https://orcid.org/0009-0008-9588-5986
Figures
Abstract
Flood-prone Bangladesh requires accurate short-term rainfall prediction, yet conventional numerical weather prediction (NWP) models remain computationally intensive and regionally limited in skill. This study evaluates and compares four deep learning (DL) architectures — Simple Long Short-Term Memory (LSTM), Stacked LSTM, Bidirectional LSTM (BiLSTM), and Gated Recurrent Unit (GRU) — against the physics-based Weather Research and Forecasting (WRF) model for one-hour-ahead rainfall forecasting across six climatologically diverse stations in Bangladesh. Each model was trained independently per station using hourly ERA5 reanalysis data spanning 2018–2024, and performance was assessed using both regression and binary detection metrics. Across the full test period, Simple LSTM achieved the strongest overall performance, explaining approximately 88% of ERA5 rainfall variance with the lowest prediction errors, while GRU produced nearly comparable results. Performance was consistently stronger at wetter stations such as Chittagong and Sylhet than at drier northwestern stations such as Rajshahi and Rangpur, where class imbalance and rainfall intermittency limited model skill. Event-category analysis across 10 WRF simulation events — spanning extreme, moderate, and weak rainfall regimes — showed that DL models demonstrated closest agreement with ERA5 reference values during extreme events, while skill declined notably during weak events due to near-zero rainfall conditions. WRF exhibited systematic overestimation, temporal displacement of peak rainfall, and strongly negative R² values across all event categories and stations. A Friedman test confirmed statistically significant differences in RMSE ranks among the four DL architectures (χ² = 10.60, p = 0.014), with a Nemenyi post-hoc test showing Simple LSTM significantly outperformed Stacked LSTM (p = 0.0095).
Citation: Akter S, Muaz JR, Awal R (2026) Comparison of data-driven and numerical weather prediction models for rainfall nowcasting in major cities of Bangladesh. PLOS Clim 5(10): e0001088. https://doi.org/10.1371/journal.pclm.0001088
Editor: Bharath Haridas Aithal, Indian Institute of Technology Kharagpur, INDIA
Received: March 17, 2026; Accepted: September 8, 2026; Published: October 9, 2026
Copyright: © 2026 Akter et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The ERA5 reanalysis data used in this study are publicly available from the Copernicus Climate Data Store. Processed datasets for this study are available in the given data URL link (https://github.com/jayeef-muaz/ERA_5_data_Major_cities_of_Bangladesh-).
Funding: The authors received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
Accurate rainfall prediction has become a critical need as the world faces increasingly frequent extreme weather events. In countries like India, around 60% of farmland relies entirely on monsoon rains, illustrating how precise forecasts are essential for maintaining food security and economic stability [1]. More broadly, the intensification of the global hydrological cycle due to climate change has made short-term rainfall forecasting a central challenge for meteorologists and disaster managers worldwide. According to the Intergovernmental Panel on Climate Change (IPCC), climate systems are shifting toward wetter seasonal patterns marked by stronger heat waves and more intense rainfall episodes in some regions of the world, making accurate predictions progressively more difficult [2].
Considerable research effort has been directed toward improving rainfall prediction methods globally and regionally. Traditional numerical weather prediction (NWP) approaches, particularly the Weather Research and Forecasting (WRF) model, demand massive computing power and rely heavily on choosing the right parameterization settings. Studies have found that the MM5 model (Fifth-Generation PSU/NCAR Mesoscale Model) can forecast rainfall reasonably well for one to two days ahead, but its accuracy drops noticeably at longer lead times [3]. The performance of NWP models depends greatly on the microphysics and cumulus schemes selected; research indicates that although the Kain-Fritsch cumulus parameterization scheme typically gives better results, its effectiveness varies considerably depending on time and location [4]. Comparative evaluations of numerical models including the National Centers for Environmental Prediction (NCEP), National Centre for Medium-Range Weather Forecasting (NCMRWF), European Centre for Medium-Range Weather Forecasts (ECMWF), China Meteorological Administration (CMA), UK Met Office, and Japan Meteorological Administration (JMA) have shown varying levels of success, with ECMWF and JMA generally outperforming the others, though they still require multi-model ensembles and better calibration to become dependable for heavy rainfall prediction [5]. Deep learning methods have started showing real promise as alternatives to traditional NWP techniques, demonstrating strong abilities to capture temporal patterns in weather data without needing complicated parameterization schemes or detailed terrain information [6]. Hybrid models have also delivered encouraging results; the M5P-SVR hybrid model, for instance, has produced more accurate rainfall predictions than either individual M5P or SVR model alone [7]. Long Short-Term Memory (LSTM), its Bidirectional variant (BiLSTM), and Stacked LSTM architectures have been shown to outperform conventional machine learning algorithms such as XGBoost and AutoML [8]. Research comparing WRF with ML and DL models indicates that the relative performance of physics-based and data-driven approaches is highly context-dependent, varying with forecast horizon, target variable, and regional characteristics; DL models have shown particular promise for short-term and nowcasting applications where training data are abundant [9,10]. For specific regional contexts such as Bangladesh’s southeastern hilly areas, studies have found that combining the Stony Brook University microphysics scheme with the Tiedtke cumulus scheme works best for simulating heavy rain events [11].
Despite these global advances, Bangladesh stands out as a country where the need for accurate short-term rainfall forecasting is especially urgent. Its deltaic terrain, sitting close to sea level, makes it highly susceptible to water-related disasters driven by heavy monsoon rains, tropical low-pressure systems, and sudden intense thunderstorms. The catastrophic Sylhet floods of June 2022 demonstrated the devastating consequences of inadequate rainfall warnings, affecting roughly 7.2 million people across nine northeastern districts, with studies showing that approximately 90% of Sylhet division went underwater [12–14]. The August 2024 floods in eastern Bangladesh were equally severe: Feni recorded 832 mm of rainfall within just five days between August 19–23, while Comilla received 513 mm and Noakhali 605 mm [15]. Around 5.8 million people in northeastern and southeastern areas were impacted, with over 1 million people completely isolated by floodwaters [16,17]. The economic consequences are equally serious: floods and waterlogging affect 34.5% and 66% of the country respectively, causing significant damage to livestock and agriculture, particularly in southwestern districts such as Satkhira, Khulna, and Jessore, where economic losses ranged from $1 to $35 million USD during 2009–2014 alone [18]. As extreme weather events become more frequent and intense, the urgency for improved forecasting tools to reduce both human casualties and economic damage continues to grow [19].
Within Bangladesh and the broader South Asian context, machine learning and deep learning approaches have been increasingly applied for hydrometeorological prediction. Support vector machines have demonstrated prediction accuracy comparable to or better than artificial neural networks for flood forecasting at Dhaka across multi-day lead times [20]. LSTM-based deep learning applied for monthly rainfall prediction at Sylhet and Srimangal stations outperformed both ANN and SVM across RMSE, MAE, and R² metrics, representing one of the first applications of deep learning specifically for rainfall prediction in Bangladesh [8]. For multi-target weather prediction across eight Bangladeshi cities including Dhaka, Rajshahi, and Chittagong, the Gated Recurrent Unit (GRU) demonstrated the best overall performance compared to LSTM and Conv1D architectures [21]. A CNN-XGBoost hybrid model applied for daily rainfall prediction in northern Bangladesh achieved R² of 0.99 at Rajshahi, highlighting the potential of hybrid deep learning approaches for region-specific forecasting [22]. However, these studies predominantly focus on daily or monthly timescales using station observations, and direct comparison between multiple deep learning architectures and physics-based NWP models at hourly resolution remains largely unexplored in the Bangladesh context.
Although several studies have applied machine learning techniques for rainfall prediction in Bangladesh, most research has focused either on flood susceptibility mapping or individual model approaches at coarse temporal resolution. Comparative evaluation between deep learning architectures and NWP models for short-term hourly rainfall forecasting across multiple rainfall regimes in the country remains limited. Furthermore, many previous studies relied on daily rainfall data rather than the high temporal resolution required for nowcasting applications. This study therefore addresses these gaps under the following assumptions and scope. ERA5 reanalysis data are assumed to reliably represent the hourly rainfall patterns of Bangladesh, and the six selected stations are assumed to represent the major rainfall regimes across the country. This study evaluates and compares four deep learning architectures — Simple LSTM, Stacked LSTM, BiLSTM, and GRU — against the WRF numerical weather prediction model for short-term hourly rainfall forecasting in six major cities of Bangladesh using ERA5 data from 2018–2024. The research focuses on evaluating predictive accuracy and computational efficiency across rainfall regimes and event intensities, and on identifying which architecture performs best under Bangladesh’s diverse climatological conditions. While Transformer-based architectures have recently demonstrated strong performance in meteorological forecasting tasks [23,24], their inclusion was beyond the scope of this study. Transformer models typically demonstrate their greatest advantage over recurrent architectures at longer sequence lengths where the self-attention mechanism can exploit dependencies across distant time steps; the 24-hour input window used here limits this advantage. Furthermore, this study prioritises computational feasibility for low-resource operational environments, where LSTM-family models offer a more practical balance between predictive skill and computational demand.
2. Data and methodology
This study evaluates four Long Short-Term Memory (LSTM) architectures — Simple LSTM, Stacked LSTM, BiLSTM, and GRU — distinguished by their layer configurations, and benchmarks their performance against the Weather Research and Forecasting (WRF) model for short-term one-hour-ahead rainfall forecasting across six stations in Bangladesh. Each model was trained independently for each station to preserve site-specific climatological signals. Given the lack of high-frequency surface observations from the Bangladesh Meteorological Department (BMD), hourly ERA5 reanalysis data were retrieved from the Copernicus Climate Data Store (CDS). The methodological framework illustrated in Fig 1 describes a systematic pipeline: data organisation, preprocessing, feature engineering via correlation analysis, model training and validation, and final evaluation against WRF simulations using standard performance metrics.
Deep learning models are multi-layered neural networks capable of learning hierarchical representations directly from raw data, automatically discovering features needed for classification or prediction without manual feature engineering. Among data-driven models, LSTM is particularly effective in capturing long-term temporal dependencies and has shown superior performance compared to traditional machine learning models [8]. Different LSTM architectures show varying levels of performance, but all outperform traditional machine learning models in rainfall prediction [6].
2.1. Data collection
This study focuses on short-term rainfall forecasting with a lead time of one hour. Higher temporal resolution is essential for capturing the dynamic evolution of convective systems and producing actionable weather guidance. However, reliable high-frequency station-level observational data are not available in Bangladesh. Although ERA5 reanalysis data are known to underestimate convective rainfall intensity in some tropical regions, they reliably capture the temporal patterns and spatial variability of precipitation, making them widely used as input for data-driven forecasting models [25,26]. This study uses hourly ERA5 reanalysis point data extracted for selected station locations (latitude and longitude) in CSV format for the period 2018–2024. The stations were chosen to represent a range of climatic conditions across the country, from very wet to dry zones, based on long-term climate variability. For the DL models, training data were selected focusing on surface meteorological parameters. We use 80% of the total data to train our models, as this proportion is widely considered suitable for developing data-driven rainfall forecasting models [25–27]. The following 10% are used for validation during model development, and the remaining 10% are reserved as an independent test set, consistent with the approach adopted in previous studies [28,29].
The study region includes six major cities selected based on the distribution of annual average rainfall to include all major precipitation regimes: Dhaka, Sylhet, Chittagong, Khulna, Rajshahi, and Rangpur. Fig 2 illustrates the long-term average rainfall pattern derived from 44 years of CHIRPS data (1981–2024) [30], highlighting that maximum rainfall occurs in the northeastern and southeastern parts of Bangladesh, the central part experiences moderate rainfall, and the lowest rainfall occurs in the west-central and northwestern regions.
The map was created using ArcMap 10.8 (Esri). Administrative boundary shapefile sourced from geoBoundaries (https://www.geoboundaries.org).
2.2. Data preprocessing
To reduce model complexity and minimize the risk of overfitting, only relevant and non-redundant meteorological variables were selected as input features. Since the developed models are primarily based on surface meteorological observations, parameters that are physically associated with rainfall formation were considered during feature selection. In addition to the current atmospheric conditions, antecedent rainfall information was also included because recent rainfall strongly influences atmospheric moisture availability and short-term precipitation persistence. Therefore, rainfall observations from 1 hour and 3 hours prior to the prediction time were incorporated as temporal features.
A correlation analysis was conducted over training data only (first 80%) to examine the relationship between each predictor and rainfall (Fig 3). Variables showing very weak correlations with rainfall (|r| < 0.03), such as year, month, and date, were excluded from the final feature set. However, feature selection was not based solely on statistical relationships; physical relevance to precipitation processes was also considered. As a result, variables such as Convective Available Potential Energy (CAPE) and temperature were retained despite their relatively weaker linear correlations, because they play important roles in atmospheric instability and convective rainfall development. Although some degree of inter-predictor correlation is evident in Fig 3, particularly among moisture-related variables such as dew point temperature, relative humidity, and specific humidity, multicollinearity does not pose the same concern for deep learning models as it does for linear statistical methods. LSTM-family architectures learn nonlinear feature interactions through gated mechanisms and are not sensitive to correlated inputs in the same way as regression-based models, as the network implicitly weights redundant information during training [6].
To capture the short-term temporal evolution of atmospheric conditions, a sequence length of 24-time steps was used. In this approach, the model receives meteorological observations from the previous 24 hours to predict rainfall for the next hour. The use of a 24-hour input window allows the model to learn diurnal rainfall variability, persistence characteristics of rainfall events, and the evolution of convective systems, which are particularly important in tropical regions such as Bangladesh. Previous studies on precipitation nowcasting and deep learning–based rainfall forecasting have shown that incorporating historical meteorological information over several hours significantly improves short-term rainfall prediction performance [31].
The dataset was transformed into sequential samples using a sliding-window approach implemented through NumPy array operations. For each prediction target at time step , the model received meteorological observations from time steps (
) to (
as input, creating overlapping sequences that maximize the utilization of available data. The target variable for each sequence was defined as the one-hour-ahead rainfall amount at time
. This formulation enables the model to learn the temporal evolution of atmospheric conditions leading to precipitation events [31].
To preserve the chronological nature of the time-series dataset and avoid data leakage, the data were divided sequentially into training, validation, and testing subsets. The first 80% of the data were used for training, the following 10% for validation during model development, and the final 10% of unseen data were reserved for independent model evaluation.
The training dataset was scaled using MinMaxScaler (Equation 1) in the scikit-learn library between [0,1], which is mostly used in rainfall forecasting models, then this scaling was applied to the validation and test set.
2.3. Model architecture
The LSTM architecture fundamentally differs from Recurrent Neural Network (RNN)’s through its cell state (Ct), a parallel information pathway that runs through the entire sequence alongside the hidden state (ht). The cell state function carries information across time steps with minimal transformations, enabling the preservation of long-term dependencies. Three specialized gates are the forget gate, input gate, and output gate which regulate information flow into, within, and out of the cell (Fig 4). These gates are neural network layers with sigmoid activation Functions (σ) that produce values between 0 and 1. This architecture allows LSTMs to selectively retain relevant information, discard irrelevant information, and expose appropriate information as output at each step. The LSTM cell at time step t receives three inputs: the current observation Xt, the previous hidden state ht-1, and the previous cell state Ct-1. Through the coordinated operation of its three gates, the LSTM produces two outputs: the current hidden state ht, which serves as the cell’s output and feeds into the next layer or final prediction, and the updated cell state Ct, which propagates to the next time step to maintain long-term memory [32].
The Simple LSTM model is a single layer LSTM model and a stacked LSTM model is a stack of more than one layer. When two or more sequential LSTM layers connect to each other it’s called stacked LSTM. Simple LSTM model configured with 128 unit and tanh activation function in initial cell state. Most researchers employ tanh Activation in LSTM architectures for rainfall and hydrological forecasting Applications, as it effectively captures the bidirectional influences of meteorological parameters on precipitation processes [8,33]. Recurrent dropout used to minimize overfitting. Linear activation function was used in the final dense layer, for continuous rainfall forecasting linear function work well. For the stacked LSTM model, a similar architecture is followed, with an additional dense layer using a ReLU activation function. The model is optimized using the Adam Optimizer.
The hyperparameters listed in Table 1 were determined through a combination of values reported in previous deep learning rainfall forecasting studies [8,33] and manual tuning based on validation loss performance. The difference in epoch counts between architectures reflects convergence behaviour observed during preliminary training: Simple LSTM and Stacked LSTM required more epochs (100) to reach stable convergence, while BiLSTM and GRU achieved stable validation loss within fewer epochs (50), consistent with their more compact parameter structures.
BiLSTM is a variant of the RNN, which uses both forward and backward serial data simultaneously. This structure is made up of two distinct layers of hidden nodes, where one performs the sequence between the start and the end of the sequence and the other performs the sequence between the end and the start of the sequence. Gated Recurrent Unit (GRU) is a simplified form of the LSTM network. It also performs similarly to the LSTM on most tasks with fewer parameters and is therefore computationally more efficient. It combines the LSTM’s forget and input gates into a single update gate, and it merges the cell state and hidden state into a single hidden state.
2.4. Event selection for WRF simulation
To compare the DL model’s prediction with WRF, we used ERA5 reanalysis data [34] as input so that the baseline of both models remains the same. The WRF events were selected from the independent test period (post-May 2024). Event selection was performed using aggregated rainfall statistics derived from six major cities. Hourly precipitation data from each station were first converted into daily accumulated rainfall totals. Subsequently, two regional rainfall indicators were calculated for each day: (i) the spatially averaged rainfall (Rmean) across all stations and (ii) the maximum station rainfall (Rmax). The spatially averaged rainfall was used to characterize widespread precipitation conditions, while the maximum station rainfall was used to identify localized intense rainfall events. Rainfall events were categorized into low, moderate, and extreme rainfall regimes using a combination of percentile-based regional rainfall statistics and station-level rainfall intensity thresholds, as summarized in Table 2.
where P represents the percentile rank of the spatially averaged rainfall distribution during the independent test period. The percentile-based classification of rainfall extremes follows commonly adopted climate-extreme methodologies, including ETCCDI precipitation indices and previous studies on extreme rainfall variability in Bangladesh [35–37]. To better represent evolving synoptic-scale weather systems, consecutive rainfall days satisfying the event-selection criteria were grouped into single multi-day events. This approach also facilitated continuous WRF simulations with appropriate atmospheric evolution, thereby reducing the fragmentation of rainfall episodes. For each selected event, a 24-hour spin-up period was included in all WRF simulations to minimize the influence of initial condition adjustment, consistent with recommended spin-up ranges reported in previous studies [38,39]. The final event set consisted of ten representative events spanning extreme, moderate, and low-rainfall conditions (Showed in Table 3).
The input predictors include surface meteorological variables from ERA5, such as 2 m temperature, 2 m dew point temperature, sea level pressure, relative humidity, wind speed, and total precipitation. These variables were selected based on their physical relevance to precipitation processes and their statistical correlation with rainfall. The WRF Preprocessing System (WPS) geogrid utility was employed to interpolate static terrestrial data to the model domain. Land use and land cover data were obtained from the 30-second resolution USGS (United States Geological Survey) dataset, providing a detailed representation of surface characteristics including vegetation types, urban areas, water bodies, and topography. Moreover, WRF initialization requires two distinct categories of atmospheric data: single-level (surface) parameters and pressure-level (upper-air) parameters with 33 levels. These three-dimensional atmospheric fields characterize the complete thermodynamic and dynamic state of the atmosphere, enabling WRF to simulate the evolution of weather systems through numerical solution of primitive equations [4].
2.5. WRF configuration
Numerical simulations were performed using the Advanced Research Weather Research and Forecasting model (WRF‑ARW) version 4.6.0 [40], a non‑hydrostatic mesoscale numerical weather prediction system widely applied for monsoon and tropical weather studies over South Asia and the Bay of Bengal [41,42]. The model was configured with a single Mercator-projection domain centered at 22.9°N, 89.9°E, encompassing Bangladesh and its adjacent regions. The horizontal grid comprised 129 × 173 points at 6 km spacing, sufficient to resolve mesoscale circulations associated with organized convection and monsoon depressions [43].
The physics configuration follows combinations shown to perform well for South Asian monsoon rainfall. Convective processes were represented with the Kain–Fritsch (KF) cumulus scheme [44], which has been identified as a suitable choice for Indian summer monsoon and heavy rainfall events when combined with Noah land surface and RRTM/Dudhia radiation packages [42,45] and is also routinely used in operational monsoon forecasting over Sri Lanka and the Bay of Bengal region [46]. Cloud microphysical processes were simulated with the WRF Single‑Moment 6‑class (WSM6) scheme [47], which explicitly predicts multiple hydrometeor species and is widely adopted in regional heavy‑rainfall and cyclone simulations over the Bay of Bengal and India [41,43] and in sensitivity studies for Bangladesh rainfall [48].
Planetary boundary layer turbulent mixing in the lower troposphere was represented by the Yonsei University (YSU) PBL scheme [49], which has repeatedly shown good performance for monsoon rainfall and convection over the Indian subcontinent and surrounding basins [42,50]. Land–atmosphere exchanges of heat and moisture were treated using the Unified Noah land surface model [51], which has been widely implemented and verified within WRF for monsoon climates [42,46]. Longwave and shortwave radiative transfer were handled by the Rapid Radiative Transfer Model (RRTM) [52] and Dudhia schemes [53], respectively, a radiation package combination that has been recommended for Indian summer monsoon simulations and is standard in many South Asian WRF studies [41,42].
This KF–WSM6–YSU–Noah–RRTM–Dudhia configuration thus follows a physics suite that has produced realistic monsoon precipitation, tropical cyclone, and heavy rainfall simulations across the broader Bay of Bengal–South Asia region for evaluating WRF rainfall skill over Bangladesh [11,54]. The model was integrated using non‑hydrostatic dynamics and hybrid pressure–sigma vertical coordinates, enabling improved representation of vertical atmospheric structure and mesoscale circulations associated with convective rainfall systems [43]. While a sensitivity analysis of alternative physics combinations was beyond the scope of this study, the selected configuration follows a well-validated suite for the Bangladesh region and serves as a representative benchmark.
For verification purposes, gridded WRF precipitation output was extracted at the nearest grid point to each station’s latitude and longitude coordinates. Given the 6 km horizontal resolution of the model domain, the nearest grid point was within approximately 3–4 km of each station location, which is considered acceptable for mesoscale precipitation verification at the city scale [43].
2.6 Evaluation metrics
To evaluate the performance of data-driven models and compare these to WRF this study used two types of metrics. Model performance is evaluated using regression metrics-MSE, MAE, RMSE, and R2 (Equation 2–5) to assess how well the model captures both the amount and pattern of rainfall. Binary detection metrics (Equation 6–9) were employed to evaluate the model’s ability to predict the occurrence of rainfall, irrespective of its capacity to estimate the actual precipitation amount; these are Precision, Recall, F1, and Accuracy. For detection metrics, rainfall predictions and observations are converted to binary classifications using a threshold of 0.1 mm (rain/no-rain), creating a confusion matrix with four categories: True Positives (TP, correctly predicted rain), True Negatives (TN, correctly predicted no-rain), False Positives (FP, predicted rain but observed no rain), and False Negatives (FN, predicted no-rain but observed rain). Precision (Positive Predictive Value) measures the reliability of positive (rain) predictions. Recall (Sensitivity, True Positive Rate) measures the proportion of actual rainfall events successfully detected. F1-Score harmonically combines precision and recall and Accuracy measures overall correctness.
Regression Metrics:
Binary Detection Metrics:
2.7. Statistical significance testing
To formally assess whether the observed differences in RMSE among the four DL architectures were statistically meaningful, a Friedman test was applied across the six stations, followed by a Nemenyi post-hoc test for pairwise comparison. The Friedman test ranks each model’s RMSE within each station (rank 1 = lowest RMSE) and tests whether the average ranks differ significantly across models, without assuming a normal distribution of the underlying scores.
Let rᵢʲ denote the rank of the j-th of k models on the i-th of N stations, and let Rⱼ = (1/N)∑ᵢ rᵢʲ be the average rank of model j. Under the null hypothesis that all models are equivalent, the Friedman statistic is computed as:
which follows a χ² distribution with k − 1 degrees of freedom [55].
When the Friedman test indicated a significant overall difference, the Nemenyi post-hoc test was used to identify which specific pairs of models differed significantly. Two models are considered significantly different if their average ranks differ by at least the critical difference (CD):
where qa is the critical value of the Studentized range statistic at significance level α = 0.05 (for k = 4 models, q₀.₀₅ = 2.569) [55]. This non-parametric approach was chosen over the paired t-test because RMSE values across only six stations do not reliably satisfy normality assumptions.
3. Results
3.1. Overview of rainfall data
The rainfall distribution in Bangladesh is characterized by extreme spatio-temporal variability, with the highest volumes concentrated in the northeastern and southeastern hilly regions during the monsoon season [56]. Fig 5 illustrates the temporal distribution of rainfall (mm) across the six major cities selected for this study in Bangladesh. The time-series bar charts demonstrate that the ERA5 reanalysis data effectively capture the region’s inherent seasonal variability, aligning with established climatic patterns. For instance, heavy rainfall occurs during June, July, August, and September, while dry conditions prevail during December, January, and February.
3.2. Loss curve analysis
Training and validation loss curves are essential for diagnosing deep learning model convergence, generalization, and issues like overfitting or underfitting. Fig 6 displays the training and validation loss curves for the simple LSTM, stacked LSTM, GRU, and bidirectional LSTM models for all six stations, showing region-specific convergence associated with rainfall regimes.
The horizontal axis represents the training epochs, and the vertical axis denotes the mean squared error (MSE).
In Fig 6, the monotonic decrease in training loss with increasing epochs indicating that the optimisation process successfully reduces prediction error during training. Similarly, the decrease in validation Loss with increasing epochs demonstrates that the models eventually stabilise, suggesting improved model performance on validation data during the training period. The relatively larger and more persistent gap between the training and validation curves found in figures 3.2 (b) and (d) suggests less stable convergence compared to the other architectures, indicating limited overfitting and reasonably stable learning behaviour. Validation loss remaining close to or slightly lower than training loss may be associated with the use of regularization techniques during training observed in Fig 6. This phenomenon is often attributable to regularization techniques such as dropout or batch normalization applied during training, which can increase training loss relative to validation loss, as reported by Jabir & Falih [57].
However, the stacked LSTM model (Fig 6(b)) exhibits more pronounced fluctuations in both training and validation loss throughout the training period. In Fig 6 (b), fluctuation in training loss reflects variability, perhaps due to mini-batch sampling and stochastic regularization methods and becomes negligible with the increase of epochs, indicating the convergence of the model. The stacked LSTM model (Fig 6b) demonstrates less stable convergence behaviour compared to the simple LSTM (Fig 6a) and GRU (figure 6(c)) models, as reflected by the more variable loss trajectories throughout training. Similarly, the BiLSTM (Fig 6d) model shows occasional spikes in validation loss for some stations, particularly Sylhet, suggesting temporary instability during optimisation. However, the validation loss subsequently decreases and stabilises, indicating recovery of the training process.
In relatively drier regions such as Rajshahi and Khulna, all models exhibit a relatively larger gap between training and validation loss curves compared to the gap in other stations. This behaviour may indicate comparatively weaker generalisation performance or higher variability in the underlying rainfall characteristics at these stations. On the other hand, the smaller gap in wetter regions indicates comparatively more stable validation performance, possibly because of more consistent or abundant data, allowing the models to capture underlying patterns more robustly.
Overall, the simple LSTM and GRU models demonstrate more efficient and stable learning, as evidenced by faster convergence and lower final loss values, whereas the stacked LSTM model shows comparatively less stable learning during the training period.
3.3 Simulation of Events
The performance of the models varied across rainfall regimes, as illustrated in the time series comparisons for extreme (Fig 7), moderate (Fig 8), and weak events (Fig 9).
During extreme events, all DL models closely tracked the ERA5 reference values in terms of timing, magnitude, and duration of peak rainfall, particularly at Dhaka, Sylhet, and Rangpur. For instance, during E01, DL models successfully captured the sustained rainfall peak at Dhaka exceeding 15 mm, and the multi-peak structure at Sylhet during E03. The DL predictions remained within a physically plausible range throughout all three extreme events. In contrast, WRF exhibited sharp, isolated spikes temporally misaligned with the ERA5 reference, producing peak values exceeding 100 mm at Chittagong during E01 and above 40 mm at Sylhet, while the ERA5 reference remained below 25 mm. These spikes indicate systematic overestimation and temporal displacement rather than a consistent underestimation bias.
During moderate events, DL models generally captured the overall rainfall pattern but showed some underestimation of peak intensities, particularly at Rajshahi and Rangpur where rainfall magnitudes were low. WRF continued to produce spurious high-intensity spikes largely decoupled from the ERA5 reference signal, with peaks exceeding 50 mm at Sylhet during E05 and E07 while ERA5 reference values remained near zero.
During weak events, ERA5 reference values were near zero across most stations and hours. DL models produced low but non-zero predictions in several cases, reflecting a tendency to overestimate under near-dry conditions. This is particularly visible at Sylhet and Rajshahi during E08 and E09, where Stacked LSTM and GRU generated notable phantom predictions against a flat ERA5 reference. WRF predicted near-zero rainfall throughout all weak events, appearing conservative but failing to detect any rainfall occurrence whatsoever.
3.4. Evaluation of models
An all-inclusive set of measures was utilized in determining the predictive quality of each model structure on the station level (Tables 4–7). Always, models had their best results in the wet, northeastern areas. In the case of the Simple LSTM, Chittagong performed best on the overall performance (R2 = 0.88, Accuracy = 0.89, F1-score = 0.86), and Sylhet came in second place.
This comparative analysis shows a major trade-off, with the Stacked LSTM having the lowest absolute MSE values at some stations, which is evidence that it is more able to capture non-linear relationships, but this is sometimes at the cost of a little greater generalization error, as indicated by its slightly lower R2 and F1-values at Sylhet and Rajshahi compared to the Simple LSTM.
GRU and simple LSTM models were very well balanced in terms of accuracy and stability. Simple LSTM showed marginally stronger results than GRU on the major measures such as Accuracy and F1-score in dominant stations with the exception of a few. But the result of GRU and Simple LSTM is outstanding. In all of the models, predictive accuracy declined monotonically according to the dry to wet gradient, with drier northwest locations (Rajshahi, Rangpur) being most difficult as shown by their lower R2 and F1-scores. The marginal differences in performance metrics across the four DL architectures do not conclusively establish the superiority of any single architecture due to the absence of formal significance. To address this, a Friedman test was conducted on station-level RMSE ranks, which revealed a statistically significant difference among the four architectures (χ² = 10.60, p = 0.014). A Nemenyi post-hoc test showed that Simple LSTM (mean rank = 1.33) achieved significantly lower RMS E ranks than Stacked LSTM (mean rank = 3.67; p = 0.0095), while no other pairwise comparison reached significance (all p > 0.05) (Table 8). GRU (mean rank = 2.17) and BiLSTM (mean rank = 2.83) occupied intermediate positions and were not statistically distinguishable from either the best- or worst-ranked architecture (Fig 10).
Models connected by a horizontal bar are not significantly different (Nemenyi test, α = 0.05).
Fig 11 presents scatter plots of ERA5 reference values against model predictions aggregated across all stations and events. Simple LSTM, Stacked LSTM, GRU, and BiLSTM achieved R² values of 0.840, 0.816, 0.838, and 0.805, respectively, with predictions clustering tightly around the 1:1 (Perfect Prediction) line. WRF produced an R² of −11.374, with predicted values reaching up to 200 mm against ERA5 reference values below 25 mm, confirming the systematic overestimation identified in Section 3.3. The WRF subplot uses a different axis scale to accommodate these extreme outliers.
The following table (Table 9) presents per-station performance metrics across all 10 events. Chittagong showed the strongest DL performance overall, with GRU achieving the lowest RMSE (0.613), MAE (0.277), and R² (0.920), while BiLSTM recorded the best F1-score (0.926) and accuracy (0.918). Sylhet also performed well, with Simple LSTM yielding the highest F1-score (0.889) and recall (0.969), and GRU achieving R² = 0.853 and RMSE = 0.954. At Khulna, Stacked LSTM led with R² = 0.837, F1 = 0.830, and the lowest RMSE (0.833).
Performance weakened considerably at Rajshahi and Rangpur. At Rajshahi, Simple LSTM achieved the best R² (0.707) but F1-score dropped to 0.632 with precision of only 0.522, reflecting difficulty in correctly identifying rainfall occurrence. At Rangpur, GRU recorded the best R² (0.453) and F1 (0.759), yet RMSE remained at 0.759, indicating substantial magnitude errors despite reasonable classification skill. At Dhaka, Simple LSTM achieved an R² of 0.856 and the lowest MAE (0.393), although the F1-score of 0.753 suggests a weaker detection skill compared to coastal stations. WRF performed poorly across all stations, with the worst regression performance at Chittagong (R² = −35.907, RMSE = 13.173, MAE = 2.855) and weakest classification skill at Rangpur (F1 = 0.329, recall = 0.211).
Notable performance differences among the four DL architectures are relatively small. For instance, R² values at Chittagong range from 0.918 to 0.923, and F1-scores differ by no more than 0.057 at any single station.
Table 10 presents metrics stratified by event category. During extreme events, Simple LSTM achieved the strongest regression performance, with the lowest MAE (0.695), lowest RMSE (1.418), and highest R² (0.812), while GRU recorded the highest F1-score (0.897), indicating a slight trade-off between magnitude accuracy and rainfall-occurrence detection among the top two models. WRF showed R² = −11.565, RMSE = 11.582, and MAE = 3.326, with recall dropping to 0.456 despite relatively high precision (0.849), indicating it detected some rainfall events but missed the majority.
During moderate events, DL model performance declined, with Simple LSTM recording R² = 0.488, RMSE = 0.387, and F1 = 0.721. WRF deteriorated sharply with R² = −112.756 and RMSE = 5.777, more than 14 times higher than Simple LSTM, reflecting near-complete decoupling from the ERA5 reference signal during these conditions.
During weak events, all DL models yielded negative R² values, with Stacked LSTM performing worst (R² = −0.407, F1 = 0.118, recall = 0.231). GRU showed the best classification skill with F1 = 0.244 and recall = 0.769, though precision remained low at 0.145. Accuracy was high across all DL models (0.929–0.949) but reflects the dominance of zero-rainfall timesteps rather than genuine detection skill. WRF recorded F1 = 0.000, precision = 0.000, recall = 0.000, and accuracy = 0.960, indicating it predicted no rainfall occurrence during weak events while appearing accurate due to the same class imbalance effect.
4. Discussion and conclusion
The ERA5 reanalysis dataset provided a consistent and temporally continuous basis for training deep learning models at hourly resolution across six climatologically diverse stations in Bangladesh. Although ERA5 is a model-based reanalysis product and not equivalent to ground observations, previous studies have demonstrated its ability to capture precipitation variability and temporal patterns in data-sparse tropical regions, supporting its use as a training input in the absence of high-frequency station data [25,26]. Consequently, the reported skill scores quantify how well each model reproduces ERA5 precipitation patterns, not how accurately they predict independently measured rainfall, a distinction that becomes particularly relevant when considering the operational applicability of these findings.
All four DL models demonstrated consistent predictive skill across the 10 evaluated events, showing closer agreement with ERA5 reference values than WRF across both regression and classification metrics. This is consistent with findings from previous studies showing that data-driven models can achieve higher skill scores than NWP systems against reanalysis-based references in short-term rainfall prediction tasks, particularly when trained on datasets that share the same statistical structure as the evaluation reference [9,10]. The superior agreement of DL models with ERA5 reference values can be partly attributed to the nature of the training framework: the models were trained directly on ERA5 reanalysis data and therefore learned the internal statistical relationships and spatial patterns embedded within that product. WRF, by contrast, operates through independent physical parameterizations that do not account for ERA5-specific biases or regional rainfall characteristics, making systematic disagreement between WRF and ERA5 reference values expected rather than exceptional.
Among the models evaluated, Simple LSTM achieved the best overall performance across the full test period, with the highest R² and lowest MAE, MSE, and RMSE at most stations, while GRU followed closely with comparable accuracy. Both models explained approximately 88% of the variance in ERA5 rainfall across the full test period (Tables 4–7), with low prediction errors and high classification accuracy. A Friedman test on station-level RMSE ranks confirmed that these performance differences were not uniform across all model pairs (χ² = 10.60, p = 0.014); a Nemenyi post-hoc test showed that Simple LSTM significantly outperformed Stacked LSTM (p = 0.0095), whereas the differences between Simple LSTM and GRU, and among the remaining pairs, did not reach statistical significance (p > 0.05, Table 8, Fig 10). This indicates that while Simple LSTM and GRU are statistically indistinguishable from each other, Simple LSTM’s advantage over Stacked LSTM is a genuine, non-random effect. The marginal differences between all four architectures, however, reflect the inherently similar capacity of recurrent models to capture hourly rainfall dynamics from ERA5 inputs. The relatively stronger performance of simpler recurrent architectures is consistent with findings in the rainfall and hydrological forecasting literature, where lightweight models have been shown to generalise more effectively than deeper architectures when training data are limited or highly variable [7,33].
Model performance showed clear spatial dependence linked to regional rainfall climatology. Chittagong and Sylhet consistently yielded the strongest results across all DL models, while Rajshahi and Rangpur showed the weakest performance. This pattern is consistent with the rainfall climatology of Bangladesh, where the northeastern and southeastern regions receive higher annual totals and more frequent intense events [55], providing richer and more balanced training signals. In contrast, the intermittent and low-intensity rainfall regime of northwestern Bangladesh results in greater class imbalance and fewer high-rainfall training samples, which plausibly limits the models’ ability to learn robust prediction patterns. This explanation remains a hypothesis, however, as a formal quantitative analysis linking training sample distribution to model performance was beyond the scope of this study.
The Stacked LSTM and BiLSTM models demonstrated comparatively less stable convergence and slightly weaker performance. The noisier loss curves of Stacked LSTM suggest that greater architectural depth does not necessarily improve performance for this forecasting task, which is consistent with observations in the broader deep learning literature that overly complex models can exhibit poorer generalisation with limited or highly variable training data [6]. BiLSTM showed smoother convergence than Stacked LSTM but showed signs of instability at Sylhet and Rangpur.
Event-category analysis revealed important performance variations across rainfall regimes. During extreme events, all DL models showed strong agreement with ERA5 reference values, with Simple LSTM achieving R² = 0.812 and F1 = 0.883, while GRU achieving R² = 0.809 and F1 = 0.897. Performance declined during moderate events, and all DL models yielded negative R² values during weak events, indicating they could not reproduce ERA5 patterns better than a simple mean predictor under near-dry conditions. The high accuracy scores during weak events (0.929–0.949) reflect the dominance of zero-rainfall timesteps rather than genuine detection skill, a direct consequence of class imbalance in low-rainfall periods. These results raise practical concerns about operational deployment: DL models trained on ERA5 show limited skill during near-dry conditions, which are common in northwestern Bangladesh for much of the year. A rainfall occurrence pre-screening step or probability threshold calibration is recommended to reduce false alarms during near-dry conditions in any future operational application in northwestern Bangladesh.
WRF performance deteriorated consistently across all event categories and stations, with the most severe overestimation at Chittagong (R² = −35.907, RMSE = 13.173) and the weakest classification skill at Rangpur (F1 = 0.329, recall = 0.211). The negative R² values for WRF indicate that its mean squared error exceeded that of a simple climatological mean baseline for the evaluated events, reflecting systematic overestimation and temporal displacement rather than outright model failure. These results are specific to the 10 selected events and the single WRF physics configuration used; a broader evaluation across more events and alternative parameterization schemes would be needed before generalising conclusions about WRF skill in Bangladesh.
Despite these promising results, this study has several limitations that merit consideration when interpreting the findings. First, models were trained and evaluated using ERA5 reanalysis data rather than independent ground observations, which limits the direct applicability of results to operational forecasting contexts. Second, the study is restricted to six urban stations and may not fully represent rural or complex terrain regions. Third, the WRF comparison is based on 10 selected events under a single physics configuration, and a more systematic multi-configuration evaluation would strengthen the benchmark.
On the whole, all DL models demonstrated substantially lower prediction errors and higher classification skill against the ERA5 reference than WRF across the evaluated events. Simple LSTM emerged as the strongest model, a difference confirmed as statistically significant against Stacked LSTM by the Friedman–Nemenyi test, while GRU produced nearly comparable results. Both architectures demonstrated computationally efficient and stable learning behaviour, making them promising candidates for further development and validation toward operational short-term rainfall forecasting in Bangladesh. These results represent a proof-of-concept pending validation against ground observations and real-world operational testing. Future research should incorporate radar observations, additional atmospheric predictors including pressure-level variables, and Transformer-based architectures to further improve and broaden the scope of the comparison, with particular focus on improving performance in the dry northwestern regions where current models show limited skill.
Acknowledgments
The authors would like to express their sincere gratitude to the Department of Meteorology, University of Dhaka, for providing the facilities and academic support required for this research.
References
- 1. Dash Y, Abraham A. Integrating empirical analysis and deep learning for accurate monsoon prediction in Kerala, India. Applied Computing and Geosciences. 2024;24:100211.
- 2.
IPCC. Climate Change 2021: The Physical Science Basis. Geneva, Switzerland: Cambridge University Press. 2021. https://www.ipcc.ch/report/ar6/wg1/
- 3. Ahasan MN, Chowdhury MAM, Quadir DA. Prediction of high impact rainfall events over Bangladesh using high resolution MM5 model. Sri Lankan J Phys. 2012;12(0):43.
- 4. Mannan MdA, Chowdhury MdAM, Karmakar S. Application of NWP Model in Prediction of Heavy Rainfall in Bangladesh. Procedia Engineering. 2013;56:667–75.
- 5. Nitha V, Pramada SK, Praseed NS, Sridhar V. Performance Evaluation of Numerical Weather Prediction Models in Forecasting Rainfall Events in Kerala, India. Atmosphere. 2025;16(4):372.
- 6. Barrera-Animas AY, Oyedele LO, Bilal M, Akinosho TD, Delgado JMD, Akanbi LA. Rainfall prediction: A comparative analysis of modern machine learning algorithms for time-series forecasting. Machine Learning with Applications. 2022;7:100204.
- 7. Di Nunno F, Granata F, Pham QB, de Marinis G. Precipitation Forecasting in Northern Bangladesh Using a Hybrid Machine Learning Model. Sustainability. 2022;14(5):2663.
- 8. Billah M, Adnan MdN, Akhond MR, Ema RR, Hossain MdA, Md. Galib S. Rainfall prediction system for Bangladesh using long short-term memory. Open Computer Science. 2022;12(1):323–31.
- 9. Schultz MG, Betancourt C, Gong B, Kleinert F, Langguth M, Leufen LH. Can deep learning beat numerical weather prediction?. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences. 2021.
- 10. Shirali E, Nikbakht Shahbazi A, Fathian H, Zohrabi N, Mobarak Hassan E. Evaluation of WRF and artificial intelligence models in short-term rainfall, temperature and flood forecast (case study). J Earth Syst Sci. 2020;129(1).
- 11. Hasan MA, Islam AKMS. Evaluation of Microphysics and Cumulus Schemes of WRF for Forecasting of Heavy Monsoon Rainfall over the Southeastern Hilly Region of Bangladesh. Pure Appl Geophys. 2018;175(12):4537–66.
- 12. Islam MdJ, Abdullah MS, Alam M. Flooding crisis in Bangladesh: urgent measures required. Biodiversity. 2024;25(2):95–8.
- 13. IFRC. Bangladesh: Flash flood – 2022-05. 2022. https://go.ifrc.org/field-reports/15815
- 14.
UN Women. Rapid gender analysis of flood situation in North and North-Eastern Bangladesh. UN Women – Asia-Pacific. 2022. https://asiapacific.unwomen.org/en/digital-library/publications/2022/07/rapid-gender-analysis-of-flood-situation-in-north-and-north-eastern-bangladesh
- 15. Rahman MM, Alam S, Shobuj IA, Santo MMH, Hossain MT, Rahman F, et al. Mental health consequences of the 2024 Feni flash flood in Bangladesh: prevalence and risk factors. Front Public Health. 2025;13:1687943. pmid:41439097
- 16.
Bangladesh: Eastern flash floods 2024 situation report No. 02. OCHA. 2024. https://www.unocha.org/publications/report/bangladesh/bangladesh-eastern-flash-floods-2024-situation-report-no-02-30-august-2024
- 17. WHO. Rising Waters, Rising Challenges - WHO’s Response to Severe Flooding in Bangladesh. https://www.who.int/bangladesh/news/detail/04-11-2024-rising-waters--rising-challenges-who-s-response-to-severe-flooding-in-bangladesh 2024. Accessed 2025 December 31.
- 18. Biswas JC, Haque MM, Maniruzzaman M, Ali MH, Kabir W, Kalra N. Natural hazards and livestock damage in Bangladesh. Nat Hazards. 2019;99(2):705–14.
- 19. Atiqul Haq SM, Arno AT, Aziz Lalin SA, Ahmed MNQ. Extreme Weather Events and Expected Parental Roles in Bangladesh and Beyond: A Review of Literature. Weather, Climate, and Society. 2024;16(2):257–74.
- 20. Liong S, Sivapragasam C. FLOOD STAGE FORECASTING WITH SUPPORT VECTOR MACHINES1. J American Water Resour Assoc. 2002;38(1):173–86.
- 21. Kabir MA, Chakma C. Redefining multi-target weather forecasting with a novel deep learning model: Hierarchical temporal convolutional long short-term memory with attention (HTC-LSTM-Attn) in Bangladesh. PLoS One. 2026;21(3):e0342431. pmid:41871152
- 22. Islam MS, Shafiuzzaman M, Mahmud G, Nowshin N, Reza P, Hasan J, et al. Explainable deep learning for rainfall prediction: A CNN-XGBoost hybrid approach in the northern region of Bangladesh. Neural Comput & Applic. 2025;37(33):28125–60.
- 23. Akinsehinde BO, Shang C, Shen Q. Achieving reliable rainfall forecasting through ensemble deep learning, fuzzy systems, and advanced feature selection. Int J Gen Syst. 2025.
- 24. Li W, Liu C, Hu C, Niu C, Li R, Li M, et al. Application of a hybrid algorithm of LSTM and Transformer based on random search optimization for improving rainfall-runoff simulation. Sci Rep. 2024;14(1):11184. pmid:38755303
- 25. Xu Y, Yang Z, Zhang F, Chen X, Zhou H. A rainfall prediction model based on ERA5 and Elman neural network. Advances in Space Research. 2025;75(2):1732–46.
- 26.
Li Y, Akrami N, Dev S. Harnessing ERA5 Reanalysis Data for Improved Long-Term Rainfall Forecasting in Southern Iran. In: 2023 IEEE 7th Conference on Energy Internet and Energy System Integration (EI2), 2023. 2045–9. https://doi.org/10.1109/ei259745.2023.10512483
- 27. da Silva ST, Milani LC, Gabrick EC, Iarosz KC, Viana RL, Caldas IL, et al. Rainfall forecast in Brazil using machine learning. Chaos. 2025;35(7):073116. pmid:40623172
- 28. Tanvir MA, Naiema N. Rainfall Prediction Using Machine Learning: LSTM. JCSIT. 2025;2(2):23–33.
- 29. Oliveira ECL de, Carvalho EC de, Jesus E dos S, Rocha R de L, Arruda HM, Alves RC de O, et al. A statistical and machine learning approach for monthly precipitation forecasting in an Amazon city. Front Earth Sci. 2025;13.
- 30.
Climate Hazards Center. CHIRPS3: Climate Hazards Center Infrared Precipitation with Stations version 3. Santa Barbara, CA, USA. 2025. https://chc.ucsb.edu/data/chirps3
- 31. Xiang Z, Yan J, Demir I. A Rainfall‐Runoff Model With LSTM‐Based Sequence‐to‐Sequence Learning. Water Resources Research. 2020;56(1).
- 32. He L, Yao Y, Xu C, Huan Z, Tang F, Ji C, et al. A New Global ZTD Forecast Model Based on Improved LSTM Neural Network. IEEE J Sel Top Appl Earth Obs Remote Sens. 2024;17: 9606–14.
- 33. Kratzert F, Klotz D, Brenner C, Schulz K, Herrnegger M. Rainfall–runoff modelling using Long Short-Term Memory (LSTM) networks. Hydrol Earth Syst Sci. 2018;22(11):6005–22.
- 34. Hersbach H, Bell B, Berrisford P, Hirahara S, Horányi A, Muñoz-Sabater J. The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society. 2020;146:1999–2049.
- 35. Haider S, Karim MR, Islam MS, Megumi TA, Rahnama QS. Extreme weather events and Spatio-temporal characterization of climate change variables in Bangladesh during 1975-2019. Heliyon. 2024;10(5):e27118. pmid:38455535
- 36. Ezaz GT, Zhang K, Li X, Shalehy MH, Hossain MA, Liu L. Spatiotemporal changes of precipitation extremes in Bangladesh during 1987–2017 and their connections with climate changes, climate oscillations, and monsoon dynamics. Global and Planetary Change. 2022;208:103712.
- 37. Khan MJU, Islam AKMS, Bala SK, Islam GMT. Changes in climate extremes over Bangladesh at 1.5 °C, 2 °C, and 4 °C of global warming with high-resolution regional climate modeling. Theoretical and Applied Climatology. 2020;140:1451–66.
- 38. Liu Y, Zhuo L, Han D. Developing spin-up time framework for WRF extreme precipitation simulations. J Hydrol (Amst). 2023;620.
- 39. Jin J, Miller NL, Schlegel N. Sensitivity study of four land surface schemes in the WRF model. Advances in Meteorology. 2010.
- 40. Skamarock WC, Klemp JB, Dudhia J, Gill DO, Barker DM, Duda MG. A description of the advanced research WRF version 4. 2019.
- 41. Baki H, Chinta S, Balaji C, Srinivasan B. Determining the sensitive parameters of WRF model for the prediction of tropical cyclones in Bay of Bengal using global sensitivity analysis and machine learning. Geosci Model Dev. 2021;15:2133–55.
- 42. Ratnam JV, Behera SK, Krishnan R, Doi T, Ratna SB. Sensitivity of Indian summer monsoon simulation to physical parameterization schemes in the WRF model. Clim Res. 2017;74:43–66.
- 43. Chawla I, Osuri KK, Mujumdar PP, Niyogi D. Assessment of the Weather Research and Forecasting (WRF) model for simulation of extreme rainfall events in the upper Ganga Basin. Hydrol Earth Syst Sci. 2018;22(2):1095–117.
- 44. Kain JS. The Kain–Fritsch Convective Parameterization: An Update. J Appl Meteor. 2004;43(1):170–81.
- 45. Sharma A, Sharma D, Panda SK, Kumar A. Sensitivity analysis of different parameterization schemes of the Weather Research and Forecasting (WRF) model to simulate heavy rainfall events over the Mahi River Basin, India. Agricultural and Forest Meteorology. 2024;346:109885.
- 46. Rodrigo C, Kim S, Jung I. Sensitivity Study of WRF Numerical Modeling for Forecasting Heavy Rainfall in Sri Lanka. Atmosphere. 2018;9(10):378.
- 47. Hong SY, Lim JO. The WRF single-moment 6-class microphysics scheme (WSM6). Asia Pac J Atmos Sci. 2006;42:129–51.
- 48. Khan MS, Alam M. Study on Sensitivity of Microphysics for the Simulation of Rainfall for the Month of May 2015 over Bangladesh using High Resolution WRF-ARW Model. J of Eng Sci. 2020;11(1):43–60.
- 49. Hong S-Y, Noh Y, Dudhia J. A New Vertical Diffusion Package with an Explicit Treatment of Entrainment Processes. Monthly Weather Review. 2006;134(9):2318–41.
- 50. Wu J, Lu J, Chen X. Performance of the WRF model in simulating convective rainfall events in the humid subtropical monsoon climate region—Poyang Lake basin. Theor Appl Climatol. 2023;153:889–911.
- 51. Ek MB, Mitchell KE, Lin Y, Rogers E, Grunmann P, Koren V, et al. Implementation of Noah land surface model advances in the National Centers for Environmental Prediction operational mesoscale Eta model. J Geophys Res. 2003;108(D22).
- 52. Mlawer EJ, Taubman SJ, Brown PD, Iacono MJ, Clough SA. Radiative transfer for inhomogeneous atmospheres: RRTM, a validated correlated‐k model for the longwave. J Geophys Res. 1997;102(D14):16663–82.
- 53. Dudhia J. Numerical study of convection observed during the winter monsoon experiment using a mesoscale two-dimensional model. J Atmos Sci. 1989;46:3077–107.
- 54. Shenoy M, Raju PVS, Prasad J. Optimization of physical schemes in WRF model on cyclone simulations over Bay of Bengal using one-way ANOVA and Tukey’s test. Sci Rep. 2021;11(1):24412. pmid:34952900
- 55. Demšar J. Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research. 2006;7(1):1–30.
- 56. Shahid S. Rainfall variability and the trends of wet and dry periods in Bangladesh. Intl Journal of Climatology. 2010;30(15):2299–313.
- 57. Jabir B, Falih N. Dropout, a basic and effective regularization method for a deep learning model: a case study. IJEECS. 2021;24(2):1009.