Why Statistically-Based Road Weather Modeling Supersedes Physical Modeling

Every year, state and local governments spend billions of dollars on snow removal during the winter season. In the early 2000s, the federal prototype Maintenance Decision Support System (MDSS) was developed by the National Center for Atmospheric Research (NCAR) with funds from the United States Department of Transportation (USDOT) Federal Highway Administration.

Today, government agencies (international, state and local) have adopted this technology and have implemented MDSS within their year-round road maintenance and wintertime snow-removal operations. MDSS is an example of government funds being used in a unique and useful way, and its adoption across many of the snow-belt states in a relatively short period of time is remarkable.

While states that have implemented MDSS across winter maintenance operations have become more effective and efficient as a result, new road weather forecasting models are now being developed. Here we explain why statistically-based road weather modeling is more scalable and accurate for MDSS than traditional physical modeling, enabling state and local governments to make significant savings on snow removal.

 

The Current Data and Compute-Intensive Model

The core of any MDSS (Maintenance Decision Support System) is the road forecast weather modeling toolset and infrastructure. Forecast modeling of pavement temperature and condition (dry, wet, snow, etc.) are typically generated through a set of physical-based algorithms that are driven by the composition and thermal characteristics of the road. The algorithms model heat exchange from heat stored in the road-base (sub-surface) to the above surface air temperature measured by a RWIS (Road Weather Information System) or estimated by the forecast. These physical models are also highly reliant on road type (asphalt, concrete and respective road-base composition) and require bridge and highway overpass coordinates to improve forecast results. This modeling approach is extremely data and compute-intensive, using tools originally designed for RWIS datasets (low volume).

This system generates tuned forecasts for those RWIS sites and consistently produces good results for near real-time and day-ahead planning. However, physical models are data challenged when scaled to high spatial and temporal resolution applications (sub-kilometer along any road segment on the globe at 15 minute forecast resolution), resulting in false positives and higher forecast error over the forecast time series.

Advanced forecast methods are quickly emerging to solve this data and compute challenge by tapping into connected vehicle data and other sensor/video technologies, thereby allowing state and local agencies to provide MDSS services along any road segment without the need for higher cost RWIS stations. There is still value in managing and operating MDSS services despite the scalability issues. However, public service agencies are missing out on maximizing the potential value from the incremental investment into higher performing technologies.

 

A More Accurate, Highly-Scalable Model

Since MDSS was first introduced at NCAR (the National Center for Atmospheric Research) in the early 2000s, the vertical compute platform (supercomputer) has fully transformed into highly distributed cloud service architecture with Graphics Processing Unit (GPU) enabled compute power that revolutionizes big data management and deep learning analytics. These new tools are significant enablers for new advanced forecast engines that exploit edge services and infrastructure to meet the demands of real-time systems (automated services) and situational awareness of road surface conditions for public safety alerts.

Agencies that have adopted MDSS over the past decade have not realized the value of these emerging technologies, as many of the original MDSS vendors have not adapted their systems to ingest big data. Scalable, statistically-based road weather forecast systems are now available for MDSS deployments (as well as traveler information or 511 alert systems) to provide more accurate and highly-scalable road weather information to end-users (sub-kilometer road state forecasts and alerts).

Statistical modeling within the forecast engine relies on a smaller dataset (ground truth, national forecast models) for model training (bias correction), which enables the model to generate more consistent results over the forecast time series. Physical models are data driven. They require more raw input for training on a specific forecast point (road type, road base, heat transfer, shade, UV), which yields slower processing cycles and limits scale across cloud resources (regional distribution and geo-redundant services).

Although the physical model will provide a more accurate forecast at one specific point, it loses its value as the user scales across thousands or millions of points along a road segment. The statistical model consistently outperforms the physical model in this use case. Advanced machine learning and model training techniques are readily available that produce high-resolution, scalable road weather forecasts and eliminate the guesswork between RWIS locations, allowing for much more efficient use of resources and materials when adverse weather affects operations and public safety.

 

WeatherCloud: A Statistically-Based Road Weather Solution

Fathym’s WeatherCloud solution offers a statistically-based, machine learning road weather forecast that is dynamically-generated, globally-scaled, and enables sub-kilometer forecasts, alerts and weather risk variables, such as delay and slick roads. Route and geographic point forecasts are available via API. 

 

We have produced a free, downloadable White Paper on road weather solutions and their impact on transportation, maintenance and driver safety. Download it now to learn in depth how next generation road weather solutions are saving time, money, and most importantly, lives.