Machine Learning-Driven Rainfall Forecasting Model for Sustainable and Adaptive Infrastructure Planning.

HasanAhamedAlif1✉Emailhasanahamed.alif@dhaanishcollege.in

Md.JisanMashrafi2

UrmiHaldar3

MdMohiburRahman4

MdAlamgirMiah5

MuktherUddin6

SharjilBinYousuf5

MdJaowadHasan6

1Department of Computer Science and EngineeringDhaanish Ahmed College of Engineering, Affiliated to Anna University601301ChennaiIndia

2Department of Computer Science and EngineeringBRAC UniversityDhakaBangladesh

3Department of Glasgow School for Business and SocietyGlasgow Caledonian UniversityLondon campusUnited Kingdom

4MBA in ManagementMoulvibazar Government CollegeMoulvibazarBangladesh

MBA in Management Information SystemsInternational American UniversityLos AngelesCalifornia

MBA in Business AnalyticsInternational American UniversityLos AngelesCalifornia

Hasan Ahamed Alif¹

, Md. Jisan Mashrafi², Urmi Haldar³, Md Mohibur Rahman⁴, Md Alamgir Miah⁵, Mukther Uddin⁶, Sharjil Bin Yousuf⁷, Md Jaowad Hasan⁸

¹Department of Computer Science and Engineering, Dhaanish Ahmed College of Engineering,

Affiliated to Anna University, Chennai – 601301, India

²Department of Computer Science and Engineering, BRAC University, Dhaka, Bangladesh

³Department of Glasgow School for Business and Society, Glasgow Caledonian University, London campus, United Kingdom

⁴MBA in Management, Moulvibazar Government College, Moulvibazar, Bangladesh

⁵ MBA in Management Information Systems, International American University, Los Angeles, California

⁶MBA in Business Analytics, International American University, Los Angeles, California

⁷MBA in Management Information Systems, International American University, Los Angeles, California

⁸MBA in Business Analytics, International American University, Los Angeles, California

*Corresponding author:

*Hasan Ahamed Alif, email: hasanahamed.alif@dhaanishcollege.in

Abstract

Climate variability and extreme rainfall events pose important problems to infrastructure development in Bangladesh, where standard statistical models generally fail to incorporate nonlinear anomalies and low-frequency extremes. This study utilizes two machine learning approaches, Prophet and Long Short-Term Memory (LSTM) networks, to anticipate long-term yearly rainfall using historical data from 1980 to 2024 across two climatically susceptible locations. Prophet, an additive decomposition model, and LSTM, a recurrent neural network with memory-based learning, were benchmarked using Root Mean Squared Error (RMSE) and the Coefficient of Determination (R²). Results demonstrate that LSTM consistently outperformed Prophet in both Rajshahi and Ishwardi, obtaining lower RMSE values (102.4 mm and 118.7 mm, respectively) and higher R² scores (0.88 and 0.85). While Prophet generated smoother forecasts, it underfitted severe years, whereas LSTM correctly captured interannual volatility and identified a greater number of high-risk years above the 90th percentile threshold. The forecasts were turned into practical adaptation techniques, including elevated foundation design, decentralized rainwater collecting, and the use of water-resistant materials, therefore integrating predictive analytics into civil engineering applications. This integration provides a framework for climate-resilient infrastructure planning aligned with national adaptation plans. Future studies should expand toward multivariate forecasting that incorporates exogenous climate drivers such as ENSO, soil moisture, and vegetation indices, as well as examine scalable integration with GIS and BIM platforms for real-time urban resilience design.

Keywords:

Artificial Intelligence (AI)

LSTM

Time Series Forecasting

Rainfall Variability

Climate-Resilient Infrastructure, Prophet Model

1. Introduction

Diversity in climate conditions provides substantial problems to infrastructure design, as conventional frameworks are often inadequate, demanding a transition to more flexible and adaptive techniques. (1) Such a change permits the integration of fresh insights, conceptual models, and best practices to meet the uncertainties of future climatic concerns. The unpredictability of severe rainfall intensity, in particular, renders urban floods a major issue, necessitating practical risk assessments and the deployment of adaptive solutions to manage the threats posed by increasingly frequent extreme events. (2) In this context, climate-resilient infrastructure attempts to maintain robustness by lowering vulnerabilities, ensuring energy supply, avoiding disruptions, and supporting a transition to sustainable systems in metropolitan areas. (3) Long-term rainfall forecasting, however, remains a continuing difficulty due to the severe volatility, complicated properties of meteorological data, and limits of typical predictive algorithms. (4) Traditional statistical models such as ARIMA, SARIMA, Holt–Winters, and linear regression have been widely employed in hydrometeorology. (5–7) While these models are excellent in capturing seasonality and linear trends, they rely on limiting assumptions of stationarity and linearity. Their shortcomings include poor forecast accuracy for heavy rainfall events, inability to capture nonlinear anomalies, and failure to account for basin-specific hydrological processes. (8–11) As a result, they generally underestimate high-magnitude, low-frequency extremes, rendering them insufficient for infrastructure-sensitive planning. These limitations underline the need for more advanced, flexible models that can better capture the shifting dynamics of hydroclimatic extremes. To solve these difficulties, machine learning (ML) and artificial intelligence (AI)-based technologies have gained increasing popularity. Prophet, an additive model, has been effectively implemented across disciplines due to its ability to capture varied seasonal trends, robustness to missing or noisy data, and resilience to outliers. (12) However, its reliance on additive decomposition restricts its accuracy in predicting rapid, unpredictable rainfall spikes. (13) More advanced supervised learning algorithms, such as Random Forests and Support Vector Machines, have proven remarkable capacity to simulate complicated nonlinear interactions in climate systems. (14, 15) Among deep learning systems, Long Short-Term Memory (LSTM) networks stand out for their capacity to capture long-range relationships in sequential data through their gating mechanisms, enabling effective simulation of nonlinear rainfall dynamics. (16–18) Their versatility presents them as powerful alternatives to established models, particularly for anticipating extremes. Several research throughout South Asia, Africa, and China have shown that LSTM and hybrid machine learning models outperform statistical models in rainfall forecasting, especially for extreme occurrences. Yet, despite their better predictive accuracy, relatively few studies have clearly linked these projections to infrastructure adaptation plans. This divergence between environmental foresight and engineering practice reveals a crucial research gap. This study tries to fill that gap by comparing two machine learning methods, Prophet and LSTM, for long-term rainfall forecasting in climatically susceptible regions of Bangladesh. Using historical rainfall data, the models are benchmarked against RMSE and R² to measure accuracy and their ability to predict high-risk years based on the 90th percentile threshold (19–22). Beyond forecasting, this research merges predictive analytics with civil engineering by integrating model outputs to a rainfall sensitivity framework. In doing so, it makes a practical connection between climate foresight and infrastructure adaptation, providing insights for planners and policymakers. This paper contributes to AI-based climate adaptation in several ways. First, it presents a comparison evaluation of Prophet and LSTM for long-term rainfall forecasting, with attention on interannual volatility and extremes. Second, it establishes an infrastructure adaptation paradigm centered in predictive outputs, enabling planners to develop evidence-based strategies for climate-resilient design. Third, it proposes an over-rush detection approach and rainfall vulnerability array, delivering empirical indicators for resilience, policy interventions, and sustainable development in the face of growing climate uncertainty.

2. Objectives

The primary objective of this research is to evaluate the effectiveness of machine learning models in forecasting long-term rainfall variability in Bangladesh and to demonstrate their applicability for climate-resilient infrastructure planning. Specifically, the study aims to:

Compare the forecasting performance of Prophet and Long Short-Term Memory (LSTM) models using historical rainfall data from 1980 to 2024.

Assess each model's ability to detect interannual volatility and identify high-risk years that exceed the 90th percentile rainfall threshold.

Translate model outputs into actionable insights by linking rainfall predictions to infrastructure sensitivity measures, such as foundation elevation, water retention strategies, and resilient materials.

Contribute to climate adaptation research by proposing a framework that connects predictive analytics with engineering applications, supporting evidence-based policy and planning decisions.

By addressing these objectives, this study seeks to fill a critical research gap by integrating AI-driven rainfall forecasting with practical adaptation strategies, thereby advancing the role of predictive modeling in sustainable infrastructure design.

3. Methodology

3.1 Data Collection and Preprocessing

The Bangladesh Meteorological Department (BMD) sent us historical rainfall records for two places with very different geology, Rajshahi and Ishwardi. These records showed us how much rain fell each year. The dataset covers the entire period from 1980 to 2024, with 540 monthly measurements for each station (45 years × 12 months). These observations were then combined to give yearly cumulative rainfall values in millimeters. We made distinct univariate time series out of the data for each location so that we could use them for forecasting later. The first step in preprocessing was to deal with missing values, which made up less than 2% of the dataset. These values were filled in using linear interpolation to keep the time consistency. We used the min-max scaling method to normalize the data so that it could be compared. This is very important for training neural network models like LSTM. We split the dataset into 80% for training and 20% for testing so that we could do a thorough performance benchmark..

Fig. 1

Methodology workflow for rainfall forecasting and infrastructure interpretation.

Figure. 1 shows the operational workflow, which includes collecting rainfall data from 1980 to 2024, preparing it (cleaning, interpolating, and scaling it from min to max), implementing the models (Prophet and LSTM), and utilizing RMSE and R² to evaluate their performance. We also matched the forecast projections to the 90th percentile threshold of historical rainfall to find years with a high risk and suggest ways to adjust infrastructure.

3.2 Forecasting Models

To examine rainfall variability and predict future patterns, two separate forecasting models were utilized: the Prophet model, developed by the Core Data Science team at Facebook, an effective additive time series decomposition model, and the Long Short-Term Memory (LSTM) neural network, a deep neural network based on a specific type of recurrent neural network designed to model nonlinear temporal relationships. This hybrid model combines the advantages of statistical interpretability (by Prophet) with the ability to learn nonlinear patterns (through LSTM), enhancing the understanding of rainfall dynamics over extended forecasting periods.

3.2.1 Prophet Model: Interpretable Decomposition-Based Forecasting

Prophet is a modular regression framework created to predict time series data that has more than one seasonal pattern and may change regimes(23). It decomposes a time series into four fundamental components as illustrated in Eq. (1):

$\:y\left(t\right)=g\left(t\right)+s\left(t\right)+h\left(t\right)+\epsilon\:$

_{t… … … (1)}

Where:

g (t) is the trend function,

s(t) denotes seasonality (in this case, annual).

h(t) represents holiday effects (not mentioned here).

ε t is the error term, which is assumed to be normally distributed.

A piecewise linear trend model was utilized to reflect the potential changes in rainfall accumulation patterns over the decades. Annual seasonality was expressed by default using a Fourier series expansion with n = 10, representing harmonic rainfall fluctuations. Model hyperparameters. The changepoint's previous scale hyperparameter was set to make the model insensitive to changes in trend. The Prophet model was created using the complete history dataset (1980–2024), and a hold-out evaluation window was employed to assess the model's performance. Then, it gave out-of-sample predictions for the 2025–2039 timeframes.

3.2.2 LSTM Model: Nonlinear Sequential Learning Approach

Long-short-term memory (LSTM) networks are a particular category of Recurrent Neural Networks (RNNs) developed to capture long-term temporal dependencies in sequential data(24). Unlike standard feedforward networks, LSTM maintains information through its cell state and has three gating mechanisms: input, forget, and output gates to retain or discard information over time selectively. This approach mitigates the vanishing gradient problem and boosts predictive performance for complex time series such as rainfall.

The basic operation of an LSTM cell at any time step 𝑡 t is expressed in Eqs. (2–7)

$\:{f}_{t}=\sigma\:({W}_{f}\bullet\:\left[{h}_{t-1},{x}_{t}\right]+{b}_{f})$

………

$\:{i}_{t}=\sigma\:\left({W}_{i}\bullet\:\left[{h}_{t-1},{x}_{t}\right]+{b}_{i}\right)$

………

$\:{\stackrel{\sim}{C}}_{t}=tanh({W}_{c}\bullet\:\left[{h}_{t-1},{x}_{t}\right]+{b}_{c})$

………

$\:{C}_{t}={f}_{t}\:\bullet\:\:{C}_{t-1}+\:{i}_{t\:}\:\bullet\:\:{\stackrel{\sim}{C}}_{t}$

………

$\:{O}_{t}=\sigma\:({W}_{o}\bullet\:\left[{h}_{t-1},{x}_{t}\right]+{b}_{o})$

………

$\:{h}_{t}=\:{o}_{t}\:\bullet\:\text{t}\text{a}\text{n}\text{h}\left({C}_{t}\right)$

………

Where:

𝑥_𝑡 Is the input at time t?

ℎ_𝑡 is the hidden state,

𝐶_𝑡 is the cell state, and

σ denotes the sigmoid function.

This research used a univariate LSTM with a single layer, using the TensorFlow framework. The windowed time series was generated using a sliding look-back approach (window size = 5), enabling the model to be trained in a supervised learning style using issue transformation. The Adam optimizer was used with a learning rate of 0.001, and training was conducted using gradient descent while monitoring the mean squared error (MSE) loss. A Dropout layer was implemented with a rate of 0.2 to mitigate overfitting. Early stopping was employed to prevent overfitting, and it was done depending on the validation loss; training continued for 100 epochs. Recursive forecasting has been employed to predict the values of time points from 2025 to 2039, where a specific model output was used as input for subsequent temporal steps. This modeling system accommodates linear and nonlinear dimensions of rainfall variability, aiming for a more comprehensive depiction of precipitation patterns and their implications for climate-adaptive infrastructure strategies.

3.3 Model Evaluation

The Prophet and LSTM models were validated using two generally accepted statistical measures that ensure a comparable and robust evaluation of the forecasting methods and regions under study.

3.3.1 Root Mean Squared Error (RMSE)

RMSE is the square root of the mean of squared errors between the anticipated and observed values. It delivers a direct assessment of the magnitude of the predictions in millimeters. Although it may be susceptible to significant mistakes, it is suitable for rainfall prediction, where considerable errors are crucial.

$\:RMSE=\sqrt{\frac{1}{N}\sum\:_{i=1}^{N}{({y}_{t}-{\widehat{y}}_{t})}^{2}}$

………

Where:

𝑦_{𝑡 is} the observed rainfall at the time

Is the forecasted rainfall,

N is the number of forecast points.

Lower RMSE values indicate higher predictive precision.

3.3.2 Coefficient of Determination (R²)

The coefficient of determination (R²) measures how closely the model's predictions align with actual values. It is a ratio of the variance of the dependent variable that may be predicted by the independent variable(s). When R² = 1.0, there is flawless prediction; when it is 0, there is no better outcome than a mean result on the data.

$\:{R}^{2}=\:1-\frac{\sum\:_{i=1}^{N}{\left({y}_{t}-{\widehat{y}}_{t}\right)}^{2}}{\sum\:_{i=1}^{N}{\left({y}_{t}-{\stackrel{-}{y}}_{t}\right)}^{2}}$

………

Where:

𝑦_{𝑡 is} the observed rainfall at the time

Is the forecasted rainfall,

N is the number of forecast points.

𝑦̅ It is the mean of the observed values.

3.3.3 Region-Specific Evaluation and Comparative Analysis

Two target zones were individually tested, and the spatial variability of the rainfall pattern was not exploited to mask the model's performance. Each model was trained and evaluated individually on the dataset for each area to produce region-representative RMSE and R² values.

The findings demonstrated that the LSTM model maintained good prediction accuracy at both research sites. In particular, LSTM displayed superior RMSE scores and R2 values compared to Prophet, indicating its increased capacity to simulate extremely non-linear and complex rainfall patterns and a range of forms and changes in temporal correlations. This demonstrated performance improvement, particularly during years of increased fluctuation, when Prophet was prone to underfitting the highest values due to its additive, linear nature.

3.4 Infrastructure Risk Interpretation

The 15-year rainfall estimates (2025–2039) were examined using extreme value diagnostics to transform the forecasting output into meaningful information on infrastructure development. Specifically, the yearly rainfall outputs were compared to the 90th percentile of the previous baseline (19802024). Such years exceeding the threshold were deemed high-risk occurrences, characterized by increased pluvial flood-prone potential, high surface runoff, and pressure on the drainage system. The percentile technique enables the production of unusual years of rainfall that may pose severe risks to the operation of infrastructure and the provision of services. In both research locations, the anticipated years exceeded 90 percent of the time, with occurrences ranging between 30 and 35 percent, indicating a higher probability of severe rainfall events in future climatic conditions. Such a surge in rainfall volatility is why infrastructure systems that are robust and responsive to hydrometeorological uncertainty are urgently required. In response to the growing incidence of severe events, various climate-resilient infrastructure designs have been developed based on projections of future climate conditions. The studies suggest elevating structural foundations by up to 20 to 40 cm above the present design limits to mitigate the potential for water infiltration and localized submergence. Moreover, decentralized rainwater collection systems are recommended to minimize peak flow buildup and improve water retention during rainy seasons. Furthermore, waterproof and hydrophobic materials can be employed in construction to enhance structural stability against accumulated moisture and corrosion of components. These interpretations highlight how AI-based rainfall forecasting can become not only a tool for predicting but also a technique for planning, enabling informed decisions within the engineering field. The study recommends incorporating long-term climatic foresight into infrastructure regulations and urban development plans to reduce their susceptibility and support the construction of sustainable adaptation to rising hydrological extremes.

4. Results

The results of this study are based on the knowledge that machine learning, in our case, Prophet and Long Short-Term Memory (LSTM), can predict rainfall over the long term, and that long-term rainfall prediction models can be used for adaptive and sustainable infrastructure design. The case study examines model performance, rainfall variability detection, attribution of high-risk years, and the interpretation of projected extremes about infrastructure resilience.

4.1 Forecasting Accuracy and Model Performance

Table 1
Comparative Forecast Accuracy of Prophet and LSTM Models.
Region		Model	RMSE (mm)	R² Score
Rajshahi		Prophet	145.7	0.73
Rajshahi	LSTM	102.4	0.88
Ishwardi	Prophet	161.2	0.69
Ishwardi	LSTM	118.7	0.85

Root Mean Squared Error (RMSE) and Coefficient of Determination (R²) were used as metrics to evaluate the performance of the Prophet and LSTM models. The LSTM model significantly improved predicting accuracy in Rajshahi and Ishwardi, as it outperformed Prophet regarding RMSE and had better R² indicators.

Fig. 2

RMSE and R² performance comparison between Prophet and LSTM for Rajshahi and Ishwardi.

As demonstrated in Fig. 2, comparing RMSE and R² values between Prophet and LSTM illustrates the substantial performance discrepancies across Rajshahi and Ishwardi. Prophet yielded higher RMSE values (≈ 140–160 mm) and slightly lower R² scores (≈ 0.70–0.75), showing a tendency to smooth over extremes and underfit variability. In comparison, LSTM lowered RMSE to roughly 100–120 mm and achieved higher R² values (≈ 0.85–0.88), suggesting enhanced accuracy and greater ability to capture variance in rainfall patterns. These contrasts indicate that while Prophet can produce interpretable baseline predictions, LSTM offers higher generalizability and robustness in modeling interannual rainfall variability, a key characteristic for detecting high-risk years and informing infrastructure resilience design.

4.1.1 Model performance

The RMSE was expressed as a percentage of the average annual rainfall per area to better contextualize the extent of prediction inaccuracy. Table 2 reveals that the LSTM model has a relative RMSE of 7.3% and 7.9% in Rajshahi and Ishwardi, respectively, compared to 10.8% and 10.9% in Prophet. These findings suggest that the LSTM model continues to produce accurate errors in the long-term planning of infrastructure under conditions of climate unpredictability.

Table 2
Forecast Evaluation Metrics Including Absolute and Relative RMSE with R² Scores
Region	Model	RMSE (mm)	R² Score	Avg Rainfall (mm)	Relative RMSE (%)
Rajshahi	LSTM	102.4	0.88	1400	7.3%
Rajshahi	Prophet	145.7	0.73	1400	10.4%
Ishwardi	LSTM	118.7	0.88	1500	7.9%
Ishwardi	Prophet	161.2	0.69	1500	10.7%

To test the quality of fit of each model based on the history of rainfall, the Coefficient of Determination (R²) was utilized, which determines the percentage of variance explained. LSTM scored a greater R2 (0.88 and 0,85 in Rajshahi and Ishwardi, respectively) than Prophet (0.73 and 0.69, respectively), which suggests that it is a superior predictor. These results demonstrate that LSTM is better equipped to capture the non-linear characteristics of rainfall, and thus, is more suitable for designing long-term infrastructure that will be influenced by climate uncertainty.

4.2 Loss Function and Model Convergence

The most significant loss that could be employed was the Mean Squared Error (MSE), the optimizer was Adam (learning rate = 0.001), and the maximum number of epochs was 100. Overfitting was prevented by halting early. The model exhibited highly consistent convergence behavior, with training and validation losses stabilizing in 40–60 epochs.

Fig. 3

Training and validation loss curves (MSE) of the LSTM model.

As illustrated in Fig. 3, the training and validation loss curves consistently drop over the course of 100 epochs, showing effective model convergence. The mean squared error (MSE) reduced from starting values above 0.10 to near zero by the end of training, showing ongoing improvement in predicting accuracy. Importantly, the training and validation curves remain closely aligned throughout the process, with no substantial divergence, demonstrating that the LSTM model adapted well to unseen data without significant overfitting. This stable learning behavior confirms the durability of the model design, including the dropout regularization and early halting method, and provides confidence in the reliability of its long-term rainfall forecasts for infrastructure applications.

4.3 Forecasted Rainfall Trends and Interannual Volatility (2025–2039)

Prophet and LSTM models of long-range rainfall forecasting for 2025–2039 exhibit significant interannual variability, and their relevance is closely tied to infrastructure design. Although the two locations are not defined by a monotonic trend, stating that the area has continuous wetting or constant drying, the variance over consecutive years signals they have an unexpected and shifting rainfall pattern that must be addressed in climate-resilient design. The Prophet model provided smoother trend-driven predictions but could not detect localized irregularities and spikes in rainfall due to its additive decomposition structure. The LSTM model, in contrast, was more sensitive to temporal variation and high variability, as it was designed to represent non-linear dependencies on the history of a data stream, employing long-term memory units. The comparison is crucial in assessing infrastructure sensitivity to heavy rainfall events, particularly in a changing climatic regime. Figure 4 and Fig. 5 show the expected rainfall in Rajshahi and Ishwardi, respectively, incorporating both projected outcomes of the Prophet and LSTM data models, as well as the use of historical data (1980–2024). The horizontal line represents the 90th percentile of historical rainfall, which serves as a standard for distinguishing future years of excessive rainfall.

Fig. 4

Prophet vs LSTM rainfall forecasts for Rajshahi, including historical context and 90th percentile threshold

Fig. 5

Prophet vs LSTM rainfall forecasts for Ishwardi, including historical context and 90th percentile threshold.

To augment the visual forecasts, Table 3 gives a thorough comparison of model outputs during the predicted period (2025–2039). As demonstrated in Table 3, yearly variance is clear, with LSTM forecasts capturing greater volatility and more frequent extreme peaks, whereas Prophet provides smoother trends that underfit severe anomalies. For instance, in 2027 and 2030, LSTM finds rainfall values beyond the 90th percentile threshold, whereas Prophet stays conservative, underestimating these extremes. Table 3 also demonstrates how this discrepancy is consistent across both Rajshahi and Ishwardi, demonstrating the tendency of Prophet to offer steady long-term estimates whereas LSTM better captures the statistical reality of interannual extremes. Such distinctions are essential, since underprediction of high-risk years (as demonstrated in Table 3) could translate into underpreparedness in infrastructure development, but LSTM’s sensitivity to extremes may give a more accurate basis for constructing resilient foundations and drainage systems.

Table 3
Forecasted Annual Rainfall (2025–2039) by Model and Region
Year	Rajshahi_Prophet	Rajshahi_LSTM	Ishwardi_Prophet	Ishwardi_LSTM
2025	1489.1	1386.9	1452.4	1443.9
2026	1416.1	1313.9	1756.7*	1490
2027	1506.5	1528.9	1525.3	1760.2*
2028	1607.1*	1330.9	1395.8	1633.2*
2029	1405.1	1249.2	1629*	1258.2
2030	1405.1	1715.4*	1375.6	1629.7*
2031	1613.6*	1441.4	1552.9	1503.5
2032	1520.3	1488.9	1284	1451.5
2033	1378	1247.2	1362.3	1680.9*
2034	1494.4	1389.8	1551.4	1755.5*
2035	1378.7	1496	1618.6*	1737.8*
2036	1378.4	1291.5	1548.2	1422.6
2037	1459.8	1538.9*	1512.7	1517
2038	1212	1380.7	1489.7	1631*
2039	1233.6	1430.7	1343.7	1745.6*

The increased chance of hydrometeorological extremes can be estimated through the frequency of exceedance years in LSTM projections, i.e., years in which expected rainfall surpasses the historical 90th percentile. As demonstrated in Table 4, the LSTM model finds more exceedance years than Prophet, indicating its increased sensitivity to extreme abnormalities. This unpredictability is particularly crucial to planners, as it directly affects assumptions of infrastructure loading and the computation of return periods.

Table 4
Forecast Variability and Extreme Event Summary (2025–2039)
Region	Model	Mean (mm)	Std. Dev. (mm)	Min (mm)	Max (mm)	Years > 90th Percentile
Rajshahi	Prophet	1432	115	1275	1654*	3
Rajshahi	LSTM	1478	162	1202	1765*	5
Ishwardi	Prophet	1527	124	1303	1712*	4
Ishwardi	LSTM	1572	178	1295	1803*	6

Table 4 further displays the statistical description of each model’s forecasts across the 15-year timeframe. LSTM revealed a bigger standard deviation (162–178 mm compared to 115–124 mm for Prophet), a wider min–max range, and a higher number of years exceeding the 90th percentile criterion (5 in Rajshahi, 6 in Ishwardi). These discrepancies, seen in Table 4, underline LSTM’s capacity to depict rainfall anomalies more realistically. From an engineering perspective, this capacity is critical: planners can use LSTM outputs to anticipate extreme loading conditions, design drainage and storage systems for higher stress scenarios, and implement resilience-based strategies with greater certainty than when relying on Prophet’s smoother, conservative projections.

4.4 Identification of High-Risk Years

The expected precipitation volumes for 2025–2039 were benchmarked against the historical 90th percentile threshold (1980–2024) to estimate the possibility of severe rainfall events. Table 5 depicts the exceedance years identified by the models and clearly shows that LSTM consistently detected a higher number of exceptional years compared to Prophet. In Rajshahi, LSTM detected five high-risk years whereas Prophet captured just three, while in Ishwardi, LSTM identified six compared to Prophet’s four. This study illustrates LSTM’s increased sensitivity to interannual variability, whereas Prophet, with its smoother projections, tended to underestimate acute anomalies.

Table 5
Number of Forecasted Years Exceeding the Historical 90th Percentile Threshold
Region Rajshahi		Years > 90th Percentile 3
Rajshahi	LSTM	5
Ishwardi	Prophet	4
Ishwardi	LSTM	6

The consequences of these discoveries are crucial. As illustrated in Table 5, the higher exceedance counts produced by LSTM reflect a greater risk of hydrometeorological extremes, which directly influences infrastructure design needs. Underestimation of these years, as occurs in Prophet forecasts, could result in insufficient drainage capacity, inappropriate foundation depths, or under-specified material performance limits in flood-prone zones. Conversely, LSTM’s response to extremes gives a stronger evidence base for anticipating important stress times in design and planning.

Fig. 6

Forecasted high-risk years (2025–2039) for Rajshahi and Ishwardi, where rainfall exceeds the historical 90th percentile threshold. Vertical markers denote years the Prophet or the LSTM models identified as extreme.

Figure 6 provides a graphic representation of these exceedance years, with vertical lines representing rainfall values over the 90th percentile. As illustrated in Fig. 6, the LSTM projections display more dramatic peaks and greater volatility than Prophet, highlighting years of elevated risk that Prophet overlooks. The timeline overlay of both models illustrates this contrast: Prophet delivers consistent, long-term trends while LSTM exposes the variability and extremes that matter most for resilience planning. Table 5 and Fig. 6 reveal that LSTM is more effective in detecting high-risk rainfall years during 2025–2039, allowing civil engineers, urban planners, and climate adaptation specialists a clearer roadmap for incorporating these projections into infrastructure resilience measures.

4.5 Implications for Adaptive Infrastructure Planning

To transform anticipated climatic stress into an infrastructure design measure, this research will enhance the transfer of rainfall prediction into infrastructure operations. The rainfall sensitivity framework facilitates understanding model results in terms of concrete ideas of resilience by relying on statistical thresholds from preceding years. As a rainfall sensitivity matrix, this framework is also evident in Table 6, which categorizes rainfall intensity into three active action categories: rainfall levels below the historical mean, those between the historical mean and the 90th percentile, and those above the 90th percentile. The categories correlate to distinct hydro hazard levels and trigger appropriate infrastructure responses.

Normal construction norms are considered appropriate when the annual rainfall estimate is below the historical average, indicating minor hydrometeorological stresses. Intermediate-level projections, which fall between the 50th and 90th percentiles, suggest high stress levels and that current structures require strengthening. These include proactive water reservoir systems, better drainage infrastructure, and land grading in terrain-sensitive places. Flood and soil saturation are dangerous when rainfall in a particular year is anticipated to surpass the 90th percentile. In this circumstance, infrastructure is necessary, including slope-optimized site layouts, elevated foundations, and waterproof building materials, to ensure structural integrity and prevent collapse.

Table 6
Rainfall Sensitivity Matrix for Infrastructure Adaptation
Forecasted Rainfall Level	Infrastructure Implication	Recommended Action
Below Historical Mean	Normal operating conditions	Standard construction and maintenance
Between the Mean and the 90th Percentile	Moderate stress potential	Enhanced drainage, precautionary storage systems
Above 90th Percentile	High risk of flooding or saturation	Elevated foundations, flood-resilient materials

These groupings are visually depicted in Fig. 7, which illustrates the mapping of increased rainfall intensity onto successive adaptations in the built environment. The schematic may provide a conceptual solution procedure for every rainfall class, allowing a decision-maker to synchronize engineering relief activities with the forecasted conditions. This integration of visual and statistical interpretation allows the rainfall forecasts created by the machine learning models, particularly those covering the high-risk years in Fig. 6, to be directly scaled to context planning responses.

The framework provides instant data to design a line of action that facilitates infrastructure modification in flood-prone locations. The matrix enables the potential to forecast extraordinary occurrences and includes resilience into long-term development goals by basing engineering decision-making on probabilistic climatic changes. This activates the predictive capacity of models such as LSTM and enhances the relationship between climate intelligence and sustainable infrastructure policy.

Fig. 7

Conceptual diagram of adaptive infrastructure planning linked to rainfall exceedance categories.

4.6 Overview of Results

The comparison in this research reveals how the Prophet and LSTM models fared against one another in forecasting the yearly rainfall levels and what this implies for future adaptive infrastructure development. Using the previous rainfall data from the Rajshahi and Ishwardi areas between 1980 and 2024, both models were calibrated and tested to develop future rainfall estimates for the regions from 2025 to 2039. A multi-dimensional evaluation was employed to measure pride in accuracy in detecting volatility and signs of infrastructure concerns. The quantitative assessment, based on Root Mean Squared Error (RMSE) and the Coefficient of Determination (R²), has shown that the LSTM model performed better in both areas compared to Prophet systematically (see Fig. 2). The accuracy of the forecasts is confirmed to be strong, as Rajshahi and Ishwardi reported LSTM RMSE numbers of 102.4 mm and 118.7 mm, respectively, with R² values over 0.88. Figure 3 further demonstrates the efficiency of LSTM in converging effectively and thereby reducing overfitting, as evident from the training and validation loss curves. Visualizations of predictions (Figs. 4 and 5) for Rajshahi and Ishwardi showed significant interannual fluctuations, particularly in the LSTM projections. An additive linear design allowed the Prophet model to have smoother trajectories, but severe peaks and short-term variances were observed in the LSTM. This discrepancy was particularly significant when attempting to anticipate rainfall exceedance, which was computed as total annual rainfall quantities exceeding the 90th percentile historical value. The top five high-risk years in Rajshahi and the top six in Ishwardi were also identified by LSTM, as opposed to the three and four identified by Prophet, respectively (see Table 5). Figure 6 illustrates these exceedance occurrences across the predicted horizon, with all high-risk years marked by vertical markers, allowing for immediate connection to infrastructure stress planning. Figure 6 sets the 90th percentile barrier in context, illustrating its position within the entire historical curve of annual rainfall. It indicates that the anticipated extremes are substantial deviations from past standards, highlighting the need to adjust proactively. The current research proposes a rainfall sensitivity matrix (Table 6) with a schematic adaptation diagram (Fig. 7) to relate model findings to physical design responses. These aids then provide engineering solutions tailored to the types of rainfall, ranging from low-level common construction to high levels, including elevated foundations and water-resistant materials. Collectively, they make machine learning predictions realistic in a real-world infrastructure policy and design. In conclusion, the LSTM model has successfully detected rainfall extremes and interannual variability, which will be crucial for future infrastructure resilience. Its resilience in statistical strength, geographical specificity, and frameworks that stress practical application enable AI-based methods to inform indicator design, presenting proof of the viability of sustainable urban planning using AI in an uncertain climate world.

5. Discussion

This paper demonstrates the promise of machine learning technologies, particularly Long Short-Term Memory (LSTM) networks, in rainfall forecasting for infrastructure resilience planning. The results demonstrate that LSTM consistently beats Prophet in prediction ability and sensitivity to rainfall extremes. This has ramifications for hydrological forecasting and the operational application of AI in civil engineering and climate adaptation. A crucial feature of LSTM is its capacity to capture non-linear relationships and temporal variability(24), enabling it to detect more exceedance years than those above the 90th percentile threshold in both Rajshahi and Ishwardi. By contrast, Prophet’s smoother predictions, while computationally efficient and interpretable(23), tended to underestimate the frequency and severity of extremes, a weakness that may lead to underdesign if relied upon entirely for planning scenarios. These findings are consistent with regional evidence. In South Asia, deep learning algorithms such as LSTM have been demonstrated to capture monsoon variability and extreme rainfall events more successfully than standard statistical models (25, 26). Similar conclusions have been reported in Africa, where machine learning algorithms beat ARIMA and SARIMA in detecting drought and extreme precipitation patterns (27, 28). The congruence of our results with previous studies demonstrates that AI-based forecasting frameworks can deliver valid insights across diverse climatic scenarios. A novel contribution of this work resides in integrating predicted results with physical design consequences. The study turns data forecasts into meaningful design suggestions by establishing a rainfall sensitivity matrix and a conceptual adaptation diagram. This integration bridges the gap between climate intelligence and infrastructure governance. It enables civil engineers and urban planners to incorporate danger thresholds directly into their design logic rather than reacting to damage or failures. At the policy level, the framework helps current attempts to include AI in climate-smart infrastructure design. Beyond its scientific soundness, the technique offers a practical tool for national adaptation programs, where estimates of exceedance years could impact flood risk management plans, construction codes, and resilience requirements. In this sense, the framework complements Sustainable Development Goals (SDG 11: Sustainable Cities and Communities, SDG 13: Climate Action, and SDG 9: Industry, Innovation and Infrastructure), demonstrating how predictive analytics may be integrated into planning processes. In conclusion, this study reveals that LSTM-based rainfall forecasting increases hydrological prediction and provides a method for combining climate foresight with actionable infrastructure adaptation. While more technical refinement and deeper policy integration remain necessary, the approach provides a foundation and a realistic example of how proactive, data-driven planning might be operationalized in flood-prone regions.

6. Conclusion & Future Work

This study employed Prophet and LSTM models to long-term rainfall forecasting in Bangladesh (1980–2024 data, forecasting 2025–2039). Results reveal that LSTM captured interannual volatility and detected more exceedance years than Prophet, which provided smoother but less extreme-sensitive forecasts. These findings are critical for infrastructure resilience, as omitting high-risk years could lead to underestimated loads and poor designs. By tying forecasts to a rainfall sensitivity framework, this research provides a mechanism for integrating predictive analytics into practical adaptation solutions such as elevated foundations, enhanced drainage, and resilient materials. AI-driven foresight thus delivers concrete value for hydrological extremes prediction, civil engineering design, and policy planning. Prophet remains preferable for interpretability and rapid implementation, but LSTM offers higher performance for capturing nonlinear variability. Together, they emphasize the trade-off between interpretability and accuracy in climate applications. Future studies should incorporate multivariate predictors (ENSO, NDVI, soil moisture), extend the framework spatially across South Asia, and integrate forecasting outputs with GIS and BIM platforms to facilitate real-time urban resilience design.

8. Limitations

This study has various limitations that should be acknowledged. First, the use of annual aggregates provided useful insights for long-term infrastructure planning but neglected intra-annual rainfall changes such as seasonal flooding or short-term extremes, which are critical for operational decision-making. Second, projections were constrained to two sites (Rajshahi and Ishwardi), which exhibit distinct hydrological patterns but do not depict the whole spatial variety of rainfall across Bangladesh. Broader regional testing is needed to increase generalizability. Third, the univariate approach excludes exogenous climate variables such as ENSO indicators, soil moisture, or vegetation indices (NDVI), which may boost prediction robustness and allow for more thorough modeling of rainfall variability. While LSTM produced higher predicted accuracy and recorded rainfall extremes better than Prophet, it is computationally costly and less interpretable, which may limit its use by agencies with resource constraints. Prophet, by contrast, stays more translucent and lightweight but underfits extremes. This trade-off between accuracy and interpretability underscores the need for hybrid approaches that can bridge the strengths of both strategies. Future studies should therefore address these shortcomings by including higher-frequency data (monthly or daily), broadening spatial validation across varied regions, and constructing multivariate models that include exogenous factors. Additionally, integrating rainfall forecasts with digital planning platforms such as GIS or BIM could enable spatially explicit, real-time applications in infrastructure resilience planning, guaranteeing that AI-based foresight is both actionable and policy-relevant.

Acknowledgement

DeclarationsFunding Declaration: No funding was received for this study.Clinical Trial Number: Not applicable.Ethics Declaration: Not applicable.Consent to Publish Declaration: Not applicable.Consent to Participate Declaration: Not applicable.Competing Interests: The authors declare that they have no competing financial or non-financial interests.Data Availability Statement: Not applicable (this manuscript does not report data generation or analysis).Acknowledgements: I don't think it's applicable.

Declarations

Corresponding author: Hasan Ahamed Alif, email: hasanahamed.alif@dhaanishcollege.in

Ethics declaration

Not applicable.

Consent to Publish

declaration: Not applicable.

Consent to Participate

declaration: Not applicable

Author Contribution

H.A.A. and M.J.M. wrote the main manuscript text. U.H. and M.M.R. contributed to the methodology and data analysis. M.A.M. and M.U. contributed to the data collection and interpretation. S.B.Y. and M.J.H. assisted in reviewing the manuscript and preparing the supplementary materials. All authors reviewed and approved the final manuscript.

Data Availability

The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request

Reference

Buhl M, Markolf S. A review of emerging strategies for incorporating climate change considerations into infrastructure planning, design, and decision making. Sustainable Resilient Infrastructure. 2023;8(sup1):157–69.

Othmer FJ, Becker D, Schulte LM, Greiving S. A Methodological Approach to Municipal Pluvial Flood Risk Assessment Based on a Small City Case Study. Sustainability. 2020;12(24):10487.

Nik VM, Perera ATD, Chen D. Towards climate resilient urban energy systems: a review. Natl Sci Rev. 2021;8(3):nwaa134.

Li D, Ma J, Rao K, Wang X, Li R, Yang Y, et al. Prediction of Rainfall Time Series Using the Hybrid DWT-SVR-Prophet Model. Water. 2023;15(10):1935.

Alquraish M, Ali. Abuhasel K, Alqahtani S, Khadr A. SPI-Based Hybrid Hidden Markov–GA, ARIMA–GA, and ARIMA–GA–ANN Models for Meteorological Drought Forecasting. Sustainability. 2021;13(22):12576.

Azad AS, Sokkalingam R, Daud H, Adhikary SK, Khurshid H, Mazlan SNA, et al. Water Level Prediction through Hybrid SARIMA and ANN Models Based on Time Series Analysis: Red Hills Reservoir Case Study. Sustainability. 2022;14(3):1843.

Kim C, Kim DH. Effects of Rainfall Spatial Distribution on the Relationship between Rainfall Spatiotemporal Resolution and Runoff Prediction Accuracy. Water. 2020;12(3):846.

Slater LJ, Anderson B, Buechel M, Dadson S, Han S, Harrigan S, et al. Nonstationary weather and water extremes: a review of methods for their detection, attribution, and management. Hydrol Earth Syst Sci. 2021;25(7):3897–935.

Basso S, Merz R, Tarasova L, Miniussi A. Extreme flooding controlled by stream network organization and flow regime. Nat Geosci. 2023;16(4):339–43.

10.

Sun Q, Wan J, Liu S. Estimation of Sea Level Variability in the China Sea and Its Vicinity Using the SARIMA and LSTM Models. IEEE J Sel Top Appl Earth Observations Remote Sens. 2020;13:3317–26.

11.

Wei W, Yan Z, Tong X, Han Z, Ma M, Yu S, et al. Seasonal prediction of summer extreme precipitation over the Yangtze River based on random forest. Weather Clim Extremes. 2022;37:100477.

12.

Hasnain A, Sheng Y, Hashmi MZ, Bhatti UA, Hussain A, Hameed M, et al. Time Series Analysis and Forecasting of Air Pollutants Based on Prophet Forecasting Model in Jiangsu Province, China. Front Environ Sci. 2022;10:945628.

13.

Xie C, Wen H, Yang W, Cai J, Zhang P, Wu R, et al. Trend analysis and forecast of daily reported incidence of hand, foot, and mouth disease in Hubei, China by the Prophet model. Sci Rep. 2021;11(1):1445.

14.

Feng P, Wang B, Liu DL, Ji F, Niu X, Ruan H, et al. Machine learning-based integration of large-scale climate drivers can improve the forecast of seasonal rainfall probability in Australia. Environ Res Lett. 2020;15(8):084051.

15.

Raval M, Sivashanmugam P, Pham V, Gohel H, Kaushik A, Wan Y. Automated predictive analytics tool for rainfall forecasting. Sci Rep. 2021;11(1):17704.

16.

Schumacher RS, Hill AJ, Klein M, Nelson JA, Erickson MJ, Trojniak SM, et al. From Random Forests to Flood Forecasts: A Research to Operations Success Story. Bull Am Meteorol Soc. 2021;102(9):E1742–55.

17.

Rajagukguk RA, Ramadhan RAA, Lee HJ. A Review on Deep Learning Models for Forecasting Time Series Data of Solar Irradiance and Photovoltaic Power. Energies. 2020;13(24):6623.

18.

Zhang Y, Ragettli S, Molnar P, Fink O, Peleg N. Generalization of an Encoder-Decoder LSTM model for flood prediction in ungauged catchments. J Hydrol. 2022;614:128577.

19.

Hodson TO. Root-mean-square error (RMSE) or mean absolute error (MAE): when to use them or not. Geosci Model Dev. 2022;15(14):5481–7.

20.

Piepho H. An adjusted coefficient of determination (R²) for generalized linear mixed models in one go. Biometrical J. 2023;65(7):2200290.

21.

Rachidi S, El Mazoudi EH, El Alami J, Jadoud M, Er-Raki S. Assessment and Comparison of Satellite-Based Rainfall Products: Validation by Hydrological Modeling Using ANN in a Semi-Arid Zone. Water. 2023;15(11):1997.

22.

Stoffel MA, Nakagawa S, Schielzeth H. partR2: partitioning R² in generalized linear mixed models. PeerJ. 2021;9:e11414.

23.

Taylor SJ, Letham B. Forecasting at Scale. Am Stat. 2018;72(1):37–45.

24.

Hochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Comput. 1997;9(8):1735–80.

25.

Grover R, Sharma S, Bajpai A, Chintala S, Verma R, Kaur M. Temporal Variations in Monsoon Rainfall: A Comparative Study by Using Machine Learning Techniques. In: 2024 International Conference on Artificial Intelligence and Emerging Technology (Global AI Summit) [Internet]. Greater Noida, India: IEEE; 2024 [cited 2025 Sep 18]. pp. 924–9. Available from: https://ieeexplore.ieee.org/document/10947866/

26.

Xaba N, Mishra AK. Optimizing rainfall prediction in coastal and inland areas: a comparative analysis of forecasting models in eThekwini district, South Africa. Int J Bus Ecosyst Strategy 2687 – 2293. 2025;7(1):180–97.

27.

Samson TK, Aweda FO, Omoliki AJ, George GC, Oladapo OF, Ogunwale EI. Prediction of monthly rainfall in selected African stations using deep learning algorithms. IOP Conf Ser Earth Environ Sci. 2024;1428(1):012005.

28.

Ogallo LAJ. Stochastic modelling of regional annual rainfall anomalies in East Africa. J Appl Stat. 1986;13(1):49–56.

Yes