💻 computer science

Traffic flow forecasting with explicit priors and prior-conditioned adaptive fusion

This paper proposes a traffic flow forecasting framework that explicitly separates dynamic observations from structural and periodic priors, integrating them via a novel Prior-Conditioned Adaptive Fusion (PCAF) mechanism to achieve state-of-the-art accuracy across multiple datasets.

Original authors: Peibei Jiang, Xiang Gu, Yuxiang Guo, Jing Wen, Dongjie Wang

Published 2026-08-12

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Peibei Jiang, Xiang Gu, Yuxiang Guo, Jing Wen, Dongjie Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather for next Tuesday. You have three different tools in your kit: a live radar showing rain clouds moving right now, a map of the local mountains and valleys that never change, and a calendar reminding you that Tuesdays are usually windy. For a long time, computer scientists trying to predict traffic have been throwing all three of these tools into a single, giant blender. They mix the live traffic data, the road map, and the time of day together in one big soup, hoping the computer can figure out which ingredient is important when. But just like a smoothie where you can't taste the strawberry from the banana, this mixing makes it hard for the computer to learn exactly how the road layout or the daily routine helps predict the future. This paper tackles that problem by asking: what if we kept the tools separate until the very last second, letting them guide the prediction only when needed?

This research, led by Peibei Jiang and colleagues at Nantong University, proposes a new way to forecast traffic flow that treats these three ingredients—live traffic, road structure, and daily patterns—as distinct guides rather than a blended mess. The team built a system that first learns from the "live radar" (recent traffic history) on its own. Then, it brings in the "map" (the fixed layout of the roads) and the "calendar" (the time of day) as special helpers. Instead of just mixing them in, the system uses a clever "adaptive fusion" mechanism. Think of it like a smart traffic director who listens to the live report but also checks the map and the calendar to decide how much weight to give each piece of information. If the traffic is behaving normally, the director might trust the daily pattern more; if there's a sudden jam, they might rely more on the live data. The paper suggests that by keeping these sources separate and letting a smart gate decide how to combine them, the computer can make much sharper predictions about where traffic will be in the next hour, two hours, or even eight hours.

The Traffic Time-Travel Problem

Predicting traffic is a bit like trying to guess the ending of a movie while it's still playing. You have the current scene (the traffic right now), the script (the road layout), and the genre (it's rush hour, so it's probably busy). The goal is to see the next few scenes before they happen. In the real world, this matters because traffic agencies need to know if a road will clog up in 30 minutes so they can change traffic lights or tell drivers to take a different route.

The problem is that traffic data is messy. It changes fast (a car brakes suddenly), but the roads themselves are slow to change (a bridge doesn't move). Plus, traffic has a rhythm; it's always heavy at 8:00 AM and light at 3:00 AM. Previous computer models tried to solve this by shoving all this information into one big neural network. They treated the road map, the time of day, and the live traffic numbers as if they were all the same kind of data. The authors of this paper argue that this is like trying to read a book, a map, and a clock all at the same time by gluing them together. It works, but it's messy and confusing for the computer.

The New Strategy: Separate Tools, Smart Mixing

The authors propose a framework called PCAF (Prior-Conditioned Adaptive Fusion). Instead of blending everything, they build three separate paths:

The Dynamic Path (The Live Radar): This part of the model only looks at the recent history of traffic. It learns how cars move and change speed without worrying about the road layout or the time of day. It's purely reacting to what just happened.
The Spatial Prior (The Map): This is a fixed "map" of the road network. The researchers use a mathematical tool called a Laplacian spectral basis to create a stable, unchanging representation of how the roads connect. It's like a skeleton of the city that never changes, providing a constant background context.
The Temporal Prior (The Calendar): This part looks at the time of day. The model uses a special code based on the first and second "harmonics" (think of them as the main beat and the next beat of a song) to understand if it's morning rush hour or late night.

The Magic Glue: The Adaptive Fusion

Here is where the paper gets really clever. Once the model has its "live traffic" reading, its "map," and its "calendar," it doesn't just smash them together. Instead, it uses a Prior-Conditioned Adaptive Fusion mechanism.

Imagine a traffic director standing at a control panel. They have a live feed of the traffic (the dynamic part) and a set of notes from the map and the calendar (the priors).

First, the director looks at the live feed and the notes together to create a candidate prediction. This is a "what if" scenario based on the context.
Then, the director calculates a gating coefficient (a number between 0 and 1). This number decides how much to trust the live feed versus the "what if" scenario.
If the traffic is doing something weird (like a sudden accident), the director might say, "Ignore the calendar, trust the live feed!" (a high gate value).
If the traffic is calm and predictable, the director might say, "The live feed is noisy, let's lean on the map and the time of day" (a low gate value).

This allows the model to be flexible. It doesn't use a fixed rule for every situation; it adapts the mix based on the specific traffic conditions at that moment.

What the Numbers Say

The team tested their new method on four real-world traffic datasets (PEMS03, PEMS04, PEMS07, and PEMS08), which contain data from hundreds of sensors across different cities. They compared their method against several other popular models, including some that were re-run under the exact same conditions to ensure a fair fight.

The results were quite strong. The new method achieved the lowest error in 10 out of 12 different comparisons (looking at different cities and different error metrics). Specifically, it had the lowest Mean Absolute Error (MAE) on all four datasets.

Compared to a version of their model that only used the live traffic data (without the map or calendar), the new method reduced the average error by 5.37% for MAE and 6.49% for RMSE (Root Mean Squared Error).
On the PEMS03 dataset, the improvement was even more dramatic, reducing errors by nearly 9% compared to one of the top competitors.

The paper also checked how well the model worked for longer predictions (up to 8 hours into the future). As expected, the further out you look, the harder it is to predict, and the error grew. However, the model remained stable and accurate across all timeframes.

Ruling Out the "Magic"

One of the most interesting parts of the paper is what the authors ruled out. They wanted to make sure their success wasn't just because they added more math parameters or random numbers.

It's not just the math: They tested the model with "random" road maps and "shuffled" road connections. When they did this, the model got worse. This proves that the specific shape of the real road network matters, not just the fact that the model has a map.
It's not just the time: They tested using only the first "beat" of the daily cycle. It didn't work as well as using both the first and second beats. This suggests that traffic patterns are complex and need more than just a simple "morning vs. night" signal.
It's not just the gate: They compared their fancy "candidate-based" fusion to a simpler "gated" method (where you just mix the two inputs directly). Their method still won. This suggests that creating a "what if" candidate scenario before deciding how to mix it is a crucial step, not just the act of gating itself.

The Bottom Line

This paper suggests that the secret to better traffic prediction isn't just building a bigger, more complex brain. It's about organizing the information better. By keeping the live traffic, the road map, and the daily schedule separate and then using a smart, adaptive system to decide how to combine them, the model can see the future more clearly.

The model is also surprisingly efficient. It only has about 0.083 million trainable parameters (which is tiny for an AI) and can make a prediction for a whole city in about 1.10 milliseconds. This means it could easily be used in real-time systems to help traffic lights change or to guide drivers to avoid jams. While the authors note that their model uses a fixed map and might struggle with sudden, unexpected changes (like a massive accident that alters traffic flow patterns instantly), the results strongly suggest that separating these different types of information is a winning strategy for forecasting traffic.

Technical Summary: Traffic Flow Forecasting with Explicit Priors and Prior-Conditioned Adaptive Fusion

1. Problem Statement

Accurate multi-step traffic flow forecasting requires the joint modeling of three heterogeneous information sources: rapidly varying historical observations, relatively stable road-network structures, and recurring intra-day traffic patterns. Existing methods typically encode these distinct signals within a shared backbone or fuse them directly. The authors argue that this approach obscures the distinct predictive roles of each source, making it difficult to determine whether improvements stem from recent traffic evidence, stable node context, or recurring temporal patterns. Furthermore, common fusion strategies, such as direct concatenation or ordinary gated interpolation, often fail to explicitly model the interaction between dynamic evidence and structural/temporal context, or they interpolate between representations without first constructing a prior-conditioned candidate.

2. Methodology

The paper proposes a framework that explicitly separates observation-driven dynamics from structural and periodic priors before fusing them via a Prior-Conditioned Adaptive Fusion (PCAF) mechanism.

2.1 Explicit Spatiotemporal Priors

The framework utilizes three distinct processing paths:

Dynamic Observation Encoder: A shared encoder processes only historical traffic sequences to learn short-term traffic dynamics ( $H_d$ ). This path is independent of graph structure or time-of-day information.
Topology-Derived Spatial Prior ( $P_s$ ): A fixed Laplacian spectral basis is computed offline from the road network's adjacency matrix. Specifically, the normalized graph Laplacian is eigendecomposed, and the first $d_s$ eigenvectors (excluding the first) are selected, standardized, and projected into the latent space. This provides stable, node-level structural coordinates that do not change during inference.
Intra-Day Temporal Prior ( $P_t$ ): A sample-specific encoding based on the time-of-day position of the last observed step. It utilizes first- and second-harmonic sinusoidal encodings (four dimensions: $\sin, \cos$ for both harmonics) projected into the latent space. This captures recurring daily patterns.

These two priors are summed to form a unified spatiotemporal context: $P_{st} = P_s + P_t$ .

2.2 Prior-Conditioned Adaptive Fusion (PCAF)

Unlike standard gated fusion that interpolates directly between the dynamic representation ( $H_d$ ) and the context ( $P_{st}$ ), PCAF employs a two-step process:

Candidate Construction: The dynamic representation $H_d$ and the unified context $P_{st}$ are concatenated and projected to generate a prior-conditioned candidate representation ( $H_c$ ).
$H_c = \tanh([H_d \parallel P_{st}]W_c + b_c)$
Adaptive Allocation: An element-wise allocation coefficient ( $\Gamma$ ) is generated from the same concatenated input. This coefficient determines the balance between the original dynamic representation and the candidate representation:
$H_f = \Gamma \odot H_d + (1 - \Gamma) \odot H_c$
where $\Gamma = \sigma([H_d \parallel P_{st}]W_\Gamma + b_\Gamma)$ .

This design allows the model to adaptively fuse information across samples, nodes, and latent dimensions, treating the priors as conditioning signals for the dynamic evidence rather than just another input feature.

2.3 Prediction

A node-shared linear head maps the fused representation $H_f$ to the future traffic states over the forecast horizon $H$ . The model is optimized using Mean Absolute Error (MAE).

3. Key Contributions

The authors summarize their contributions as follows:

Explicit Separation: A framework that separates observation-driven dynamics, topology-derived spatial priors, and intra-day temporal priors before fusion, clarifying their distinct roles in multi-step forecasting.
Unified Context Instantiation: The use of a fixed Laplacian spectral basis and sample-specific harmonic time-of-day encoding to provide structural and temporal conditioning signals.
PCAF Mechanism: The design of Prior-Conditioned Adaptive Fusion, which constructs a candidate representation from the joint dynamic-context input before allocating information, differing from direct gated interpolation.

4. Experimental Results

The method was evaluated on four real-world datasets (PEMS03, PEMS04, PEMS07, PEMS08) using a unified controlled protocol with a 96-step input window and 12-step forecast horizon.

Performance: The proposed method achieved the lowest error in 10 of the 12 dataset–metric comparisons and the lowest MAE on all four datasets.
- Compared to a "Dynamic-only" variant (no priors), the full model reduced average MAE and RMSE by 5.37% and 6.49%, respectively.
- On PEMS03, reductions relative to the strong GeoWaveNet baseline were 8.70% (MAE), 9.83% (RMSE), and 22.33% (MAPE).
Ablation Studies:
- Component Contribution: Removing either the spatial or temporal prior increased errors, confirming they provide complementary information.
- Topology Specificity: Controls using random node vectors, permuted Laplacian features, or rewired graphs resulted in higher errors than the correct Laplacian representation, proving the gains are due to actual road-network topology rather than just additional parameters.
- Fusion Mechanism: PCAF outperformed non-gated concatenation, two-stage prior fusion, and ordinary gated fusion, validating the candidate-based fusion design.
- Harmonic Encoding: Using both first and second harmonics for time-of-day encoding yielded better results than using only the first harmonic or no temporal prior.
Efficiency: The model is compact with approximately 0.083 million trainable parameters. The fixed spectral basis is computed once offline, and the online inference time averages 1.10 ms per sample on an NVIDIA GeForce RTX 4080 SUPER.

5. Significance and Claims

The paper claims that explicitly separating observation-driven dynamics from structural and periodic priors, and integrating them conditionally via PCAF, improves both model interpretability and prediction accuracy. The results support the hypothesis that treating road topology and intra-day patterns as conditioning contexts—rather than mixing them indistinguishably in a shared backbone—allows for more effective multi-step forecasting.

The authors note that while the fixed Laplacian prior provides stable structural context, it cannot fully characterize spatial relationships that change with traffic conditions (e.g., accidents or congestion). Similarly, the temporal prior is limited to intra-day harmonic patterns. Future work is suggested to combine fixed topology with adaptive structure learning and incorporate external factors like weather, holidays, and weekly periodicity while maintaining the separation between dynamic observations and contextual priors.