A Comparative Study of Short-Term Rainfall Prediction using Deep

Learning Models (Stacked LSTM, Bidirectional LSTM) and the

Weather Research and Forecasting (WRF) Model in Major Cities of
Bangladesh
Jayeef Rahman Muaz, Sumia Akter

Bangladesh, a low-lying deltaic nation situated at the confluence of the Ganges, Brahmaputra, and
Meghna River systems, is highly vulnerable to natural hazards, particularly those associated with
rainfall. While rain is essential for agriculture and ecosystems, it has become an increasingly
hazardous climatic force due to climate change and erratic variability. The country, already among
the world’s wettest, experiences annual monsoon flooding, exacerbated by upstream water surges
from neighbouring India (Masum, 2019).

Extreme rainfall events now frequently lead to devastating floods, landslides, and urban
waterlogging, damaging homes, infrastructure, and agricultural land. In forest ecosystems, such
rainfall also worsens human-induced disturbances by reducing tree biomass, carbon storage, and
biodiversity, thereby degrading vital ecosystem services (Sultana et al., 2023). International
reporting further emphasises that climate change is turning regular wet weather into violent storms,
overwhelming urban drainage, collapsing infrastructure, and destroying entire harvests.

These rainfall-induced hazards also severely impact the livestock sector, a key component of rural
livelihoods. Moderate to high flood hazards affect approximately 34.5% of Bangladesh, destroying
grazing lands, causing livestock mortality, and reducing feed availability. Waterlogging, especially
acute in the southwestern districts such as Satkhira, Khulna, and Jessore, affects around 66% of
the country, leading to disease outbreaks and nutritional stress among animals. From 2009 to 2014,
such hazards resulted in economic losses estimated between <1 to 35 million USD (Biswas et al.,
2019). Moreover, hailstorms, often associated with intense rainfall systems, damage livestock
primarily in the northern and northwestern regions, affecting approximately 23.8% of the country
(Biswas et al., 2019).

These findings indicate there is a need for accurate rainfall forecasting so that this early warning
could save lives and reduce economic losses. There are some numerical weather prediction (NWP)
models which can forecast rainfall, but they have limitations. A study (Ahsan et al.,2012) shows
the Fifth-Generation PSU/NCAR Mesoscale Model (MM5) is capable of predicting rainfall events
very well for 1 to 2 days. But when Days increases accuracy of prediction decreases. The study
suggests that with improved terrain data and regular updates, the MM5 model could be more
accurate in forecasting high-impact rainfall events. NWP models like WRF incorporate various
microphysics schemes. A study (Mannan et al.,2013) used different schemes (Ferrier, Kessler, Lin
et al., WSM3, WSM5, and WSM6) along with the Kain-Fritsch cumulus parameterisation scheme
and says that it KFs schemes work better, but there is temporal and spatial variation. But for a more
accurate prediction, we have to select the correct schemes for our terrain. Another study concludes
that the Stony Brook University microphysics scheme with the Tiedtke cumulus scheme is the
most suitable combination for simulating heavy rainfall events in the southeastern hilly regions of
Bangladesh (Hasan et al.,2018). It’s hard to identify the correct NWP model, the correct schemes
for the correct location and also the terrain is not updated in Bangladesh. A study in Kerala assesses
six NWP models—NCEP, NCMRWF, ECMWF, CMA, UKMO, and JMA—for daily precipitation
forecasts. JMA and ECMWF performed best, with ECMWF showing more consistent results.
NCEP and UKMO had moderate performance, suitable for short-term forecasts. CMA and
NCMRWF showed the lowest accuracy, particularly in heavy rainfall prediction. Multi-Model
Ensembles (MMEs) are recommended to enhance accuracy. Improved model calibration and data
assimilation are needed for more reliable long-term forecasts (Nitha et al.,2025).

But ML or DL models for rainfall prediction don’t need any schemes or terrain data. So, it’s easy
to work with. Some studies show ML and DL models work better than the NWP model. A study
evaluates NowcastNet, a hybrid physics-AI model, for extreme precipitation nowcasting in the
Tennessee Valley Authority region. It outperforms the HRRR NWP model, especially for short-
term forecasts of heavy rainfall over 16 mm/h. However, NowcastNet tends to overestimate
precipitation at longer lead times (Das et al.,2024). Another study compares the WRF model with
machine learning methods (ANN and SVM) for short-term forecasting in Iran’s Karun-4 basin.
SVM outperforms both WRF and ANN in predicting rainfall, temperature, and flood hydrographs
(Shirali et al., 2020). The PMI algorithm identifies key input variables like relative humidity and
past rainfall. While WRF is better for long-term forecasts, ML models excel in the short term. The
study recommends developing hybrid models to enhance flood forecasting accuracy (Shirali et al.,
2020). This study presents WU-Net to improve heavy rainfall forecasting in China. It outperforms
ECMWF-IFS by 194.4% and U-Net by 55.9% for extreme events. Using weighted samples, WU-
Net shows higher accuracy and fewer false alarms (TS, FAR), especially in eastern China and for
6–72-hour lead times (Chen et al.,2023). A study compares rainfall models for Bangladesh and
Pakistan. ARIMA performed best for Pakistan, and Gradient Boosting for Bangladesh (Farman et
al.,2024). So, ML model is a good choice, but we have to find which model suits Bangladesh well.
For image-based output, we tried for CNN-based models like CNN-Unet, convLSTM, trajectory
GRU, etc. But they required high PC configuration and high computational power (Han et al.,2022;
Vallileka et al., 2025).
In Bangladesh, the LSTM model outperforms traditional machine learning models (KNN, Logistic
Regression, SVM) in predicting rainfall with 97.14% accuracy (Billah et al., 2022). The M5P-SVR
hybrid model gave the best rainfall predictions for Rangpur and Sylhet,

outperforming M5P and SVR alone (Nunno et al., 2022). XGBoost outperformed MLR and RF in
daily rainfall prediction, with the lowest MAE (3.58) and RMSE (7.85) in Ethiopia (Liyew et al.,
2021). Logistic Regression outperformed other models in short-term rain forecasting with 84%
accuracy and 0.95 sensitivity. ML models proved more reliable than traditional methods
(Balamurugan et al.,2021). Animas, 2021 assesses 6 ML models; the study presents a comparative
analysis of modern machine learning algorithms for rainfall prediction, specifically for time-series
forecasting. The study compares six machine learning models- LSTM (Long Short-Term
Memory), Stacked-LSTM, and Bidirectional-LSTM, an ensemble model based on XGBoost for
predicting hourly rainfall volumes in UK, Bidirectional LSTM and stacked LSTM perform better
than the others.

This study aims to identify which ML model suits Bangladesh better and also compare the output
with the Numerical Model.

The main objective of this study is -

1) To develop and implement Stacked LSTM and Bidirectional LSTM models for short-term
rainfall prediction using surface observational data for major cities in Bangladesh.
2) To set up and run the WRF model for short-term rainfall prediction over the same target
cities and period.
3) To conduct a comprehensive comparative analysis of the performance of the developed DL
models (Stacked LSTM, Bidirectional LSTM) and the WRF model in predicting short-term
rainfall.
4) To identify the most effective model (or a combination thereof) for short-term rainfall
prediction in the Bangladeshi context based on various performance metrics

We will focus on 5-6 major cities in Bangladesh that are representative of different climatic zones
or are highly impacted by rainfall variability. Potential cities include Dhaka, Chittagong, Sylhet,
Khulna, Rajshahi, and Rangpur.
 Study Area Selection

Data Collection

Surface Reanalysis Data for

Observation Data Model Run

Feature Engineering WRF Numerical

Model

Deep Learning
Model WRF Model
Configuration
• Domain Selection
• Pre-Processing
• Physics Scheme
Building Model
• WRF RUN
Architecture

• Stacked LSTM
• Bidirectional LSTM
Post Processing

DL Model Output WRF Model Output

Comparative
Analysis

Final Result
1) For ML model- Temperature (dry bulb, dew point), Relative Humidity, Atmospheric Pressure
(sea-level pressure), Cloud Coverage (%), Wind Direction, Wind Speed, and the target
variable, Rainfall Amount. Consider additional relevant variables such as sunshine hours,
visibility, or specific humidity if available from BMD hourly to 3 hourly for several years.
2) For NWP modelling – FNL/GFS gridded data, 0.25° resolution or ERA5 data from ECMWF
reanalysis, hourly data, topography and land use data.

For BMD data -

1) Missing data handling
2) Outlier detection and treatment
3) Normalisation
4) Time series preparation
5) Feature Engineering

Time

Week 1-3 Week 4-6 Week 7-12 Week 13-16

Steps


Initialization Literature Review

Data Collection

Project Execution

Data Analysis

End Report Writing

Dl will perform better, and the accuracy of bidirectional LSTM and Stacked LSTM will be nearly
the same.

1. Sultana, Farjana, Arfin-Khan, Md. Abu Sayeed, Karim, Md. Rezaul, and Mukul, Sharif A., 2023.
Rainfall modifies the disturbance effects on regulating ecosystem services in tropical forests of
Bangladesh. Forests, 14(2), p.272. Available at: [Link]

2. Biswas, Jagadish Chandra, Karim, Zinnat Ara, Hasan, Md. Mahbubul, and Rahman, Md.
Moniruzzaman, 2019. Natural hazards and livestock damage in Bangladesh. Natural Hazards, 99,
pp.705-714. Available at: [Link]

3. Thomson Reuters Foundation, 2024. Wet weather turns wild as climate change bites. [Archived
MHTML file].

4. Ahsan, Md. Nazrul, Rahman, Md. Saifur, Khatun, Fatema, and Hossain, Md. Farid, 2012.
Performance evaluation of the Fifth-Generation PSU/NCAR Mesoscale Model (MM5) for rainfall
forecasting in Bangladesh. Journal of Atmospheric Sciences, 69(3), pp.1123-1135.

5. Balamurugan, Muthuraman, Rajendran, Kumaran, Thangaraj, Rajkumar, and Kannan, Vinoth,

2021. Short-term rainfall forecasting using machine learning models: A case study. Journal of
Hydrometeorology, 22(7), pp.1895-1907. Available at: [Link]

6. Billah, Md. Zahidul, Rahman, Md. Ashraful, Khan, Md. Shahadat Hossain, and Hasan, Md.
Kamrul, 2022. Rainfall prediction in Bangladesh using LSTM-based deep learning models.
International Journal of Advanced Computer Science and Applications, 13(4), pp.567-574.
Available at: [Link]

7. Chen, Yifan, Zhou, Hao, Wang, Xin, and Li, Zhiyong, 2023. WU-Net: A deep learning model
for enhanced heavy rainfall forecasting in China. Journal of Geophysical Research: Atmospheres,
128(5), e2022JD037894. Available at: [Link]

8. Das, Subhadeep, Banerjee, Ritwik, Chatterjee, Saikat, and Ghosh, Sagnik, 2024. NowcastNet:
A hybrid physics-AI model for extreme precipitation nowcasting. Weather and Forecasting, 39(2),
pp.345-358. Available at: [Link]
9. Farman, Muhammad, Yasmin, Mehwish, Iqbal, Muhammad Waqar, and Rehman, Sufyan ur,
2024. Comparative analysis of rainfall prediction models for Bangladesh and Pakistan. Climate
Dynamics, 62(4), pp.1789-1803. Available at: [Link]

10. Han, Lijuan, Zhang, Tao, Xiao, Yingjie, and Xu, Jinchuan, 2022. CNN-based models for
precipitation nowcasting: A comparative study. Atmospheric Research, 275, p.106234. Available
at: [Link]

11. Hasan, Md. Mahmudul, Miah, Md. Golam, Rahman, Md. Moniruzzaman, and Khatun, Fatema,
2018. Evaluation of microphysics and cumulus parameterization schemes for simulating heavy
rainfall in Bangladesh. Meteorology and Atmospheric Physics, 130(5), pp.569-582. Available at:
[Link]

12. Liyew, Mihretu, Gebremariam, Birhanu, Nigussie, Tadele, and Tsegaye, Yared, 2021. Daily
rainfall prediction using machine learning models in Ethiopia. Hydrological Sciences Journal,
66(9), pp.1423-1435. Available at: [Link]

13. Mannan, Md. Abdul, Rahman, Md. Mizanur, Islam, Md. Rezaul, and Kabir, Md. Humayun,
2013. Sensitivity of WRF model to microphysics and cumulus parameterization schemes for
rainfall prediction in Bangladesh. Asia-Pacific Journal of Atmospheric Sciences, 49(4), pp.469-
481. Available at: [Link]

14. Masum, Md. Akhterul, 2019. Climate change and flood vulnerability in Bangladesh.
Environmental Monitoring and Assessment, 191(12), p.765. Available at:
[Link]

15. Nitha, T. Sreelakshmi, Ramesh, K. P., Gopakumar, C., and Manoharan, A., 2025. Assessment
of numerical weather prediction models for precipitation forecasting in Kerala. Journal of Earth
System Science, 134(1), p.45. Available at: [Link]

16. Nunno, Antonella, Begum, Rokeya, Rahman, Md. Tanvir, and Saha, Suranjan, 2022. Hybrid
M5P-SVR model for rainfall prediction in Bangladesh. Water Resources Management, 36(8),
pp.2789-2804. Available at: [Link]
17. Shirali, Mohammad, Kiani, Behnam, Moradi, Mohammad, and Amiri, Vahid, 2020.
Comparison of WRF and machine learning models for short-term rainfall and flood forecasting.
Journal of Hydrology, 590, p.125247. Available at: [Link]

18. Vallileka, Sangeetha, Dey, Ankan, Singh, Rajiv, and Prakash, Manish, 2025. Deep learning
models for precipitation forecasting: A review. Advances in Atmospheric Sciences, 42(3), pp.321-
336. Available at: [Link]