Twin AI brains team up to predict icy roads before they freeze

Every winter, a thin film of ice on an asphalt road can turn an ordinary commute into a lethal hazard. The difference between a highway crew salting in time and drivers hitting black ice often comes down to one number: the temperature of the road surface itself. A team of Chinese researchers has now unveiled a hybrid artificial intelligence framework that predicts this critical variable with remarkable precision, outperforming ten rival models across forecasting horizons of one, three, and six hours. The work, published in Geoscientific Model Development, could reshape how road authorities anticipate icing events on instrumented highways.

The new system, called the Improved LSTMs Ensemble with Stacking, or ILES, was developed by Wanting Li of Nanjing University of Information Science and Technology and colleagues, including corresponding authors Linyi Zhou of the Chinese Academy of Meteorological Sciences and Xianghua Wu. Rather than betting on a single neural network, the framework deploys two architecturally distinct deep learning models as base learners and fuses their outputs through a statistical meta-learner. The result is a prediction engine that captures both the fine-grained local rhythms of road weather and the long-range temporal dependencies that govern how pavement heats and cools.

The physics behind the problem is deceptively complex. Road surface temperature is governed by the surface energy balance at the pavement-atmosphere interface, where net radiation is partitioned into turbulent sensible heat, latent heat from moisture, heat stored in the pavement substrate, and the energy consumed or released when surface water freezes or melts. Because these fluxes respond to air temperature, humidity, wind, and precipitation with varying lags, road temperature exhibits strong diurnal periodicity and abrupt transitions during cold-air outbreaks or rain. Physics-based models can simulate this balance accurately, but only when detailed pavement thermal properties, such as conductivity, heat capacity, and emissivity, are known, and those parameters are rarely measured at operational road weather stations.

Data-driven approaches sidestep that parameter problem by learning directly from observations, but most machine learning methods treat each time step independently, ignoring the thermal inertia embedded in the road itself. The first ILES base learner, KNN-LSTM, tackles this from an unusual angle: before feeding a sequence of observations into a three-layer Long Short-Term Memory network, it searches the historical record for the fifteen most similar past weather situations and appends their average as an extra feature. This similarity-based retrieval lets the network exploit locally recurring meteorological patterns, effectively giving it a memory of how the road responded to comparable conditions before.

The second base learner, BiLSTM-MHA, approaches the problem from the opposite direction. A bidirectional LSTM processes the 24-hour input window both forward and backward in time, while a multi-head self-attention mechanism, borrowed from the transformer architectures that revolutionised language modelling, dynamically weights which moments in the window matter most. Residual connections and layer normalisation keep training stable. Together, these components excel at extracting long-range dependencies, such as sustained cooling trends or the slow thermal response of pavement to a day of sunshine, that a unidirectional network scanning only forward can miss.

The stacking ensemble that binds these two models is built with unusual care to avoid data leakage, a common pitfall in machine learning forecasting. Base learner predictions used to train the meta-learner are generated through three-fold cross-validation in which each fold corresponds to one complete winter season, so the meta-learner is always trained on genuine out-of-sample forecasts rather than fitted values. A Bayesian Ridge Regression meta-learner then fuses the two streams of predictions, automatically tuning its own regularisation strength through evidence maximisation. Crucially, this Bayesian layer produces closed-form uncertainty estimates, decomposing forecast error into irreducible noise and parameter uncertainty, so operators receive not just a temperature number but a calibrated probability band around it.

The team trained and evaluated the framework on four consecutive winters of hourly observations, from December 2020 to February 2024, collected at road weather station M9393 near a railway overpass in the flat inland plain of Jiangsu Province, where winter minimum temperatures reach roughly minus three degrees Celsius. After rigorous quality control, including the rejection of physically implausible sensor readings and careful gap-filling drawn only from training-period data, the dataset comprised 8,664 hourly samples. The results were striking: ILES achieved coefficients of determination of 0.993, 0.923, and 0.826 at the one, three, and six hour horizons respectively, with mean absolute errors ranging from 0.373 degrees Celsius at one hour to 2.108 degrees at six hours.

Against the benchmark field, which included naive persistence forecasting, nonlinear regression, random forest, XGBoost, GRU, standard LSTM, bidirectional LSTM, and CNN-LSTM models, ILES cut mean absolute error by roughly 8 to 33 percent at the one hour horizon and by 5 to 13 percent at three hours, and it posted the lowest error of all eleven models at nearly every horizon, the lone exception being a marginally lower error by GRU at six hours. An ablation study confirmed that both innovations earn their keep: multi-head attention alone reduced error by 8.2 percent and similarity augmentation by 10.3 percent, with the combined ensemble delivering the best overall performance even though the two gains partially overlap.

Perhaps the most practically important finding concerns what goes into the model. The researchers compared three input configurations: station observations alone, station observations enriched with ERA5-Land reanalysis data, and station observations augmented with physics-motivated engineered features, namely the surface-to-air temperature gradient and multi-scale rates of change of road temperature. The engineered features won decisively, reducing error by nearly 29 percent relative to the plain station baseline at one hour, while the reanalysis augmentation actually performed worst. The explanation is spatial: reanalysis products average conditions over grid cells of roughly nine to eleven kilometres, whereas road surface temperature is a fiercely point-scale quantity. For well-instrumented sites, domain knowledge beats big external datasets, and no new sensors are required.

Interpretability and generalisation tests rounded out the validation. SHAP-based attribution analysis, stratified by temperature regime and time of day, showed that the model leans most heavily on air temperature, exactly as surface energy balance theory predicts, with its influence intensifying at thermal extremes, and the ranking of secondary drivers remained physically plausible. The framework was also retrained at two independent stations, one on a Yangtze River bridge and one near Huai’an Airport, where it beat the standard LSTM baseline by between 5.7 and 35.7 percent in mean absolute error, and its Bayesian prediction intervals proved conservatively calibrated at all three sites. The authors caution that the framework has so far been tested only in a temperate monsoon climate and functions as a station-specific predictor that cannot be blindly transferred to road segments with different shading or orientation. Still, with icing risk hanging on a single degree, a model that forecasts road temperature to within a few tenths of a degree hours ahead, and tells you how sure it is, could give winter maintenance crews the head start that saves lives.

Subject of Research: Hybrid deep learning ensemble prediction of winter road surface temperature for icy road prevention

Article Title: A hybrid method for winter road surface temperature prediction using improved LSTMs and stacking-based ensemble learning

Article References: Li, W., Zhou, L., Wu, X., Guan, Y., Guo, Y., Chen, K., Huang, W., & Zhao, W. (2026). A hybrid method for winter road surface temperature prediction using improved LSTMs and stacking-based ensemble learning. Geoscientific Model Development, 19(18), 9035-9061. https://doi.org/10.5194/gmd-19-9035-2026

Image Credits: AI Generated

DOI: 10.5194/gmd-19-9035-2026

Keywords: road surface temperature, deep learning, LSTM, ensemble learning, stacking, winter road maintenance, traffic safety, machine learning, Bayesian regression, SHAP interpretability, ERA5-Land, Jiangsu