the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 20 Oct 2025
A Hybrid Method for Winter Road Surface Temperature Prediction Using Improved LSTMs and Stacking-Based Ensemble Learning
Abstract. Wintertime low temperatures and snow cover usually diminish the friction coefficient of asphalt pavements, thereby elevating accident and congestion risks. Road surface temperature (RST) is an important parameter for maintaining traffic safety under extreme winter weather conditions, as it helps predict road icing events. Aiming to enhance the precision and robustness of RST prediction, this paper introduces a forecasting framework combining optimized Long Short-Term Memory (LSTM) architectures with a stacking-based ensemble strategy. Two improved LSTMs are constructed: (1) KNN-LSTM, integrating K-nearest neighbors to capture local spatiotemporal similarity patterns, and (2) Attention-BiLSTM, employing bidirectional temporal modeling with dynamic attention weighting mechanisms. These models function as base learners in the stacking ensemble, with Bayesian ridge regression utilized as the meta-learner to consolidate their predictions. The proposed hybrid model was trained and validated using minute-resolution winter meteorological data (2020–2024) collected from a station located on the Longhai Railway Bridge in Jiangsu, China. Experimental results show that the KNN-LSTM and Attention-BiLSTM models exhibit complementary advantages in capturing localized and global temporal features. The ensemble model demonstrates superior performance over individual models, achieving a 1-hour MAE of 0.074, MSE of 0.010, and MAPE of 46.7 % with a significant reduction compared with the best-performing single model. Under extended prediction horizons (3-hour and 6-hour), including low-temperature below 0 °C conditions and typical weather backgrounds, the ensemble model sustains high prediction accuracy and stability. These findings underscore the efficacy of integrating local pattern extraction with attention-based mechanisms via ensemble learning, thereby enhancing RST prediction. This study presents a scalable and adaptable framework for intelligent road weather management systems, offering practical insights for operational deployment.
- Preprint
(2630 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 07 Jan 2026)
- CC1: 'Comment on egusphere-2025-3638', Yue Zhou, 27 Nov 2025 reply
-
CC2: 'Comment on egusphere-2025-3638', Fan Lingli, 08 Dec 2025
reply
This manuscript presents a hybrid framework combining improved LSTM architectures with stacking-based ensemble learning for winter road surface temperature prediction. Validated using high-resolution (minute-level) data from a highway meteorological station in Jiangsu Province (2020–2024), the model demonstrates performance across multiple forecasting horizons (1-hour, 3-hour, 6-hour), sub-zero conditions (<0°C), and typical synoptic backgrounds. The research exhibits clear engineering value (e.g., road icing warning) and interdisciplinary innovation at the intersection of machine learning and meteorology. However, improvements are needed in data representativeness, physical mechanism integration, and model generalizability, as discussed below.
Major Comments
(1)Insufficient Data Representativeness and Spatial Generalizability
The model is trained and validated exclusively on data from the Longhai Railway Bridge in Jiangsu, a region with a warm temperate semi-humid monsoon climate. This limits conclusions about performance in diverse geographies where RST dynamics differ due to terrain, vegetation, or pavement materials. And, the dataset lacks explicit coverage of extreme weather years (e.g., severe cold waves), raising questions about model reliability during rare but critical events.
(2)Integrate Meteorological Physics to Enhance Interpretability
Incorporate key parameters from the road surface energy balance equation (e.g., albedo, thermal conductivity, estimated solar radiation) as inputs or constraints.
(3)Inadequate Benchmarking Against State-of-the-Art Models
The manuscript claims superiority over "individual models" (LSTM, KNN-LSTM, Attention-BiLSTM) but lacks comparisons with recent hybrid methods in RST prediction. Quantitative metrics (e.g., MAE, MSE) against these models are absent, weakening claims of methodological advancement.
Minor Comments
(1)Ambiguities in Figure and Table Presentations
Figure 6 (RST periodicity) lacks confidence intervals, making it impossible to assess the statistical significance of diurnal variations.
(2)Inconsistent Terminology and Citation Errors
The term "Attention-LSTM" is used in Figure 12's caption but not defined in the main text; it should be corrected to "Attention-BiLSTM" for consistency.
In Section 4.2, "STM" is referenced in Figure 12's legend but not defined, causing confusion.
Summary
This study contributes a novel ensemble framework for winter RST prediction, with promising results in reducing short-term prediction errors. However, its scientific impact is limited by narrow data coverage, inadequate physical grounding, and insufficient benchmarking. Addressing the major concerns—particularly data representativeness, physical mechanism integration, and comparative validation—will strengthen the manuscript’s validity and relevance to operational road weather management. I recommend minor revisions prior to reconsideration for publication.Citation: https://doi.org/10.5194/egusphere-2025-3638-CC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 266 | 68 | 21 | 355 | 13 | 13 |
- HTML: 266
- PDF: 68
- XML: 21
- Total: 355
- BibTeX: 13
- EndNote: 13
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This manuscript introduces a forecasting framework combining optimized Long Short-Term Memory (LSTM) architectures with a stacking-based ensemble strategy, aimed at predicting road surface temperature (RST). The application of this method can provide certain support for winter RST forecasting. The manuscript has a clear research motivation, and the experimental results show improvements. However, there are still several issues that require further revision.
Major Comments:
1、The discussion on the impact of weather conditions on RST is insufficient and needs to be strengthened.
2、In fact, this study uses hourly data for analysis. A detailed description of data quality and characteristics should be provided. Additionally, how does the authors’ method of converting minute-level data to hourly data differ from that used by meteorological departments?
3、The manuscript devotes substantial space to introducing methodologies. For mature methods, the focus should be on citations and brief descriptions, with emphasis on the application value and innovative points of these methods in this study.
4、A comprehensive introduction to the observation site is required: is it a station on a highway bridge or a regular road surface? The impact of surface latent heat on RST varies significantly between these two settings, and this should be clearly clarified.
5、There are almost no references cited to support the analysis in the main text. Relevant research achievements in the field should be supplemented as theoretical support to enhance the scientific rigor and credibility of the discussion, especially comparative analyses with other RST forecasting methods and results.
6、For RST prediction, forecasting under low-temperature and overcast/rainy conditions is particularly critical. However, the manuscript provides insufficient analysis of RST prediction results under these weather conditions. More relevant analyses should be added, along with physical mechanism explanations for how weather conditions influence prediction performance.
Minor Comments:
1、The descriptions of data in Table 1 are of little significance, as they only cover conventional meteorological variables. More attention should be paid to data distribution and quality control details.