Xie et al. (2025) Near-Surface Temperature Prediction Based on Dual-Attention-BiLSTM

Identification

• Journal: Atmosphere
• Year: 2025
• Date: 2025-10-10
• Authors: Mei Du, C. Li, Gang Du
• DOI: 10.3390/atmos16101175

Research Groups

• Department of Mathematical and Physics, Shijiazhuang Tiedao University, Shijiazhuang, China

Short Summary

This study developed a Dual-Attention-BiLSTM model, integrating random forest-based feature selection and two novel attention mechanisms, to improve hourly short-term near-surface temperature prediction. The model significantly enhanced prediction accuracy compared to a standalone BiLSTM network, demonstrating superior practical application value for short-term forecasts in inland areas.

Objective

• To develop and evaluate a Dual-Attention-BiLSTM model, integrating random forest-based feature selection and two self-designed attention mechanisms (Key-Value Attention and Feature Attention), for hourly short-term near-surface temperature prediction (12, 24, 36, and 48 hours ahead) in Shijiazhuang City, China, addressing limitations of existing methods in handling meteorological factor contributions and temporal context.

Study Configuration

• Spatial Scale: Shijiazhuang City, Hebei Province, China (114.5°E, 38°N), with data at a spatial resolution of 0.1° × 0.1° (approximately 9 km).
• Temporal Scale: Hourly predictions for 12, 24, 36, and 48 hours ahead, using hourly input data from the year 2022.

Methodology and Data

• Models used:
  • Developed Model: Dual-Attention-BiLSTM (Bidirectional Long Short-Term Memory network with Key-Value Attention Mechanism and Feature Attention Mechanism, initialized with Random Forest feature weights).
  • Components: Bidirectional Long Short-Term Memory (BiLSTM), Random Forest (RF) for feature importance, Key-Value Attention Mechanism, Feature Attention Mechanism.
  • Comparative Models: BiLSTM (baseline), Feature-BiLSTM, Key-Value-BiLSTM, BiLSTM-Kalman, TD-LSTM.
• Data sources: ERA5-Land reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF) Climate Data Store.
  • Variables: Hourly near-surface temperature, 10 m U-component wind, 10 m V-component wind, surface pressure, total precipitation, and near-surface dew point temperature.

Main Results

• The Dual-Attention-BiLSTM model significantly improved prediction performance compared to using BiLSTM alone.
• The mean absolute error (MAE) for the Dual-Attention-BiLSTM model ranged from 0.80 °C to 1.08 °C, with a reduction of 0.17 °C to 0.39 °C compared to BiLSTM alone.
• The root mean square error (RMSE) for the Dual-Attention-BiLSTM model ranged from 1.17 °C to 1.37 °C, with a reduction of 0.12 °C to 0.22 °C compared to BiLSTM alone.
• Scheme 4 (Dual-Attention-BiLSTM) achieved the best overall performance, with an average RMSE of 1.24 °C and MAE of 0.92 °C, representing a 12.1% reduction in RMSE and a 22.0% reduction in MAE compared to the BiLSTM-only model (Scheme 1).
• While the TD-LSTM model showed slightly lower error metrics in some cases, the Dual-Attention-BiLSTM model demonstrated better practical application value for short-term predictions (within 24 hours) by providing stable performance and rarely overestimating observed temperatures, which is crucial for disaster warnings and resource management.
• The model performed optimally for 24-hour predictions, suggesting a need for improved generalization across various forecast periods.

Contributions

• Developed a novel Dual-Attention-BiLSTM model that integrates random forest-based feature selection with two self-designed attention mechanisms for enhanced near-surface temperature prediction.
• Introduced a Feature Attention Mechanism that dynamically weights meteorological features based on random forest importance, reducing interference from redundant information.
• Proposed a Key-Value Attention Mechanism that learns contextual information across different time steps, overcoming the limitation of traditional attention mechanisms in treating features homogeneously within a time step.
• Demonstrated that the synergistic effect of both attention mechanisms significantly improves the model's predictive capability compared to single attention mechanisms or a standalone BiLSTM.
• Provided a robust model with high practical application value for short-term (within 24 hours) near-surface temperature forecasting in inland regions, particularly in avoiding critical overestimations of peak and valley temperatures.

Funding

• College Students’ Innovation and Entrepreneurship Training Program: "Research on Temperature Prediction Model Based on Artificial Neural Network" (No. S202410107108)
• National Natural Science Foundation of China (No. 42306233)

Citation

@article{Xie2025NearSurface,
  author = {Xie, Wentao and Du, Mei and Li, C. and Du, Gang},
  title = {Near-Surface Temperature Prediction Based on Dual-Attention-BiLSTM},
  journal = {Atmosphere},
  year = {2025},
  doi = {10.3390/atmos16101175},
  url = {https://doi.org/10.3390/atmos16101175}
}