Optimizing winter traffic forecasting through spatially transferable models in cold regions: Insights for infrastructure management

Transportation agencies in cold regions need winter-robust traffic forecasts that can transfer across road types when monitoring sites are sparse. Using five Alberta WIM sites (FMD for model development; LED, VID, RDD, LVD for transfer testing) and winter seasons (Nov–Mar, 2005–2009), we predict the day-ahead daily volume factor (DVF) for three vehicle classes (total, passenger cars, trucks) with four model structures: Winter-weather (Ww), Naïve (Na), Base (Ba), and Para (Pa). The Ww model combines an expected daily volume factor (EDVF), continuous snowfall, and temperature-category dummies. Spatial transfer tests show high accuracy across functionally similar and distinct highways. Representative gains versus baseline structures include LVD-Ww (total traffic) MAPE 6.23 % vs Ba 7.77 % (19.8 % reduction), and LED-Ww (total traffic) 5.46 % vs Ba 6.19 % (11.8 % reduction), with R²≈ 0.994–0.996. Class-specific patterns emerge: trucks often favor simpler structures (e.g., LVD-Na 4.88 % vs Ww 5.24 %-6.9 % reduction). Results indicate that careful model choice by vehicle class and road function enables spatially transferable winter forecasts without deploying additional WIM sites, supporting resource-efficient maintenance, operations, and traveler information.