes, including 100 highway nodes (10 of which are interweaving area nodes) and 10 local segment nodes, while G60 comprises 143 segment nodes, including 130 highway nodes (13 of which are interweaving area nodes) and 13 local segment nodes. We select all data in the normal scenarios and two days of data in the overload scenarios as the training set, and the rest of the dataset (four days with 1152 time steps) in the overload scenarios as the test set. The historical 60-minute data is used to predict the traffic flow for the next 20, 40, and 60 minutes.

Metrics and Baselines. Following the previous works [Zhou et al., 2024] in traffic flow prediction, Mean Absolute Errors (MAE), Root Mean Squared Errors (RMSE), and Mean Absolute Percentage Errors (MAPE) are adopted to measure the performance of methods. We select nine representative baselines for comparison with HST-WAVE: Historical Average (HA) [Jiang et al., 2021], VAR [Lu et al., 2016], DCRNN [Li et al., 2018], STGCN [Yu et al., 2018], Graph Wavenet (GWNet) [Wu et al., 2019], Trafformer [Jin et al., 2023a], PDFormer [Jiang et al., 2023], STPGNN [Kong et al., 2024], and DCST [Zhou et al., 2024].

Model Settings. We implement our HST-WAVE by using PyTorch Lightning on the NVIDIA GeForce 3090 GPU with 24 GB memory. For the model's hyper-parameters, the hidden dimension d of HSTM is set to 64, the layers of MSWT and CHGAN are set to 2, the head of attention is set to 4, the hop decay factor λ is set to 0.5, the loss weight coefficient µ is set to 0.2, and the size T ′ of lv is set to 12. We use the Adam optimizer during the model training, the learning rate is set to 0.001, and the batch size is set to 64.

4.2 Comparison Results (RQ1)

Table 2 summarizes the experimental results of the HSTWAVE and baseline models for 20-, 40-, and 60-minuteahead predictions on the G56 and G60 datasets. The results demonstrate that HST-WAVE consistently outperforms all baseline methods across three evaluation metrics in the HIPO task. Notably, at the 60-minute prediction horizon, HST-WAVE achieves substantial improvements over the bestperforming existing methods, with MAE, RMSE, and MAPE improving by 11.1%, 11.0%, and 13.2% on G56, and by 17.7%, 11.5%, and 12.7% on G60, respectively.

HA and VAR perform the worst since they fail to handle complex spatial-temporal data. STPGNN and PDFormer achieve the second-best performance among all baseline methods. STPGNN focuses on identifying key nodes in traffic graphs, which helps capture partial interaction information in interweaving areas. On the other hand, PDFormer leverages the relationships between similar nodes and temporal patterns, enabling it to adapt to varying traffic conditions.

HST-WAVE surpasses STPGNN and PDFormer due to its tailored design for the HIPO task. The MSWT module empowers it to address irregular temporal dependencies across diverse traffic conditions, while the CHGAN module equips the model to effectively capture complex transition relationships between road segments. Furthermore, the ATECL module significantly enhances the model's generalization ability in challenging overload scenarios.

4.3 Effectiveness of Component (RQ2)

We conduct ablation studies to evaluate the effectiveness of each component in HST-WAVE. The variants are as follows. 1) w/o HG: this variant replaces the CHGAN with the GCN. 2) w/o MS: this variant removes the multi-scale window convolutional operation. 3) w/o lv: this variant drops the adaptive temporal pattern. 4) w/o CL: this variant does not adopt the ATECL strategy.

Figure 3 presents the comparison of the above variants, which illustrate that: 1) w/o HG performs worse than HSTWAVE, which reflects the necessity of the CHGAN module in capturing the dynamic flow transitions among various type segments. 2) The performance of w/o MS also decreases, which confirms the effectiveness of the multi-scale window modeling mechanism in learning complex traffic patterns. 3) Removing the ATECL strategy (w/o CL) significantly degrades the model performance. This reason is that ATECL

contributes to adapting to the irregular temporal patterns and data sparsity in overload scenarios. 4) w/o lv performs better than w/o CL in G56, but the opposite is valid on the G60. This suggests that lv in ATECL can mitigate the noise impact by preserving key temporal pattern knowledge.

4.4 Performance on Interweaving Areas (RQ3)

We verify the performance of the HST-WAVE on interweaving areas under overload scenarios, which are characterized by high traffic flow and frequent traffic behaviors. The results are shown in Figure 4, and we observe the following findings. HST-WAVE significantly outperforms PDFormer and STPGNN across all metrics on both datasets while exhibiting smaller error increments compared to the other methods. Although STPGNN achieves similar error increments to HST-WAVE on G60 (Figure 4 (d), (e), (f)), its overall performance remains substantially inferior to ours. Moreover, unlike STPGNN, our method does not rely on explicitly identifying key nodes to achieve superior performance in interweaving areas. This indicates that HST-WAVE attains high generality by effectively learning the complex and heterogeneous interactions present in interweaving areas.

4.5 Hyper-parameter Sensitivity (RQ4)

We select three important hyper-parameters of the HSTWAVE to analyze the effects on performances, including the weight coefficient µ of contrastive learning loss, the hidden dimension d of the model, and the length of the adaptive temporal pattern lv. Figure 5 shows the trend of MAE under different hyper-parameter settings on both datasets and default hyper-parameters in our experiments are marked by vertical dashed lines. We have the following observations. 1) We can obtain the optimal results when setting µ = 0.3, which strikes an effective balance between the traffic flow prediction loss and the enhancement contrastive loss. 2) The model's performance improves steadily as the d increases to 64, but larger dimensions offer no additional benefits and may increase the risk of over-fitting. 3) The optimal length for lv is set to 12, which is sufficient to capture key temporal cues and guide accurate predictions in overload scenarios.

5 RelatedWork

5.1 Traffic Prediction

Traffic prediction is a critical research area due to its significant role in traffic management and urban planning [Gomes et al., 2023]. Early approaches relied on traditional statistical models like ARIMA [Min and Wynter, 2011] and Bayesian models [Wang et al., 2014] but struggled with tackling complex spatial-temporal dependencies. Deep learning has exhibited superior potential in capturing these dependencies. For spatial modeling, Convolutional Neural Networks (CNNs) [Zhang et al., 2020] and Graph Neural Networks (GNNs), including variations like Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs) [Wu et al., 2020; Ye et al., 2021; Song et al., 2022; Ji et al., 2023; Kong et al., 2024], are utilized to effectively represent road network structures and interactions. For temporal dependencies, CNNs and recurrent architectures like LSTM and GRU [Ma et al., 2015] are used to extract sequential features.

Recent advances in traffic forecasting have been marked by the advent of Transformer-based approaches [Liu et al., 2023; Jin et al., 2023a; Jiang et al., 2023; Zhou et al., 2024], such as Trafformer [Jin et al., 2023a], DCST [Zhou et al., 2024], and PDFormer [Jiang et al., 2023]. These approaches have notably enhanced prediction accuracy by simultaneously modeling spatial and temporal dependencies, setting a new benchmark in the field. Other studies [Lu et al., 2018; Wang et al., 2019b] focus on traffic prediction during holidays simply using parameter modeling of temporal patterns, which has difficulty handling complex spatial-temporal correlations, particularly traffic behaviors in interweaving areas.

5.2 Heterogeneous Graph and Time Series Augmentation

Heterogeneous graphs [Sun and Han, 2013] effectively model and integrate diverse types of entities and relationships in complex data, enabling more accurate representation learning. In recent years, attention mechanisms have been widely applied in the research of heterogeneous graph neural networks, such as HAN [Wang et al., 2019a], HGT [Hu et al., 2020], SlotGAT [Zhou et al., 2023].

Time series augmentation techniques [Wen et al., 2021; Zanella et al., 2022; Sarkar et al., 2020; Cheung and Yeung, 2021] aim to enhance model generalization on small or imbalanced datasets. They are typically applied to domainspecific temporal data or regular temporal patterns [Eldele et al., 2021; Franceschi et al., 2019], resulting in suboptimal performance when adapted to the HIPO problem.

6 Conclusion

In this paper, we investigated a novel problem: highway traffic flow prediction under overload scenarios, which is more challenging than ordinary traffic prediction tasks. We proposed a heterogeneous spatial-temporal graph network with adaptive contrastive learning, HST-WAVE, to address this issue. The framework incorporated MSWT and CHGAN modules within HSTM blocks to capture irregular temporal patterns and complex traffic behaviors, while the ATECL strategy improved generalization and mitigated data sparsity. Extensive experimental results demonstrated the effectiveness of our HST-WAVE on the HIPO problem.

Acknowledgments

This work is supported by the National Key R&D Program for the 14th-Five-Year Plan of China (2023YFC3804104 in 2023YFC3804100).