Pan et al. (2026) Univariate vs. Multivariate Long-Short Term Memory for Daily Rainfall Forecasting at a Coastal Station in New South Wales

Identification

• Journal: Lecture notes in civil engineering
• Year: 2026
• Date: 2026-01-01
• Authors: Xiao Pan, Gokhan Yildirim, Ataur Rahman
• DOI: 10.1007/978-3-032-18708-6_4

Research Groups

• School of Engineering, Design and Built Environment, Western Sydney University, Sydney, New South Wales, Australia
• Department of Civil Engineering, Faculty of Engineering, Aksaray University, Aksaray, Türkiye

Short Summary

This pilot study evaluates univariate and multivariate Long Short-Term Memory (LSTM) models for next-day rainfall forecasting at an Australian coastal station. While the multivariate LSTM showed minor improvements over the univariate baseline, both models exhibited limited skill in reproducing daily rainfall amounts, struggling particularly with zero-inflation and extreme events.

Objective

• To evaluate and compare the performance of univariate and multivariate Long Short-Term Memory (LSTM) architectures for next-day rainfall prediction at Albion Park, a coastal station in New South Wales, Australia.

Study Configuration

• Spatial Scale: Point scale (single station: Albion Park, New South Wales, Australia).
• Temporal Scale: Daily rainfall forecasting, using a seven-day antecedent window for prediction.

Methodology and Data

• Models used: Univariate Long Short-Term Memory (U-LSTM) and Multivariate Long Short-Term Memory (M-LSTM) networks. Both models were implemented in PyTorch and trained using the Adam optimizer.
• Data sources:
  • Antecedent daily rainfall observations.
  • Co-observed meteorological drivers: temperature, humidity, wind speed, solar radiation, and evapotranspiration.

Main Results

• The Multivariate LSTM (M-LSTM) achieved a Root Mean Square Error (RMSE) of 0.987 mm and a Mean Absolute Error (MAE) of 0.378 mm on the test dataset.
• The M-LSTM showed a weak correlation coefficient (r = 0.176) and a negative Nash–Sutcliffe Efficiency (NSE = −0.077), indicating limited skill in accurately reproducing day-to-day rainfall amounts.
• Visual diagnostics revealed systematic underestimation and variance compression in the model predictions.
• The multivariate setting provided small but consistent improvements compared to the univariate baseline.
• Both U-LSTM and M-LSTM models struggled significantly with zero-inflation (dry days) and extreme rainfall events.

Contributions

• This pilot study provides a comparative evaluation of univariate and multivariate LSTM models for daily rainfall forecasting in an Australian coastal context, highlighting their current limitations.
• It identifies specific challenges for deep learning models in this domain, such as handling zero-inflation and extreme events.
• The paper proposes a prioritized roadmap for future research, focusing on data quality, architectural improvements, loss function design, and evaluation strategies tailored for Australian conditions, including multi-station learning, probabilistic outputs, and hybrid physical–statistical models.

Funding

Not explicitly mentioned in the provided text.

Citation

@article{Pan2026Univariate,
  author = {Pan, Xiao and Yildirim, Gokhan and Rahman, Ataur},
  title = {Univariate vs. Multivariate Long-Short Term Memory for Daily Rainfall Forecasting at a Coastal Station in New South Wales},
  journal = {Lecture notes in civil engineering},
  year = {2026},
  doi = {10.1007/978-3-032-18708-6_4},
  url = {https://doi.org/10.1007/978-3-032-18708-6_4}
}