Road Weather Forecasts in Norway with the METRo Model

Abstract

We present a model evaluation of road weather forecasts in Norway with the METRo model in a quasi-operational setting. The road weather forecasts are initialized with measurements made by road weather stations and driven by mesoscale weather forecast data from the Norwegian Meteorological Institute. One important source of hazardous driving conditions in Norway are freezing road-surface temperatures. We quantify the skill of our model setup to predict such conditions by computing the hit rates and false-alarm rates for incidences of freezing temperatures, relative to the climatological rates of occurrence. The METRo forecasts consistently add skill in wintertime and the crucial transitional seasons of spring and fall. Our study illustrates a successful proof-of-concept for novel, operational road weather forecasts in Norway, that could easily be realized with an open-source prediction model and readily available input data.

1. Introduction

Road weather forecasting is a special application of weather forecasting, which aims to predict meteorological conditions on a road network [1]. The most important variables in road weather forecasts consist of the road-surface temperature, near-surface temperature, near-surface humidity, and the presence of precipitation and water on the road. Hazardous driving conditions, such as icy roads, can be identified and predicted based on these variables, and ultimately applied for traffic safety, road maintenance and services, and driver warning systems. In Norway, and Nordic countries in general, these applications are particularly important during winter where half of the yearly 27,000 road and 50,000 pedestrian injuries in Finland, Norway, and Sweden can be traced back to slippery road and walkway conditions [2]. In addition, such forecasts will become even more important with the ongoing implementation of driving assistance systems in autonomously driving cars [3].

State-of-the-art road weather forecast systems use the output from numerical weather prediction models to compute the present and near-future road weather, from a few minutes to a few days ahead [4]. During the last few decades, several road weather models (RWMs) were developed and are now used operationally by weather agencies and road-service authorities. Two prominent European examples are the Royal Netherlands Meteorological Institute’s RWM [5] and the Finnish Meteorological Institute’s RWM [6], which both predict road-surface temperatures and surface conditions such as road friction. Their performance and forecast quality were compared in a recent study [4]. Both models are one-dimensional heat balance models that use predicted meteorological surface variables from mesoscale numerical weather prediction models, interpolated onto specific road points. These models differ mainly in the forecast initialization procedure, which is explained in detail in [4,7,8].

METRo (Model of the Environment and Temperature of Roads) is another widely used road weather forecast model, which has been developed in Canada in the late 1990s [8]. In 2006, the METRo code was publicly released under the GPL license, and since then, the model has been adapted for operational use in several countries, for example, in the Czech Republic (METRo-CZ) [9]. The authors evaluated the model with data from 25 road weather stations, demonstrating the crucial role of the radiation flux forecast for the prediction of road-surface temperatures. Later, Sokol et al. [10] presented an ensemble technique for road-surface temperature forecasting, which allows for the estimation of forecast uncertainty and probabilistic forecasts. The METRo model was also used in the work of Djemal et al. [11] for surface temperature forecasts at a military airport in Praha. In addition, the METRo model was integrated into a winter maintenance weather forecast system that covers Finland, Sweden, and Russia [12] as well as the United States [13]. For the 2022 winter Olympic demonstration station (Beijing Huilongguan station), road-surface temperature prediction was tested based on the METRo model [14].

In Norway, the platform Vegvær [15] is also built on the METRo model and is seen as the current benchmark for localized road weather forecasting. Vegvær’s adaptive forecasting methods have been utilized to provide actionable insights for road maintenance teams. Comparable systems, such as the Finnish Meteorological Institute’s RWM [6] and the UK Met Office’s OpenRoad [16], have similarly demonstrated the effectiveness of such solutions.

Recent studies have highlighted the growing role of advanced numerical weather models and ensemble-based approaches in improving road weather predictions. For instance, Kangas et al. [17] demonstrated the effectiveness of machine learning techniques in refining predictions of freezing road conditions across Northern Europe. Similarly, a 2020 study by Carillo et al. [18] showcased the integration of localized sensor and camera networks with weather data to enhance the accuracy of road-surface temperature forecasts in Finland and Sweden. These developments underscore the potential of using multi-source data fusion for road weather modeling.

These studies also emphasize the importance of uncertainty quantification in road weather forecasting. Advances in ensemble forecasting techniques, such as those implemented by Sokol et al. [19], allow for more reliable predictions by accounting for variations in input data and model physics. Other studies explored the role of vehicle-mounted [20,21] and mobile [22] sensors in enhancing real-time data acquisition for road weather models, offering a promising avenue for improving prediction accuracy in dynamic environments.

With Vegvær, the METRo has only been recently introduced for road weather forecasts in Norway. The heterogeneous topography of Norway, with its long coastline and significant south–north span, presents a challenge for numerical weather prediction. Hence, we analyzed the applicability of the METRo model for operational road weather forecasting in Norway by initializing the model with road weather observations from the Norwegian road authority (Statens Vegvesen, SVV). In conjunction with the operational mesoscale forecast from the Norwegian Meteorological Institute [23], we quantify the effectiveness in predicting hazardous conditions, such as freezing road-surface temperatures. Our approach builds on recent advancements exemplified by Vegvær and similar platforms, contributing to the development of a robust framework for highly accurate and localized road weather predictions.

2. METRo Overview

2.1. Physics Module

METRo was developed by the Canadian weather agency in 1999 [8] for operational road weather forecasts up to 24 h in advance at specified, fixed positions within a road network. Its main component is a surface energy balance module, estimating the present and predicting future road-surface temperatures by calculating surface energy fluxes. The energy fluxes taken into consideration are shortwave and longwave radiation, latent and sensible heat flux, and anthropogenic heat flux (for example, friction between the road-surface and passing cars).

METRo has two further modules that allow for a more realistic physical modeling of the meteorological conditions at the road level. The first module is a vertical heat diffusion module, which simulates one-dimensional heat transfer between the surface, different road layers, and the deep soil. For this module, the vertical build-up of the road can be specified in terms of layers of different materials (such as asphalt, concrete, or rock) and depth. The second module tracks the accumulation of water on the road-surface, as a result of precipitation, phase changes (freezing and thawing), and water run-off. There are separate reservoirs for liquid and frozen water, allowing for transfer from one reservoir to the other.

2.2. Input and Output Variables

To initialize the METRo model and its physics modules, initial data such as surface, sub-surface, and near-surface air temperatures, humidity, precipitation, and wind speed are required (see Section 3.1 for details). In our setup, these initial conditions are derived from in situ measurements at SVV road weather stations (see Section 4 for details).

During the model simulation, we need to prescribe atmospheric boundary conditions that describe future external forcing (see Section 3.2 for details). These boundary conditions include variables like wind speed, air temperature, humidity, cloud cover, and precipitation. Variables such as the road-surface temperature are, at this stage, not externally prescribed any longer, but calculated by METRo.

To apply bias correction to road-surface temperatures in the preprocessing phase, METRo requires some forecast data to overlap in time with the observational data. In our configuration, the forecast data overlap in time with the observation data for the previous 4 h, as illustrated in Figure 1. For example, if the desired road forecast is for the next 24 h, forecast data starting from 4 h before the initial time have to be used.

The model output contains all variables needed for initialization and forcing: surface energy fluxes, accumulated water on the road, and hazard warnings that are derived from meteorological conditions, such as freezing rain or melting snow.

3. Technical Implementation and Data Flow

The road weather station data (observations) are obtained from the Frost API (Application Programming Interface), which is maintained by the Norwegian Meteorological Institute (MET). The data can be requested internally through an API. The forecast data can be requested by the internal Forecast API, which downloads the weather forecasts from the Thredds API (also maintained by MET) and converts the parameters to those needed by METRo. Finally, the model API uses the METRo model and data from the stations and Forecast API to calculate the road weather forecast. A general overview of the system and data flow is illustrated in Figure 2.

3.1. Observational Data

We used observations of the last 4 h from SVV road weather stations to initialize the METRo model. Observations were obtained from the Frost API, which is maintained by MET. During data acquisition, the model API is provided with data types, site information, and observation data. The observational data are scraped and stored in a database that is processed to match the format expected by the METRo model.

Sometimes, the observational time series are incomplete, especially for subsurface temperature, which currently is available at only 12 stations. For these cases, we use the following analytic estimations for missing values:

• Subsurface temperature = Surface temperature minus 2 °C;
• Surface temperature = Two-meter air temperature;
• Road condition = Dry, except if precipitation rate > 0.5 mm in the last 2 h, yielding initialization as wet.

3.2. Forecast Data

The Forecast API provides the weather forecast to the model API, which is used as initial and boundary conditions for the METRo model. The list of forecast parameters used are described in

Table 1. Forecast API Forecast parameters.

Parameter	Description	Unit
altitude	height above sea level	m
air_temperature_2 m	screen level temperature	K
relative_humidity_2 m	screen level relative humidity	Normalized (0–1)
wind_speed_10 m	wind speed at 10 m height	m/s
cloud_area_fraction	total cloud cover	Normalized (0–1)
air_pressure_at_sea_level	air pressure at sea level	hPa
precipitation_amount	accumulated total precipitation	kg/m2

.

The list of parameters used by the METRo model can be found on

Table 2. METRo Forecast parameters.

Parameter Name	Parameter Description	Convertion Description	Availability in THREDDS	Unit
at	Air temperature at 1.5 m	present in THREDDS as air temperature at 2 m	Present	K
td	Dew point at 1.5 m	derived from relative humidity	Derived	K
ws	Wind speed at 10 m	wind speed at 10 m height	Present	m/s
cc	Octal cloud coverage	present in THREDDS as fractional cloud cover	Present	Normalized (0–1)
ap	Surface pressure at station altitude	calculated as explained in Table 1	Not Present	hPa
ra	Rain precipitation quantity	derived from total precipitation	Not Present	kg/m2
sn	Snow precipitation quantity	derived from total precipitation	Not Present	kg/m2

.

The conversion of variables from THREDDS API to METRo is performed as described in

Table 3. Parameter conversion between THREDDS and METRo.

METRo Parameter Name	Unit	Used THREDDS Parameters	Unit	Conversion Strategy
forecast time	ISO 8601	time	s	Javascript function that returns a string in ISO format
at	°C	air_temperature_2 m (T)	K	$at=T-273.15$
td	°C	dewpoint_temperature_2 m (T), relative_humidity_2 m (RH)	K	$td=T-273.15-\left({\displaystyle \frac{100-RH·100}{5.0}}\right)$
ws	km/h	wind_speed_10 m	m/s	$ws=$ wind_speed_10 m $· 3.6$
cc	octal (0–8)	cloud_area_fraction (C)	0–1	$cc=C·8$
ap	mbar	air_pressure_at_sea_level (P0), altitude (h)	hPa, m	$ap=P0·{\left[1-{\displaystyle \frac{0.0065·h}{at+0.0065·h+273.15}}\right]}^{5.257}$
ra	mm	precipitation_amount (prec)	kg/m2	if at > 0: ra = prec
sn	cm	precipitation_amount (prec)	kg/m2	if at ≤ 0: sn = prec

.

Given that the data from THREDDS API are spatially discretized in a horizontal grid spacing of 1 × 1 km, we use the following steps to estimate the values of the parameters at the locations in METRo:

• Grid identification: Based on the input coordinate (red dot P in Figure 3), the algorithm identifies the closest coordinates of the horizontal grid of the THREDDS API data surrounding the red dot (v1, v2, v3, v4) to create a polygon.
• Smallest triangle: The algorithm creates the vertices and finds the triangle that has the minimum distance of the point P to its vertices.
• Barycentric interpolation: We use the barycentric interpolation [24] to interpolate the vertex data across the surface of the triangle. Using the three sub-areas (u, v, and w), the algorithm calculates the sum of the multiplication of the vertices with the sub-areas.

3.3. Model API

The model API receives the data from the Forecast API and station API and runs them through the METRo road weather model yielding the road weather forecast as output. Detailed information about the input and output of the METRo model is given in the model documentation on the METRo homepage (https://framagit.org/metroprojects/metro, accessed on 15 April 2025) and by Crevier and Delage [8]. The model API requests data from the other APIs and prepares an input file to the METRo model that consists of a timestamped list of observation and forecast data for all the parameters. The main output that we present here is the road-surface temperature (st). Another important output parameter of the METRo model is the road condition (rc), indicating if the road conditions will be dry, wet, ice/snow, mix water/snow, dew, melting snow, frost, or icy rain.

4. Experiment Design

Using the model setup as outlined above, our experiment design aims to recreate the conditions of a routine road weather service based on our METRo configuration. We initialized 24 h road weather forecasts twice a day, at 4 and 16 local time (LT) at the locations of 27 SVV road weather stations over the course of the year 2021. Given that we performed these forecasts retrospectively, we can use SVV road weather measurements to evaluate the model performance (see Section 4).

The 27 SVV road weather stations used in our study take measurements of near-surface temperature and humidity, road-surface temperature, and precipitation. The geographical coordinates of each station, station ID, and altitude above sea level are listed in

Table 4. Station information.

Station ID	Long Name	Latitude	Longitude	Altitude [m]
SN27785	E18 Rødbøl	59.107	10.1062	83.7
SN53280	E16 Flåm	60.861	7.1048	6.1
SN63595	Rv70 Gråura	62.5818	9.2136	463.8
SN79215	E6 Yttervika	66.2298	13.8728	9.8
SN16611	E6 Fokstugu	62.1125	9.28625	976.8
SN16620	E6 Avsjøen	62.1807	9.4752	929.8
SN27055	Fv312 Hanekleiva	59.5735	10.181	75.1
SN27075	E18 Grelland	59.511	10.2078	92
SN27285	E18 Gulli Nord	59.3183	10.3748	26.4
SN27730	E18 Fokserød	59.1828	10.2077	104.8
SN50815	E39 Vågsbotn	60.4768	5.348	91.1
SN6700	Rv3 Svingen	60.9573	11.4902	202
SN67153	E39 Øysand	63.3237	10.2438	13.9
SN68175	E6 Moholtlia	63.4078	10.4388	118
SN79791	E6 Saltfjellet	66.5548	15.3235	673.9
SN84770	E6 Hålogalandsbrua	68.4683	17.482	42.1
SN84905	E10 Bjørnfjell	68.436	18.1035	502.2
SN91490	E8 Bossovarri	69.1185	20.7455	544.8
SN12280	Rv3 Stabekken	60.8165	11.2853	216.1
SN27320	E18 Hem Nord	59.3535	10.3825	56.4
SN49860	Rv7 Dyranut	60.3693	7.4945	1232.9
SN52390	E39 Ostereidet	60.626	5.4722	96.6
SN61580	E136 Brustuglia	62.2943	8.1255	452.1
SN82210	Rv80 Bertnes	67.2878	14.5932	26
SN84910	E10 Skitdalshøgda	68.516	17.8725	398.9
SN6690	Rv3 Ebru	60.8632	11.4172	269
SN91420	E8 Halsebakkan Nedre	69.2852	20.4657	153.9

. Not all stations provided usable measurements at every hour of the year, with measurement time series regularly containing missing values. We applied a conservative selection for the road weather forecasts used in our evaluation and rejected all forecasts with any missing values in the reference measurement time series. This leaves us with a total of 9586 road weather forecasts for all stations during the 24 h forecast. We checked that every station in Table 4 contributes with an almost equal amount of forecasts and removed any outliers from the study, like physically unrealistic values, such as road-surface temperatures outside a feasible range, e.g., below −50 °C or above 70 °C, or as forecasts with sudden, unexplained jumps in the time series data.

5. Results

5.1. Full-Year Evaluation Scores

The evaluation scores presented in this section are based on a comparison between METRo model output and SVV road weather station measurements. All scores are computed as a function of forecast lead time up to 24 h. We calculated the mean absolute error (MAE) and bias for road-surface temperature, near-surface air temperature, near-surface dew point, and wind speed. The bias (MAE) at a given lead time is defined as the (absolute) value of the difference between model output and the observed value, averaged over all METRo forecasts. The MAE provides us with an intuitive measure of how much our METRo forecasts deviate from the actually observed road weather on average, whereas the bias allows us to assess if these deviations are caused by systematical issues arising from our METRo configuration.

In our validation, we see the prediction of road-surface temperature (RST) as the most important METRo output, as many hazardous driving conditions in Norway are directly related to freezing RSTs, such as freezing precipitation as well as ice and snow on the road. When looking at scores of wind speed, air temperature, and dew point, it is important to remember that these variables in METRo are prescribed by the boundary conditions. With a validation of these variables, we are basically comparing the weather forecast data that we use as boundary conditions with the SVV road weather measurements.

The MAE of RST increases with forecast lead time with a considerable portion of the MAE caused by a positive bias of 1–2 °C (Figure 4a). There are local maxima in both the bias and MAE at lead times of 8 and 20 h into the forecast, corresponding to 12 a.m./p.m., as our METRo forecasts are initialized at 4 and 16 LT. This finding motivates us to look at the scores of RST for all forecasts initialized at 4 and 16 LT separately (red and blue lines in Figure 4a). We can see that the maxima in RST bias and MAE are caused by a positive temperature bias during daytime, while the values of bias and MAE around midnight are rather low. This daytime bias could indicate an issue related to the surface energy balance, with solar irradiance being a potential error source. Possible sources of error include local shadow effects, an incorrect surface albedo, or errors in the cloud coverage. However, other processes such as surface heat fluxes or errors in long-wave radiation could also play a role, potentially causing insufficient cooling at the surface. Further investigation would be required to determine the dominant source of this bias.

The near-surface air temperature has a negative bias (Figure 4b). This bias, however, is rather small, consistently being less than −0.5 °C. The MAE of air temperature is not that large either, with a mean of 1.51 °C over all lead times. Similar to those of air temperature, the dew point also has a small negative bias and a rather low MAE (Figure 4c). The error growth for both variables is already saturated after 5–10 h of lead time.

The wind speed MAE is clearly dominated by a positive bias of 1.05 m/s, consistent over all lead times after a spinup of around 5 h (Figure 4d). This positive bias is most likely associated with the modeled wind being at 10 m, whereas most SVV road weather stations measure wind speed at a height of 2 m above ground. Assuming a logarithmic wind profile close to the ground will automatically yield the observed bias.

Even more important than the MAE and bias is the model’s accuracy under certain conditions, for example, as a tool to predict hazardous driving conditions associated with freezing surface temperatures, i.e., RST ≤ 0 °C. For this purpose, we computed the model’s hit rate and false-alarm rate for the prediction of RST being ≤ 0 °C. Here, the hit rate (HR) is defined as the ratio of all cases in which both the observed and the modeled RST are ≤0 °C. The miss rate (MR), accordingly, is the ratio of all cases in which the observed RST is ≤0 °C, but where METRo is >0 °C (MR = 1 − HR). We also calculated the false-alarm rate (FAR), which is the ratio of all cases in which the observed RST is >0 °C, while that in METRo is ≤0 °C. Accordingly, one would find the correct rejection rate (CRR) in which both the observed and modeled RST are >0°C (CRR = 1 − FAR).

A well-performing road forecasting system is characterized by a high HR and a low FAR. As a reference for the performance of our METRo configuration, rather than the rates of a completely random predictor (each equal to 0.5), we use the climatological rate of occurrence, CRO, of freezing road-surface temperatures. This mimics a reference that randomly predicts freezing road-surface temperatures based on the CRO of RST ≤ 0 °C. This CRO is in our case derived from the SVV measurement data and defined as the ratio of observations that are ≤0 °C over a given time frame (in this case, the full year 2021). Such a reference system with random predictions based on the CRO would have both a hit rate and a false-alarm rate equal to the CRO. However, it is crucial to note that this climatology-based reference does not account for seasonality. This approach would predict freezing events even in the middle of summer, which is highly questionable and clearly unrealistic. Consequently, the model’s high HR and low FAR compared to climatology are expected. The climatology is relatively easy to outperform since it lacks the temporal context necessary for accurate seasonal predictions.

Figure 5 shows the HR and FAR for freezing RSTs over the full set of road weather forecasts in 2021. The model achieves a high HR and a significantly lower FAR compared to the climatological reference, highlighting its capability to capture freezing conditions accurately. The low FAR is particularly encouraging, as it minimizes false alarms and reduces unnecessary interventions by road authorities. The slight decrease in HR at longer lead times may indicate increasing uncertainty in the forecast, which is typical for most prediction systems. Despite this, the model’s overall performance confirms its reliability as a tool for road-safety applications.

5.2. Seasonal Mean Performance

In investigating seasonal biases and MAEs of RST during winter (DJF), spring (MAM), summer (JJA), and fall (SON), our METRo configuration is subject to a positive surface temperature bias for spring, fall, and particularly summer, where the two peaks are visible again (Figure 6). While this could suggest potential issues related to solar radiation or cloud cover representation, directly attributing the origin of the bias to radiation is challenging due to competing effects in the model. Radiation is most likely accurate, but underestimated surface fluxes may lead to excessive surface heating. Errors in RST in winter are lower than in summer. Keeping in mind that the main purpose of the road weather forecasts is the prediction of hazardous driving conditions, such as freezing precipitation on the road, it is reassuring to find that the RST verification scores for winter and fall are much better than those averaged over the full year.

In contrast, winter RST errors are generally smaller than in summer. However, the model exhibits a negative bias during this season, indicating that it tends to predict temperatures that are too cold. This likely results in an overprediction of freezing conditions, which can negatively impact model performance and forecast accuracy. Given that the primary purpose of road weather forecasts is to predict hazardous driving conditions, such as freezing precipitation on the road, it is reassuring to find that RST verification scores for winter and fall are better than those averaged over the full year.

Errors in air temperatures are the largest in wintertime, but overall remain relatively low and consistent across all seasons (Figure 7). The biases in air temperature are small, indicating no systematic deviations. However, the MAE is quite large in all seasons, suggesting that positive and negative errors compensate each other in the mean error calculation. This compensation effect highlights the need to analyze error distributions more closely to understand the underlying variability.

Similarly, seasonal differences in errors of dew point exhibit the same pattern, Figure 8. Dew point errors are generally low, but seasonal variations still exist. Given the relationship between dew point and air temperature, these errors may reflect the combined influence of atmospheric humidity and temperature variations, which should be further explored to assess their implications for predictive modeling.

The seasonal errors in 10 m wind speed are lower in summer and fall, and larger in winter and spring (Figure 9).

Seasonal Hit Rates for Predictions of Freezing RST Conditions

Overall, METRo has a good performance for hazardous driving conditions due to freezing temperatures on Norwegian roads (Figure 10). Even in wintertime, when the climatology-based hit rate is already over 0.9, the METRo forecasts add additional forecast skill to the predictions (Figure 10) In the shoulder seasons, spring and fall, where freezing temperatures occur more seldomly, the METRo forecasts have a hit rate much higher than the climatology. It is only in summer, when the RST forecasts are subject to positive temperature biases (Figure 4 and Figure 6), that the METRo forecasts do not add additional skill, though the occurrence of freezing events at the SVV weather stations is extremely rare.

The increases in hit rate are complemented by low false-alarm rates in winter, spring, and fall (Figure 11). For these seasons, the FARs of METRo are consistently lower than the values derived from the climatological occurrence of freezing RSTs. The low false-alarm rate in summer is, in this case, not a sign of good model performance, as it is paired with a low, almost non-existent, hit rate. The positive temperature bias seems to make it impossible for the model to reach freezing RSTs.

6. Summary and Conclusions

We set up and evaluated a road weather prediction system for Norway based on the road weather model METRo. Our METRo configuration is initialized with road weather measurements from stations belonging to the Norwegian Road Authority. Boundary conditions during the forecast are derived from numerical weather forecasts by the Norwegian Meteorological Institute. The performance of our system is evaluated based on a quasi-operational setup over the course of the year 2021. road weather forecasts at the locations of 27 road weather stations are initialized twice a day, at 4 a.m./p.m., with a forecast length of 24 h. We used the time series from the road weather station measurements to verify the METRo model output.

The forecast of near-surface air temperature, wind speed, and dew point from the proposed system are similar to the measured values (Figure 4). There are small negative biases in both the air temperature and the dew point. The wind speed is subject to a certain positive bias, which may be caused by an irregular measurement height for wind speed at the road weather stations. The MAEs for these variables are, however, not large. As these variables are prescribed by the boundary conditions in METRo, our validation shows that the data from the weather forecasts by the Norwegian Meteorological Institute have low errors at the locations of the road weather stations, and they are a suitable source for the boundary conditions in our METRo road weather forecasts.

In order to evaluate the performance of the METRo model itself, we compute the MAEs and bias for the road-surface temperature (RST, Figure 4). The MAE increases over the forecast lead time, with a considerable contribution of a bias of 1–2 °C. This positive bias seems to be related to the solar radiation, as it is revealed by looking at the forecasts initialized at 4 a.m. and 4 p.m., separately. A seasonal analysis of RST verification scores confirms the relation between positive RST bias and solar irradiance (Figure 6). Errors in the summer months are especially high and the mean errors over the whole year of 2021 deteriorate considerably because of this. METRo’s performance in winter, spring, and fall, the most critical seasons for hazardous driving conditions related to freezing RST, is much better than that in summer.

More importantly, the plain magnitude of the absolute error in RST is the model’s capability to distinguish between freezing and non-freezing road-surface temperatures. We evaluate METRo’s skill by computing hit rates and false-alarm rates for the occurrence of RST ≤ 0 °C. METRo’s performance is characterized by an increased hit rate and a low false-alarm rate, averaged over the full year of 2021 (Figure 5), and in winter, spring, and fall (Figure 10 and Figure 11). For the application of traffic safety in wintertime, it is noteworthy that the hit rate for freezing surface temperatures in our METRo configuration lies above the climatological rate of occurrence, even though the latter is already at about 95%. The large temperature bias in summer prevents successful summertime predictions of RST ≤ 0 °C. The quasi-operational road weather prediction system, which is developed and verified in this study, would provide valuable and much needed information on hazardous driving conditions and safety on the Norwegian road network.

More specifically, our METRo configuration represents a successful proof-of-concept for road weather forecasts initialized by Norwegian road weather station data and driven by weather forecasts from the Norwegian Meteorological Institute. An operational setup following our description would not be difficult to realize, as we use an open-source model, and all input data are already routinely collected and prepared by Statens Vegvesen and the Norwegian Meteorological Institute. For future work on this subject, we recommend a simple bias correction for the boundary conditions of air temperature, dew point, and wind speed, which are all subject to small, but systematic errors. We also recommend the use of more accurate boundary conditions for solar irradiance, instead of octal cloud cover.