Software System for Road Condition Forecast Correction

Dmitrii Smolyakov, Evgeny BurnaevarXiv 2020 · OpenReview

Abstract: In this paper, we present a monitoring system that allows increasing road safety by predicting ice formation. The system consists of a network of road weather stations and an intelligent data processing program module. The results were achieved by combining physical models for forecasting road conditions based on measurements from stations and machine learning models for detecting incorrect data and forecast correction.

Introduction

Predicting ice formation is a critical, yet difficult, aspect of road safety. Traditional approaches generally fall into two categories: physical models and machine learning (ML) models. Physical models use complex partial derivative equations to simulate heat exchange (energy balance). While interpretable, they are computationally expensive, deterministic, and struggle to account for non-deterministic factors like heat generated by traffic. Conversely, ML models excel at pattern recognition but often lack physical interpretability and require constant refitting for new environments.

Furthermore, both methods rely on Road Weather Information Systems (RWIS)—networks of sensors that frequently malfunction. Feeding faulty data from a broken sensor into a forecast model leads to dangerous or costly errors. This research proposes a hybrid system that combines the strengths of both modeling worlds while automatically filtering faulty sensor data.

Methodology

The proposed software pipeline utilizes a multi-stage approach to improve forecasting accuracy:

  1. Physical Core (METRo): The "Model of the Environment and Temperature of Road" (METRo) serves as the baseline. It decomposes temperature changes into radiation flux, turbulence, and latent heat.
  2. Residual Correction via Gradient Boosting: Instead of replacing the physical model, Gradient Boosting on Decision Trees (using the lightgbm library) is used to predict the residual (error) of the METRo model. The final forecast is the sum of the physical model's output and the ML-predicted correction.
  3. Feature Engineering: Temporal data is encoded using trigonometric functions (sine and cosine) to capture the cyclic nature of time (e.g., ensuring the model understands that Hour 23 is adjacent to Hour 0).
  4. Synthetic Anomaly Detection: To identify broken sensors, a detector was developed to compare predicted values against actual readings. Due to the rarity of real-world failure labels, the researchers artificially generated three types of anomalies—single outliers, short-term Poisson noise, and long-term Gaussian noise—to tune the detection threshold.

Datasets

The study utilized seven years of data (2012–2019) from 120 RWIS stations across Russia. The data was split into training (35 stations), threshold tuning (15 stations), and final testing (9 stations). For the test set, data was manually labeled as "normal" or "anomaly" to verify detector performance.

Key Results

The hybrid approach consistently outperformed standalone models across 1, 2, and 3-hour forecast horizons. As shown in the performance metrics, the Mean Absolute Error (MAE) for road temperature at a 1-hour horizon dropped from 0.84°C using METRo alone to 0.48°C with the hybrid correction.

Mean Absolute Error (MAE) comparison across different models and forecast horizons.

The anomaly detection module also showed superior performance. The proposed Gradient Boosting detector achieved an $F_1$-score of 0.774, significantly outperforming standard industry techniques such as Isolation Forest (0.308) and One-Class SVM (0.227).

Performance metrics for various anomaly detection algorithms.

Conclusion

The most significant insight from this work is that ML is most effective when used to "mop up" the non-deterministic errors that physical equations cannot easily capture, rather than replacing them entirely. Additionally, the use of synthetic anomalies proved to be a robust solution for the "small data" problem common in industrial IoT failure detection. Future work will focus on refining the decision-making module to better optimize road maintenance costs, such as reducing unnecessary salting or sanding based on these high-accuracy forecasts.

Summary updated

Abstract

In this paper, we present a monitoring system that allows increasing road safety by predicting ice formation. The system consists of a network of road weather stations and intelligence data processing program module. The results were achieved by combining physical models for forecasting road conditions based on measurements from stations and machine learning models for detecting incorrect data and forecast correction.
BibTeX
@misc{smolyakov2020software,
  author = {Dmitrii Smolyakov and Evgeny Burnaev},
  title = {{Software System for Road Condition Forecast Correction}},
  year = {2020},
  eprint = {2003.09957},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2003.09957},
}
Figure