Quantile-Physics Hybrid Framework for Safe-Speed Recommendation under Diverse Weather Conditions Leveraging Connected Vehicle and Road Weather Information Systems Data

This study proposes a hybrid framework combining Quantile Regression Forests with physics-based safety constraints to generate real-time safe-speed recommendations for freeways by leveraging high-resolution Connected Vehicle and Road Weather Information System data to reduce weather-related crash risks.

The Gist

Imagine you are driving down a highway. On a sunny day with dry pavement, you know exactly how fast you can go safely. But what happens when it starts pouring rain, or when a sudden snowstorm hits, turning the road into a sheet of ice? The posted speed limit sign (say, 55 mph) doesn't change. It stays the same, even though the road conditions have become dangerous. This mismatch is often why accidents happen.

This paper proposes a smart solution: a digital co-pilot that doesn't just tell you the legal limit, but calculates the actual safe speed for the exact moment you are driving, based on the weather and the road's grip.

Here is a breakdown of how this "Smart Speed Co-pilot" works, using simple analogies.

1. The Ingredients: A Massive Digital Kitchen

To build this system, the researchers needed a lot of fresh ingredients. They gathered data from three main sources:

• Connected Vehicles (The "Crowd"): They used data from thousands of cars driving around Buffalo, NY. Think of these cars as thousands of eyes and ears on the road, constantly reporting how fast everyone is actually driving.
• Road Weather Stations (The "Sensors"): They used special roadside sensors that measure things like how slippery the road is (grip), how far you can see (visibility), and the temperature.
• Digital Maps (The "Blueprint"): They used a digital map of the roads to know exactly where the highways are and what the legal speed limit is.

They cooked up a massive dataset containing over 6.6 million records from 73 days, covering everything from sunny days to heavy snowstorms.

2. The Brain: The "Weather-Savvy" Predictor

The core of the system is an AI model called Quantile Regression Forests (QRF).

• The Analogy: Imagine a weather forecaster who doesn't just say, "It will rain at 3 PM." Instead, they say, "There's a 25% chance the rain will be light, a 50% chance it will be moderate, and a 75% chance it will be heavy."
• How it works: Instead of guessing a single speed (like "drive at 45 mph"), this AI looks at the current weather and road grip and predicts a range of speeds. It tells you, "Based on what other drivers are doing right now, the middle 50% of people are driving between 42 and 48 mph." This captures the natural variation in traffic.

3. The Safety Net: The "Physics Brake"

Here is the clever part. The AI is great at guessing what people are doing, but sometimes people drive too fast for the conditions. To fix this, the researchers added a Physics-Based Safety Rule.

• The Analogy: Imagine you are walking on a icy sidewalk. Your brain (the AI) might see everyone else walking fast, but your legs (the Physics Rule) know that if you run, you will slip and fall. So, your legs say, "No matter what everyone else is doing, you must slow down to a speed where you can stop before hitting a wall."
• How it works: The system calculates the Stopping Sight Distance. It asks: "If the road is wet and visibility is low, how fast can a car go and still stop safely within the distance the driver can see?" This creates a "Hard Ceiling" for the speed. Even if the AI thinks people are driving at 50 mph, if the physics says you can't stop safely above 35 mph, the system caps the recommendation at 35 mph.

4. The Final Recipe: The "Safe Speed Interval"

The system combines the AI's prediction (what's normal) with the Physics limit (what's safe) and the Legal Limit (the sign on the pole).

• It creates a recommended speed range, like [38 mph, 42 mph].
• If the weather is perfect, the range might be close to the legal 55 mph.
• If it's a blizzard, the range might drop to [25 mph, 30 mph].

5. Did It Work? (The Taste Test)

The researchers tested this system on real highways in Buffalo.

• Accuracy: The system was incredibly accurate. When it predicted the "middle" speed, it was usually within 1.5 mph of the actual average speed of cars.
• Reliability: In 96% of cases, the predicted speed was within 5 mph of the real speed.
• Comparison: They compared it to old methods (like just saying "drive 10% slower than the sign") and found their new method was much smarter. It didn't just guess; it learned from the weather and the road conditions.

Why This Matters

Think of this system as a dynamic speed limit sign that updates every 10 minutes.

• For Drivers: It gives you a realistic, safe speed to aim for, reducing the panic of "Am I going too fast?" in bad weather.
• For Safety: By keeping cars moving at a speed that matches the road's ability to stop them, it prevents the "chain reaction" crashes that happen when one car brakes hard on ice and others can't stop in time.

In short, this paper presents a way to turn static, unchanging speed limits into a living, breathing safety guide that adapts to the weather, the road, and the traffic, all to keep us safer on the road.

Technical Summary

1. Problem Statement

Traffic safety is severely compromised during inclement weather due to reduced visibility and diminished tire-road friction. Current static posted speed limits fail to account for real-time variations in environmental and roadway conditions, often leading to a mismatch between legal limits and actual safe driving speeds. While Variable Speed Limits (VSLs) exist, they face two primary technical challenges:

• Data-Driven Models: Many machine learning approaches (e.g., neural networks) act as "black boxes," producing point forecasts that lack physical interpretability. They may recommend speeds that are physically unattainable (unsafe) or overly conservative.
• Physics-Only Models: Purely physics-based approaches often ignore the complex, nonlinear interactions between traffic flow, weather, and driver behavior.
• Gap: There is a lack of frameworks that successfully integrate probabilistic data-driven predictions with mandatory physical safety constraints (specifically stopping sight distance) to generate adaptive, safe speed intervals.

2. Methodology

The authors propose a Hybrid Quantile-Physics Framework that fuses data-driven probabilistic modeling with physics-based safety constraints.

A. Data Sources & Preprocessing

The study utilizes a spatiotemporally aligned dataset from the Buffalo-Niagara metropolitan area (73 days, 2022–2023) containing 6.6 million records:

• Connected Vehicle (CV) Data: High-frequency (1–5s) GPS and speed data from Wejo, covering ~5–10% of traffic.
• Road Weather Information System (RWIS): 10-minute interval data from a Vaisala station, including surface temperature, grip level, visibility, and precipitation history.
• OpenStreetMap (OSM): Road network geometry and posted speed limits.
• Processing: CV data was filtered to freeway segments (55 mph zones) using OSM geometry buffers. Data was aggregated into 10-minute windows, calculating average vehicle speeds and aligning them with corresponding RWIS weather conditions.

B. Core Modeling: Quantile Regression Forests (QRF)

Instead of predicting a single mean speed, the framework uses Quantile Regression Forests (QRF) to estimate the conditional distribution of vehicle speeds.

• Inputs: 26 features including meteorological variables (visibility, grip, temp), pavement conditions, and temporal context (hour, day).
• Outputs: The model predicts the 0.25, 0.50 (median), and 0.75 quantiles of the speed distribution for each 10-minute window. This provides a statistically calibrated prediction interval ($Q_{0.25} - Q_{0.75}$) capturing uncertainty.

C. Physics-Based Safety Constraint

To ensure the recommended upper speed limit is physically safe, a Stopping Sight Distance (SSD) constraint is applied:

• Formula: The total stopping distance is modeled as the sum of braking distance, reaction distance, and a safety headway buffer.
• Constraint: The recommended speed must allow a vehicle to stop completely within the driver's visible range ($d_{visible}$), which is capped by RWIS visibility data and a maximum AASHTO standard (495 ft for 55 mph).
• Calculation: A physics-derived upper bound ($v_{phys}$) is calculated numerically for every 10-minute interval based on real-time grip and visibility.

D. Interval Fusion Strategy

The final recommended speed interval is a fusion of the statistical and physical models:

• Upper Bound: The minimum of the predicted 0.75 quantile ($Q_{0.75}$), the physics-based limit ($v_{phys}$), and the posted speed limit ($v_{posted}$).
• Lower Bound: The minimum of the predicted 0.25 quantile ($Q_{0.25}$) and the calculated upper bound.
• Result: An adaptive interval $[v_{low}, v_{high}]$ that is centered on observed traffic behavior but strictly capped by physical safety limits.

3. Key Contributions

1. Hybrid Framework: First integration of QRF (for capturing traffic/weather variability) with explicit SSD physics constraints (for safety guarantees), bridging the gap between "black box" AI and interpretable engineering.
2. High-Resolution Multi-Source Data: Creation of a massive, aligned dataset (6.6M+ records) combining CV trajectory data with granular RWIS pavement and weather data.
3. Probabilistic Safety: Moving beyond point estimates to provide safe speed intervals that are statistically calibrated (PICP) and physically feasible.
4. Generalization: Demonstration of model robustness across diverse weather conditions (clear, rain, snow) and different freeway segments without overfitting.

4. Experimental Results

The model was evaluated on a held-out test set using metrics for interval coverage, width, and point accuracy.

• Predictive Accuracy:

  • MAE (Mean Absolute Error): 1.55 mph for median speed predictions.
  • Threshold Accuracy: 96.43% of predictions fell within ±5 mph of the ground truth.
  • Generalization: The model maintained performance across weather types, with accuracy dropping only slightly in snow (88.24% within ±5 mph) compared to clear weather (97.35%).

• Interval Calibration:

  • PICP (Prediction Interval Coverage Probability): The 50% prediction interval achieved 48.55%, nearly matching the ideal 50% target, indicating excellent calibration.
  • MPIW (Mean Prediction Interval Width): Stable at ~18 mph, showing the model does not unnecessarily widen intervals under uncertainty.

• Baseline Comparisons:

  • vs. Static Rules (Posted ±10%): The hybrid model significantly outperformed fixed limits, which only captured 34.19% of actual speeds.
  • vs. Rolling IQR (Historical): The QRF model achieved similar interval calibration but better point accuracy (MAE 1.55 vs. 1.18 for short-window Rolling IQR, though Rolling IQR is more sensitive to noise).
  • vs. NGBoost: While NGBoost had similar interval calibration, its point prediction error (MAE) was double that of QRF (3.21 vs. 1.55 mph).

• Case Study: Applied to 6 major freeways in Buffalo (I-90, I-190, etc.) during two testing periods. The model successfully detected speed reductions during specific adverse weather events (e.g., low visibility on Oct 19 and Mar 23) and adjusted the physics-based upper bound accordingly, while the static 55 mph limit remained unchanged.

5. Significance

This study demonstrates the feasibility of deploying real-time, weather-aware, and physics-constrained speed guidance systems.

• Safety Impact: By ensuring recommended speeds never exceed the physical limit of stopping within sight distance, the framework directly addresses the root cause of weather-related crashes.
• Operational Utility: The system provides actionable, adaptive intervals rather than static limits, offering a more nuanced approach to traffic management that balances efficiency (staying close to observed speeds) with safety.
• Scalability: The use of standard CV and RWIS data suggests this framework can be deployed in other regions with similar infrastructure, advancing Intelligent Transportation Systems (ITS) capabilities.