TeaNet: An Enhanced Attention Network for Climate-Resilient River Discharge Forecasting Under CMIP6 SSP585 Projections

Abstract

The accurate prediction of river discharge is essential in water resource management, particularly under variability due to climate change. Traditional hydrological models commonly struggle to capture the complex, nonlinear relationships between climate variables and river discharge, leading to uncertainties in long-term projections. To mitigate these challenges, this research integrates machine learning (ML) and deep learning (DL) techniques to predict discharge in the Subernarekha River Basin (India) under future climate scenarios. Global climate models (GCMs) from the Coupled Model Intercomparison Project 6 (CMIP6) are assessed for their ability to reproduce historical discharge trends. The selected CNRM-M6-1 model is bias-corrected and downscaled before being used to simulate future discharge patterns under SSP585 (a high-emission scenario). Various AI-driven models, such as a temporal convolutional network (TCN), a gated recurrent unit (GRU), a support vector regressor (SVR), and a novel DL network named the Temporal Enhanced Attention Network (TeaNet), are implemented by integrating the maximum and minimum daily temperatures and precipitation as key input parameters. The performance of the models is evaluated using the mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and coefficient of determination (R²). Among the evaluated models, TeaNet demonstrates the best performance, with the lowest error rates (RMSE: 2.34–3.04; MAE: 1.13–1.52 during training) and highest R² (0.87–0.95), outperforming the TCN (R²: 0.79–0.88), GRU (R²: 0.75–0.84), SVR (R²: 0.68–0.80), and RF (R²: 0.72–0.82) by 8–15% in accuracy across four gauge stations. The efficacy of the proposed model lies in its enhanced attention mechanism, which successfully identifies temporal relationships in hydrological information. In determining the most relevant predictors of river discharge, the feature importance is analyzed using the proposed TeaNet model. The findings of this research strengthen the role of DL architectures in improving long-term discharge prediction, providing valuable knowledge for climate adaptation and strategic planning in the Subernarekha region.

1. Introduction

It is significantly proven that global warming has been, more often than not, human-induced. It has altered the global rainfall pattern, as well as the air temperature [1]. This, in turn, has affected condensation and precipitation cycles all over the world [2,3]. Traditional watershed management is no longer working, and the alternative management strategies are not as multifaceted as the traditional ones. This situation is exacerbated by widespread deforestation, unsustainable agricultural practices, and inappropriate land usage. Many studies carried out all over the world show that, for the above reasons, the worldwide water distribution system is being increasingly affected, particularly in regions where management strategies are lacking and ineffective [4,5]. All of this causes a discrepancy between the supply and demand for freshwater, which then affects daily human consumption, industry, and agriculture by causing resource depletion, water shortages, and increased stress on existing water sources. Many climate forecasting models are being developed to effectively aid planners in coping with this gap with efficient water resource management strategies. These efficient models, with their accurate/near-accurate river discharge patterns, will be elucidative for effective, sustainable water management, especially for regions vulnerable to climatic variability [6]. To simulate past and future climate conditions based on multi-dimensional emission scenarios, the state-of-the-art technology is global climate models (GCMs), which are gaining popularity. The World Climate Research Programme (WCRP), too, has provided strong frameworks for such assessments across different time scales. The Coupled Model Intercomparison Project (CMIP) is one such model, initiated in 1995 under the WCRP [7].

The CMIP series of models was developed to improve climate predictions, assess long-term environmental changes, and support policy decisions related to climate adaptation and mitigation [8]. These models integrate various climate factors, including atmospheric, oceanic, and land surface processes, to provide more accurate projections of future climate scenarios. In these series, CMIP5 has played a crucial role in the development of climate change scenarios [9]; however, due to the variability with respect to forecasting monsoon rainfall in South Asia, there exist significant uncertainties [10]. CMIP6, the latest iteration, integrates improved parameterizations and shared socioeconomic pathways (SSPs) for more reliable projections of these types [11,12,13]. This improved model reduces the biases in simulation processes and improves the synoptic process representation [14,15].

GCM models are being improved constantly to minimize uncertainties, but the inherent biases lead to the over- or underestimation of climate change impacts [16]. These biases necessitate statistical or dynamic downscaling techniques to refine GCM outputs and improve their spatial and temporal resolutions for hydrological applications [17]. However, selecting appropriate GCMs for hydrological modeling remains challenging, particularly in developing regions with limited computational and human resources [18,19]. Multi-model ensembles (MMEs) are commonly used to minimize uncertainty and produce more locally relevant projections [20]. Nevertheless, despite these advances, GCMs alone often fail to capture the fine-scale hydrological dynamics required for accurate water resource forecasting. To address these challenges, there is growing interest in integrating data-driven techniques, particularly machine learning and deep learning models, with GCM outputs and in situ data for improved hydrological predictions.

Hydrological models are essential in simulating rainfall–runoff processes based on the physical characteristics of a watershed [21]. These models provide a crucial understanding of water partitioning, budgeting, and inundation mapping, which are fundamental for water resource management [22]. However, traditional hydrological models are highly data-intensive and computationally demanding. Moreover, no single model has proven effective in capturing the complex interactions between climatic and hydrological processes, especially under climate change and anthropogenic activities [23]. Recent advances in ML and DL provide promising alternatives for hydrological modeling, enabling more accurate predictions of rainfall, streamflow, groundwater levels, and flood risks. These AI-driven approaches can handle complex, nonlinear relationships in hydrological systems, process large datasets efficiently, and improve decision-making in water resource management and climate change adaptation. ML techniques such as regression trees (RT), long short-term memory (LSTM), adaptive boosting (ADB), and gradient boosting (GB) have demonstrated potential in forecasting rainfall–runoff and river discharge patterns while requiring relatively smaller data inputs [24,25,26]. Despite their usefulness, these methods are frequently criticized for failing to adequately represent the nonlinearity in river discharge behavior, highlighting the need for advanced ML architectures in hydrological studies [27].

While ML applications in hydrological modeling have increased, direct comparative analyses of multiple ML methods for bias correction in climate-driven river discharge forecasting remain scarce [28,29]. Most studies have focused on individual ML techniques, without evaluating their relative performance under CMIP6-based climate scenarios [30]. A recent study [31] demonstrated that integrating multiple hydrological models (VIC, H08, CWatM, Noah-MP, and CLM) with ML-based post-processing methods such as random forest (RF), extreme gradient boosting (XGB), and LSTM significantly improved streamflow simulations. However, a comprehensive review of ML methodologies in hydrological processes [32] has pointed out that, despite these advancements, ML models often struggle with physical interpretability and generalization, particularly in extreme hydrological events. However, another work [33] highlighted the challenges of parameter uncertainty in hydrological modeling and demonstrated the effectiveness of ML-based emulators in reducing the computational costs while maintaining accuracy. Their study applied ML models like artificial neural networks (ANN), K-nearest neighbors (KNN), and RF for parameter identification, reinforcing the importance of region-specific calibration techniques [34,35]. These studies collectively emphasize the need for adaptive bias correction strategies in ML-based river discharge forecasting to improve model reliability and robustness [36,37]. Furthermore, research on the Subarnarekha River Basin, an essential yet underrepresented region in India, remains limited despite its high vulnerability to climate variability. Addressing these gaps is essential in improving climate-adaptive water resource management strategies.

In parallel, deep learning (DL) methods have increasingly been integrated into environmental and hydrological modeling due to their capacity to capture nonlinearities and complex spatiotemporal dependencies. From 2021 to 2024, several studies advanced this domain by proposing hybrid frameworks that combined DL architectures with optimization and ensemble learning strategies. For example, Ref. [38] developed a novel logic development algorithm (LDA)-optimized boosted artificial neural network (BAANN) to accurately estimate permeability in plastic waste aggregate concrete, significantly outperforming conventional ML models. Their findings demonstrated the importance of hyperparameter tuning and hybrid learning in achieving high predictive performance. Similarly, Liu [39] employed transformer-based models for streamflow forecasting, demonstrating improved accuracy over recurrent networks by effectively modeling long-range dependencies. The integration of physical constraints through physics-guided deep learning (PGDL) has further enhanced models’ generalizability and scientific validity in hydrological contexts [40]. Moreover, explainable AI (XAI) techniques such as SHAP have been increasingly adopted to improve model transparency and foster stakeholder trust [41]. Despite these advancements, a lack of comparative evaluations of DL models in CMIP6-driven hydrological prediction remains, particularly for underrepresented basins such as that of the Subarnarekha. The present study addresses this gap by employing and assessing advanced DL methods under SSP585 projections.

This study bridges the identified research gap by comprehensively comparing multiple ML techniques for bias correction in CMIP6-driven hydrological forecasting in the Subarnarekha River Basin. Specifically, this study evaluates the performance of a TCN, GRU, RF, and SVR alongside the proposed TeaNet model. Recent research [42] analyzed RF, LSTM, SVM, and multivariate adaptive regression splines (MARS) for streamflow prediction in the rivers of Southern India, demonstrating that RF outperformed other models in daily streamflow prediction, while LSTM showed superior performance for monthly streamflow forecasting. These findings validate the inclusion of RF and LSTM in our study while emphasizing the need for novel architectures like TeaNet to address model limitations in extreme flow prediction. Similarly, a study [43] applied AI-driven forecasting techniques, including RF, ANNs, and hybrid models, to predict river discharge in the Himalayan Sindh River. Their findings highlighted the superior performance of hybrid models like RF-SARIMA in improving discharge predictions by effectively capturing seasonal variations and complex hydrological dynamics. These results reinforce the need for comparative assessments of advanced ML architectures to improve the forecasting accuracy under future climate scenarios further. To extend the applicability of these ML approaches, this study integrates shared socioeconomic pathway SSP585 to analyze future hydrological scenarios, providing a further understanding of climate-driven water resource variability. By comparing various ML approaches, this study presents an integrated framework for the selection of the optimal bias correction method, thereby enhancing hydrological projections, water resource planning and sustainability, and climate resilience in underrepresented river basins.

2. Study Area

The Subarnarekha River Basin is a transboundary river system spanning the states of Jharkhand, Odisha, and West Bengal in Eastern India, covering a drainage area of approximately 18,951 km². The river originates near Nagri village (23°18′ N, 85°11′ E) in the Ranchi district of Jharkhand at an elevation of about 600 m and flows eastward through the Chhotanagpur Plateau, crossing the Dalma Hill Range and draining into the Bay of Bengal at Talshari, Odisha, with a total length of 470 km (Figure 1). The basin is bounded by the Chota Nagpur Plateau to the northwest, the Brahmani and Burhabalang basins to the south, and the Bay of Bengal to the southeast. The principal tributaries of the Subarnarekha include the Kharkai, Kanchi, and Karkari rivers, with the Kharkai joining at Sonari near Jamshedpur. The basin features diverse terrains, including the elevated Chhotanagpur Plateau and alluvial plains, as well as Tertiary and pre-Cambrian formations with phyllites, dolomites, gneiss, and granites. The region falls under elevations ranging from 0 to 1172 m and experiences an average annual temperature between 18.0 and 32.4 °C. Daily data on the precipitation and temperature were downloaded from the CMIP6 repository [44,45,46,47].

The humid subtropical climate of the basin is defined by monsoon rains occurring from June to October, which contribute around 90% of the annual rainfall, with the annual rainfall ranging from 1100 to 1400 mm, mostly during the monsoon season. The river basin includes two primary topographical regions: the Northern Plateau, comprising the Ranchi and Singhbhum districts in Jharkhand, and the Coastal Plains, extending across Odisha’s Mayurbhanj district and West Bengal’s East and West Medinipur districts. The study area, situated within the Jharkhand region, covers the upper half of the basin up to the Ghatsila gauge station (21°33′ to 23°32′ N latitudes and 85°09′ to 87°27′ E longitudes), with a catchment area of approximately 14,140 km². The basin’s geology includes a variety of soil types, such as alluvial soils, laterites, and boulder conglomerates, alongside basic and ultrabasic rocks in the southern region and Chhotanagpur granite gneiss elsewhere. The subtropical climate is marked by evenly distributed rainfall, primarily during the southwest monsoon, which brings about 90% of the total annual rainfall [44,45,46].

The Subarnarekha Basin supports a population of over 7.8 million people and is vital for agriculture, industry, and the domestic water supply in the region. It is also home to significant mineral resources and major industries, particularly around Jamshedpur and Ranchi. Several large-scale water resource projects, including the Chandil Dam, Icha Dam, and Galudih Barrage, have been constructed for irrigation, hydropower, and flood control purposes. Despite its economic importance, the basin faces environmental challenges, such as pollution from mining and industry, deforestation, and sedimentation, which threaten its ecological balance and water quality [48]. The Subarnarekha River has historically been known as the “Streak of Gold” due to gold mining near its origin [44,45,46].

The accurate prediction of river flows in the Subarnarekha River Basin is crucial for several reasons. This region is highly dependent on the river for irrigation, drinking water, industrial use, and ecological sustenance. Reliable flow forecasts enable effective water allocation, optimize reservoir operations, and support agricultural planning, especially during the critical monsoon season. Furthermore, timely and accurate discharge predictions are essential for flood risk mitigation, as the basin is prone to flash floods and extreme hydrological events, which can have devastating impacts on lives, livelihoods, and infrastructure. Advanced prediction models also play a vital role in monitoring river health, managing pollution, and ensuring the sustainability of aquatic ecosystems. In the context of climate change and increasing variability in rainfall and temperatures, robust river flow forecasting is indispensable for informed decision-making and long-term water resource planning in the Subarnarekha Basin.

3. Materials and Methodology

3.1. Data and Methods

3.1.1. Hydroclimatic Data

Meteorological data from four active gauge stations of the Subarnarekha River Basin, namely Muri, Adityapur, Jamshedpur, and Ghatsila, were obtained from the Central Water Commission (CWC). This dataset includes the maximum and minimum temperature, precipitation, and discharge records from 1980 to 2022, as presented in

Table 1. Detailed specifications of the discharge dataset used for analysis.

Gauge Station	Period of Records	Data Type	Latitude (N)	Longitude (E)	Elevation (m)
Muri	1989–2020	Daily	23.3628°	85.8747°	231
Adityapur	1980–2022	Daily	22.7861°	86.1744°	123
Jamshedpur	1980–2020	Daily	22.8156°	86.2161°	111
Ghatsila	1980–2022	Daily	22.5856°	86.4617°	72

. The daily data on precipitation and temperature (maximum and minimum) from 5 GCMs, namely ACCESS- CM 2, CNRM-CM6-1, CNRM-ESM-2-1, EC-EARTH3, and MRI-ESM2-0, were downloaded from the CMIP6 repository (https://www.nccs.nasa.gov/services/data-collections/land-based-products/nex-gddp-cmip6, accessed on 20 March 2025), whose specifics are provided in

Table 2. Specifications of the selected CMIP6 GCM climate models.

Modeling Agency	Model Name	Description	Country
Commonwealth Scientific and Industrial Research Organization, Australian Research Council Centre of Excellence for Climate System Science	ACCESS-CM2	Commonwealth Scientific and Industrial Research Organisation (CSIRO)	Australia
Centre National de Recherches Meteorologiques/Centre Européen de Recherche et Formation Avancée en Calcul Scientifique	CNRM-ESM2-1	Centre National de Recherches Météorologiques Coupled Global Climate Model, version 5	France
Centre National de Recherches Meteorologiques/Centre Européen de Recherche et Formation Avancée en Calcul Scientifique	CNRM-CM6-1	Centre National de Recherches Météorologiques Coupled Global Climate Model, version 5	France
EC-Earth Consortium	EC-Earth3	EC-Earth Earth System Model Version 3 with Dynamic Vegetation Component	Europe
Meteorological Research Institute (MRI)	MRI-ESM2-0	Meteorological Research Institute Earth System Model Version 2.0	Japan

.

3.1.2. Data Pre-Processing

• Bias correction of GCM models

While GCMs employ complex numerical representations of atmospheric, oceanic, and land surface processes, they inherently contain systematic biases due to uncertainties in parameterization, spatial resolution limitations, and boundary condition assumptions [17]. These biases can significantly affect hydrological simulations, necessitating statistical corrections to refine the accuracy of climate projections used in impact assessments and decision-making. In this study, two widely used bias correction methods, linear scaling (LS) and the Delta Change (DC) approach, were employed to adjust the precipitation and temperature data obtained from five CMIP6 GCMs (ACCESS-CM2, CNRM-CM6-1, CNRM-ESM-2-1, EC-EARTH3, and MRI-ESM2-0).

The linear scaling (LS) method is a simple yet effective bias correction technique that adjusts raw GCM outputs by applying a scaling factor derived from the ratio of observed and simulated climatological means [49]. This approach ensures that the long-term mean of GCM projections aligns with historical observations, while preserving the temporal variability of the original data. For temperature correction, the LS method applies an additive correction using Equation (1):

$$T_{corr,i}=T_{GCM,i}+\left(T_{obs,mean}-T_{GCM,mean}\right)$$

(1)

where $T_{corr,i}$ = bias-corrected temperature for day i; $T_{GCM,i}$ = raw GCM temperature for day i; $T_{obs,mean}$ = observed historical mean temperature; and $T_{GCM,mean}$ = GCM-simulated mean temperature. Other technical details include the additive shift, computed as ΔT_m = μ(T_obs,m) − μ(T_raw,m) for each month m; diurnal range preservation, i.e., the daily maximum/minimum temperatures corrected independently to maintain intra-day variability; and validation, with shift factors validated against 2016–2020 holdout data (NSE > 0.89 for all stations). For precipitation correction, the LS method applies a multiplicative factor to account for proportional scaling using Equation (2):

$$P_{corr,i}=P_{GCM,i}\times \left({\displaystyle \frac{P_{obs,mean}}{P_{GCM,mean}}}\right)$$

(2)

where $P_{corr,i}$ = bias-corrected precipitation for day i; $P_{GCM,i}$ = raw GCM precipitation for day i; $P_{obs,mean}$ = observed historical mean precipitation; and $P_{GCM,mean}$ = GCM-simulated mean precipitation. Other technical details include the calibration period, which is 1980–2020 (41 years); scaling factor computation, with monthly multipliers calculated as ${\displaystyle \frac{{\mu (P}_{obs,m})}{{\mu (P}_{GCM,m})}}$, where m denotes the calendar month; implementation, which is applied separately for each GCM and gauge station using a 30-day rolling window to smooth seasonal transitions; threshold handling, with the wet-day threshold set to 1 mm/day; and days with P_raw < 1 mm assigned P_cor = 0. The LS method applies first-order statistical corrections by aligning the monthly means of the GCM outputs with the observed climatology. While computationally efficient, LS assumes stationarity in the bias magnitude and only corrects mean discrepancies, leaving higher-order statistical moments (variance, skewness) and extreme value distributions unadjusted. This limitation becomes pronounced when analyzing climate extremes, as the LS-preserved variability patterns may not reflect the observed hydrological realities.

The Delta Change (DC) approach is another statistical downscaling method that applies a constant correction factor to GCM outputs, ensuring consistency with historical observations while preserving future climate signal projections [50]. For the temperature, the DC correction is given by Equation (3):

$$T_{corr,i}=T_{obs,i}+\left(T_{GCM,future}-T_{GCM,historical}\right)$$

(3)

where $T_{corr,i}$ = bias-corrected temperature for day i; $T_{obs,i}$ = observed historical temperature for day i; $T_{GCM,future}$ = GCM-simulated future temperature; and $T_{GCM,historical}$ = GCM-simulated historical temperature. Other technical details include delta calculation, where ΔT_i = T_fut,i − T_his,i derived from 30-year climatological means (1980–2010 vs. 2071–2100), and spatial consistency, with grid-cell-specific deltas interpolated to station locations using inverse distance weighting. For precipitation, a multiplicative correction factor is applied, as shown in Equation (4):

$$P_{corr,i}=P_{obs,i}\times \left({\displaystyle \frac{P_{GCM,future}}{P_{GCM,historical}}}\right)$$

(4)

where $P_{corr,i}$ = bias-corrected precipitation for day i; $P_{obs,i}$ = observed historical precipitation for day i; $P_{GCM,future}$ = GCM-simulated future precipitation; and $P_{GCM,historical}$ = GCM-simulated historical precipitation. Other technical details include the future period, which is 2021–2100 under SSP585; the delta factor, i.e., δP_i = ${\displaystyle \frac{P_{future,i}}{P_{historical,i}}}$, computed at a daily resolution and then smoothed using a 15-day moving average; and extreme value handling, with the 95th percentile capped at 300 mm/day to prevent unrealistic precipitation spikes. The DC method is especially valuable if long-term climate trends need to be maintained so that bias correction does not alter the projected climate signals. However, it relies on the assumption that the biases are constant over time, which is not necessarily true in all instances. While effective in preserving future climate signals, DC’s temporal invariance assumption may (1) underestimate precipitation intensification trends in warming scenarios, (2) over-smooth temperature extremes due to additive correction mechanics, or (3) fail to account for changing bias–cloud feedback relationships.

Despite the recognized limitations, the LS and DC methods were chosen based on their complementary strengths in addressing distinct aspects of bias correction. LS excels in aligning GCM outputs with observed climatological means through computationally efficient monthly scaling factors, making it ideal for large multi-model ensembles [44]. Its simplicity ensures the minimal distortion of intra-monthly variability while effectively reducing mean-state biases. Conversely, DC preserves climate change signals by applying station-specific delta factors, crucial in maintaining SSP585’s projected intensification of monsoon extremes [40]. This dual approach balances historical fidelity (via LS) with future trend preservation (via DC), addressing both distributional biases and climate signal integrity. In this study, the process of bias correction was conducted with Python version 3.10.12, utilizing the xarray and pandas libraries for data manipulation and the SciPy Stats module to make statistical adjustments. Subsequently, the corrected climate variables were validated using independent observational datasets by employing statistical performance measures.

3.2. Methodology

A visual depiction of the research methodology used in this study is presented in Figure 2. The precipitation and temperature (maximum and minimum) were utilized as input variables, and discharge was the output variable for river discharge simulation. Different ML and DL models, such as a temporal convolutional network (TCN), a gated recurrent unit (GRU), random forest (RF), support vector regression (SVR), and the proposed model named the Temporal Enhanced Attention Network (TeaNet), were employed to simulate river discharge. The historical data were divided into a 70% training dataset (1980–2000) and a 30% testing dataset (2000–2022). The proposed TeaNet architecture was assessed using four statistical indicators (MSE, RMSE, MAE, and R²) for the training and testing periods. In parallel, a bias correction was applied to the data from the GCM models.

3.2.1. Regression Models

• Temporal Convolutional Network (TCN)

The TCN [51] is carefully optimized for sequence modeling problems, making it well suited for discharge prediction. TCNs use 1D dilated causal convolutions to capture long-range dependencies in time series data without facing the vanishing gradient issue. The TCN architecture is formed of several convolutional layers, with each convolutional layer using a dilation factor d, and the receptive field increases exponentially with the depth L. The output at each layer is given by Equation (5):

$$y\left(t\right)=\sum _{i=0}^{k-1}{W}_{i}x\left(t-d\cdot i\right)$$

(5)

where W_i—convolutional filter weights, y(t)—output at time step t, k—filter size, x(t − d⋅i)—input at a delayed time step (t − d⋅i), and d—dilation rate. By layering several dilated convolutional layers, TCNs efficiently learn multi-scale temporal features, which are essential for hydrological modeling. This enables the network to capture the long-term tendencies of river discharge, making it a resilient alternative to conventional autoregressive models.

• Random Forest (RF)

RF creates multiple decision trees and combines their predictions for higher accuracy and robustness [52]. In hydrological modeling, RF performs best in detecting complex interactions among meteorological and hydrological variables like precipitation, temperature, and previous discharge levels. The RF algorithm achieves this through (i) bootstrapping B datasets from the given data, (ii) creating B decision trees by training one on each unique subset, and (iii) predicting using a voting average (regression) or by voting (classification) for every tree. For regression tasks like discharge prediction, the final RF output is computed as shown in Equation (6):

$$\widehat{y}={\displaystyle \frac{1}{B}}\sum _{b=1}^B{T}_b\left(x\right)$$

(6)

where B—number of decision trees, $\widehat{y}$—predicted discharge, and T_b(x)—prediction of the b-th tree for input x. The strength of RF is that it can process high-dimensional data, reduce overfitting, and give feature importance rankings, thus making it necessary in evaluating the impact of climatic variables on river discharge.

• Gated Recurrent Unit (GRU)

The GRU preserves temporal relationships in sequential data and prevents vanishing gradients [53]. GRUs employ two essential gates: (i) the update gate z_t, which decides how much of the previous information needs to be carried over, and (ii) the reset gate r_t, which regulates the influence of prior hidden states on the present state. The mathematical formulation of a GRU cell is given below (Equations (7)–(10)):

$$z_t=\sigma (W_z{x}_t+U_z{h}_{t-1}+b_z)$$

(7)

$$r_t=\sigma (W_r{x}_t+U_r{h}_{t-1}+b_r)$$

(8)

$$\widetilde{h_t}=tanh(W_h{x}_t+U_h(r_t\cdot h_{t-1})+b_h)$$

(9)

$$h_t=(1-z_t)\cdot h_{t-1}+z_t\cdot \widetilde{h_t}$$

(10)

where h_t—hidden state; x_t—input at time t; W, U, b are weight matrices and bias terms; $\widetilde{h_t}$—sigmoid activation function; and tanh—hyperbolic tangent activation function. By utilizing these gating mechanisms, GRUs efficiently learn long-term dependencies in hydrological time series data, capturing seasonal and periodic variations in river discharge more effectively than traditional statistical models.

• Support Vector Regression (SVR)

SVR is useful in hydrological modeling to capture nonlinear relationships between climatic variables [54] (e.g., rainfall, temperature) and river discharge while being robust to noise and high-dimensional data [55]. SVR aims to find a function f(x) that predicts discharge values while minimizing the error within a specified tolerance ϵ. The regression function is given by Equation (11):

$$f\left(x\right)=w^T\varphi \left(x\right)+b$$

(11)

where ϕ(x)—a kernel function that transforms the input data into a higher-dimensional space, allowing SVR to capture nonlinear patterns effectively. The optimization objective is to minimize the complexity of the model while ensuring that most predictions fall within the ϵ-insensitive zone, formulated as shown in Equation (12):

$${\displaystyle \frac{1}{2}}{\parallel w\parallel }^2+C\sum _{i=1}^N\left({\xi }_i+{\xi }_i^{\mathrm{*}}\right)$$

(12)

where C is a regularization parameter controlling the trade-off between model complexity and prediction errors, and ${\xi }_i$, ${\xi }_i^{*}$ are slack variables that handle deviations beyond ϵ. SVR’s ability to handle outliers and nonlinearity makes it well suited for discharge prediction, where hydrological datasets often exhibit high variability and uncertainty. However, its performance strongly depends on the proper tuning of hyperparameters such as C, ϵ and the kernel function, requiring careful calibration for optimal results.

• Temporal Enhanced Attention Network (TeaNet)

The Temporal Enhanced Attention Network (TeaNet) is designed to effectively process sequential data by integrating temporal feature extraction, increased attention mechanisms [56], and feature fusion. TeaNet introduces several architectural innovations that enhance its ability to process time series data. It achieves this through the synergistic combination of three key mechanisms: multi-dilated convolutions that capture rich temporal contexts, an enhanced attention module that models inter-feature and temporal relationships, and a residual fusion layer that retains both raw and processed information. These innovations work together to provide a more comprehensive and context-aware understanding of sequential data, ultimately improving the accuracy and generalizability of the model.

The model begins with an input layer that accepts data shaped as (T, F), where T—number of time steps and F—number of features in the sequence. Temporal feature extraction is achieved using five convolutional layers, each with causal padding and increasing dilation rates (1, 2, 4, 8, 16). The causal padding ensures that future data points are not used to predict past values, maintaining temporal causality. The dilation rates expand the receptive field, allowing the model to capture dependencies across varying time scales efficiently. Formally, the output of a dilated convolution at time t is expressed as follows (Equation (13)):

$$y_t={\sum }_{i=0}^{k-1}{w}_i\cdot x_{t-i\cdot d}$$

(13)

where k—kernel size, $w_i$—convolution weights, d—dilation rate, and x—input sequence. Dropout layers are incorporated after each convolutional layer to mitigate overfitting. Following this, an improved attention module processes the temporal features using a multi-head attention mechanism. This module captures interactions between features at different time steps. The attention mechanism calculates the relevance of each feature (Equation (14)):

$$\mathrm{Attention}\left(Q,K,V\right)=\mathrm{softmax}\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}\right)V$$

(14)

where Q, K, and V are the query, key, and value matrices, respectively, and $d_k$ is the dimensionality of the key. Residual connections and layer normalization ensure stable training and maintain the integrity of the original temporal features. The model incorporates a temporal attention mechanism to highlight the significance of specific time intervals. A dense layer with Softmax activation generates normalized attention scores for each time step (Equations (15) and (16)):

$${\alpha }_t={\displaystyle \frac{\exp \left(z_t\right)}{{\sum }_{t^{\prime }}\exp \left(z_{t^{\prime }}\right)}}$$

(15)

$$z_t=W\cdot h_t+b$$

(16)

where $z_t$—score for time step t, $h_t$—input to the dense layer that computes the attention score $z_t$, and W and b—trainable weights. These scores are used to reweight the temporal features (Equation (17)):

$$h_t^{\mathrm{attended}}={\alpha }_t\cdot h_t$$

(17)

The reweighted temporal features are then combined with the original features via a feature fusion step that uses a residual connection (Equation (18)):

$$f_t=h_t^{\mathrm{attended}}+h_t$$

(18)

This fusion ensures the preservation of both raw and attended temporal information. The fused features are passed through global average pooling, which summarizes the sequence by averaging across time steps (Equation (19)):

$$g={\displaystyle \frac{1}{T}}{\sum }_{t=1}^T{f}_t$$

(19)

This reduces the sequence to a fixed-size vector, capturing the most salient information. Then, the pooled features are processed by fully connected layers to learn nonlinear relationships. A dense layer with ReLU activation refines the features, followed by a dropout layer to prevent overfitting. The output layer predicts the target variable using a linear activation function for the effective forecasting of river discharge, represented as follows (Equation (20)):

$$\widehat{y}=W_{\mathrm{out}}\cdot g+b_{\mathrm{out}}$$

(20)

where $\widehat{y}$—predicted target variable, g—pooled feature vector obtained from global average pooling, $W_{\mathrm{out}}$ and $b_{\mathrm{out}}$—trainable weights and bias of the output layer. Figure 3 illustrates the layers of the TeaNet architecture graphically, while

Table 3. Layer-wise description and parameters of the TeaNet architecture.

Layer	Parameter	Description
Conv1D	4096	Filters = 128, kernel_size = 3, dilation_rate = 1, padding = ‘causal’, activation = ‘relu’
Dropout	0	Dropout layer with rate 0.2
Conv1D	49,152	Filters = 128, kernel_size = 3, dilation_rate = 2, padding = ‘causal’, activation = ‘relu’
Dropout	0	Dropout layer with rate 0.2
Conv1D	24,620	Filters = 64, kernel_size = 3, dilation_rate = 4, padding = ‘causal’, activation = ‘relu’
Dropout	0	Dropout layer with rate 0.2
Conv1D	12,352	Filters = 64, kernel_size = 3, dilation_rate = 8, padding = ‘causal’, activation = ‘relu’
Dropout	0	Dropout layer with rate 0.2
Conv1D	6176	Filters = 32, kernel_size = 3, dilation_rate = 16, padding = ‘causal’, activation = ‘relu’
Dropout	0	Dropout layer with rate 0.2
Dense (Processed Input)	352	Fully connected layer applied to match input shape
Multi-Head Attention	20,992	4 attention heads, key dimension = 32
Add (Residual)	0	Add input and attention output
Layer Normalization	64	Layer normalization applied to attention output
Dense (Temporal Weights)	33	Temporal attention weights (sigmoid activation)
Temporal Weighting	0	Element-wise multiplication with attention weights
Add (Feature Fusion)	0	Merge original and attended features
Global Average Pooling	0	Compute mean across the time steps
Dense	2112	Fully connected layer with 64 units, ReLU
Dense (Output)	65	Fully connected layer for regression output

provides a detailed description and the parameters of these layers.

3.2.2. Feature Importance Using TeaNet Model

Feature importance is a metric that evaluates the contribution of each input variable to the output predictions. In river discharge forecasting, examining the feature importance provides essential perspectives regarding which factors most directly influence river discharge variation. The influence of the rainfall and temperature variables on river discharge is evaluated through the analysis of feature importance employing the TeaNet DL model, where the feature importance can be extracted and quantified based on the learned temporal attention weights. Mathematically, the process is implemented as follows.

Temporal Attention Mechanism: The attention mechanism assigns importance to each time step of the input sequence, calculated as follows (Equation (21)):

$${\alpha }_t=\mathrm{sigmoid}\left(W\cdot h_t+b\right)$$

(21)

where ${\alpha }_t$—attention weight for time step t, $h_t$—hidden representation for time step t, W—learned weight matrix, and b—bias term.

Feature Weighting: The attention weights are applied to the input features as follows (Equation (22)):

$$x_{t,\mathrm{weighted}}=x_t\cdot {\alpha }_t$$

(22)

where $x_{t,\mathrm{weighted}}$—weighted feature value for time step t, $x_t$—original input feature for time step t, and ${\alpha }_t$—attention weight for time step t.

Global Feature Importance: After applying temporal attention, the feature importance for each variable can be aggregated across all time steps as follows (Equation (23)):

$$I_f={\displaystyle \frac{1}{T}}{\sum }_{t=1}^T{x}_{t,f,\mathrm{weighted}}$$

(23)

where T—total number of time steps, $I_f$—importance score for feature f, and $x_{t,f,weighted}$—weighted value of feature f at time step t. Features with higher attention weights (${\alpha }_t$) exhibit stronger relevance in predicting river discharge.

3.2.3. Performance Assessment

Various algorithms, such as RF, SVR, TCN, and GRU, were employed for discharge data collected from four gauge stations to evaluate the efficacy of machine learning models in predicting river discharge. A rigorous hyperparameter tuning process was employed to ensure the reliability and generalizability of all ML/DL models before performance evaluation. For SVR, which is particularly sensitive to parameter changes, a grid search with 5-fold cross-validation was employed to optimize the kernel type, regularization constant C, epsilon (ε), and gamma (γ). The random forest, GRU, and TCN models were also tuned in terms of the depth, structure, and learning rates using a combination of automated and manual methods. The TeaNet model was iteratively refined through the empirical testing of dilation rates, attention head sizes, and convolutional layers. The optimal hyperparameter configurations are summarized in

Table 4. Hyperparameter tuning details for ML and DL models.

Model	Tuned Hyperparameters	Optimal Values	Tuning Method
SVR	Kernel, C, ε, γ	RBF, C = 10, ε = 0.1, γ = 0.01	Grid Search + 5-fold CV
Random Forest	n_estimators, max_depth, min_samples_split	200, 15, 2	Grid Search
GRU	Layers, Units, Dropout, Learning Rate	2 layers, 64 units, 0.2, 0.001	Keras Tuner + Manual Tuning
TCN	Filters, Kernel size, Dilation Rate	64 filters, kernel = 3, [1, 2, 4, 8, 16]	Manual Tuning
TeaNet	Attention Heads, Convolutional Layers, Dilation Rates	4 heads, 5 conv layers, [1, 2, 4, 8, 16]	Iterative Experimentation

, enhancing the reproducibility and robustness of the comparative performance results discussed below. A set of performance evaluation metrics was utilized, including the RMSE, MAE, R², and MSE, to assess the performance of the models. These metrics provide a multidimensional assessment of the predictive accuracy, capturing both the error magnitude and model generalizability.

Table 5. Performance evaluation metrics utilized in this study.

Evaluation Metric	Mathematical Representation
Root Mean Squared Error (RMSE)	$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{\left(y_i-\widehat{y_i}\right)}^2}$
Coefficient of Determination (R2)	${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\widehat{y_i}\right)}^2}}$
Mean Absolute Error (MAE)	$\mathrm{MAE}={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}\left|y_i-\widehat{y_i}\right|$
Mean Squared Error (MSE)	$\mathrm{MSE}={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n}{\left(y_i-\widehat{y_i}\right)}^2$

represents the corresponding equations for the analyzed performance indicators, where $y$—predicted value and $\widehat{y}$—observed value, while $y_i$ and ${\widehat{y}}_i$ are the observed and predicted ith values. MSE, MAE, and RMSE values close to 0 and R² values close to 1 mean that the prediction performance is good [57].

4. Results and Discussion

4.1. Multi-Model Ensemble of General Circulation Models

The effectiveness of GCMs in simulating hydroclimatic variables plays a crucial role in assessing future climate scenarios and their impacts on hydrological systems. However, individual GCMs often exhibit inherent biases and structural uncertainties, necessitating an ensemble approach to improve the predictive reliability [30]. In this study, the SSP585 scenario was analyzed to identify the most accurate GCMs by comparing their outputs to observed historical discharge data. The models CNRM-CM6-1, CNRM-ESM2-1, EC-Earth3, and MRI-ESM2-0 demonstrated the strongest correlations with observed records, making them the most suitable candidates for ensemble-based forecasting.

To increase the robustness of the climate projections, an MME was constructed by aggregating the outputs of these top-performing GCMs using an arithmetic mean approach. This method effectively compensates for model-specific discrepancies, reducing systematic errors and improving the overall predictive accuracy. The ensemble-based approach is particularly advantageous in hydrological modeling, as it integrates diverse representations of atmospheric and oceanic processes, leading to a more comprehensive simulation of climate-driven variations in river discharge. The refined ensemble data generated from the MME were subsequently utilized as inputs for ML and DL models, further refining the reliability of the long-term discharge predictions.

4.2. Analysis of Regression Models’ Performance

The performance of these models varied across locations, with most demonstrating suboptimal accuracy in capturing the nonlinear and dynamic relationships inherent in the hydrological data. In contrast, the proposed TeaNet model consistently outperformed the benchmark models, effectively capturing intricate temporal patterns and improving the discharge prediction accuracy. Utilizing the predictive capabilities of TeaNet, the discharge projections were extended to the year 2100. However, long-term extrapolation inherently introduces uncertainty due to potential shifts in climate dynamics and land use patterns [58,59]. Thus, while these projections provide a critical understanding of future discharge trends, they should be interpreted with caution, particularly for long-term water resource planning and decision-making [60]. The results indicate significant variations in predictive capabilities across different models and discharge stations. Among the models tested, TeaNet consistently outperformed other approaches across all four discharge stations—Muri, Jamshedpur, Adityapur, and Ghatsila—in both the training and testing phases, as represented in Supplementary Table S1. The superior performance of TeaNet suggests its ability to effectively capture nonlinear relationships and temporal dependencies within river discharge data, which is often challenging for traditional regression-based models. During the training phase, TeaNet demonstrated strong predictive accuracy, as indicated by the high R² values (0.94, 0.91, 0.95, and 0.87) for Muri, Adityapur, Jamshedpur, and Ghatsila, respectively. The low RMSE values (2.34, 3.04, 1.98, and 2.81) and the minimal MAEs (1.13, 1.52, 1.16, and 1.47) reinforce the model’s capability to closely approximate the observed discharge levels. The MSE values (6.24, 5.33, 5.01, and 6.68) further confirm the minimal deviation of the predictions from the actual values. These findings highlight the model’s ability to effectively capture complex hydrological patterns and relationships within the training data, ensuring a high degree of accuracy in retrospective simulations.

Recent studies have explored the effectiveness of deep learning architectures in river discharge prediction [61] and investigated the performance of CNN and TCN models for streamflow forecasting in the Niger River Basin. The mentioned study demonstrated that the TCN outperformed other ML models for certain discharge stations, reinforcing the potential of convolutional architectures in hydrological modeling. Similarly, a study [62] developed a multi-attention encoder–decoder-based TCN for flood prediction, integrating a multi-head attention mechanism with a TCN to capture intricate temporal dependencies in hydrological data. Their findings indicate that incorporating attention significantly enhances the prediction accuracy, particularly for extreme hydrological events. Further extending the application of attention mechanisms in hydrological forecasting, a study [63] explored the effectiveness of spatiotemporal attention-based LSTM for water level prediction. Their study demonstrated that incorporating spatial and temporal attention significantly improved the forecasting accuracy, particularly in flood-prone regions. These findings further reinforce the role of attention-based models in capturing intricate dependencies, aligning with TeaNet’s approach of integrating an advanced attention mechanism alongside temporal feature extraction to enhance river discharge forecasting.

TeaNet’s performance was further evaluated using unseen data during testing to assess the model’s generalization capabilities. The results indicated slight performance degradation, as expected when transitioning from training to testing datasets. However, TeaNet maintained strong predictive accuracy, with good R² values (0.83, 0.80, 0.88, and 0.80) across the respective stations. The RMSE values slightly increased to 3.62, 4.12, 2.94, and 4.23, and the MAE values to 2.32, 3.58, 2.83, and 10.9, indicating moderate errors in prediction. The corresponding MSE values (9.29, 7.4, 6.28, and 9.32) reaffirm the model’s reliability, albeit with slightly reduced precision compared to the training phase. The observed reduction in the R² values during testing highlights the inherent challenges in hydrological forecasting, such as seasonal variability, extreme discharge events, and the influence of external environmental factors. Nevertheless, the consistently lower error metrics of TeaNet compared to other models suggest that it effectively mitigates these challenges by capturing key temporal patterns and reducing overfitting, which is often a limitation in traditional machine learning models.

Figure 4 depicts the comparison of the actual vs. forecasted river discharge using the TeaNet model for both the training and testing datasets across different gauge stations. Unlike conventional regression models, which struggle to adapt to the highly variable and nonlinear nature of hydrological systems, TeaNet demonstrated better learning capabilities by utilizing temporal dependencies and non-stationary relationships within the discharge data. The consistently higher R² values and lower RMSE, MAE, and MSE scores across both the training and testing phases reinforce its suitability for real-world hydrological applications. These findings align with previous research [64], which highlights the effectiveness of advanced deep learning frameworks in capturing the complexities of hydrological processes. Given its robust performance and adaptability, TeaNet presents a promising approach to long-term river discharge forecasting, contributing to improved water resource management, flood risk assessment, and climate adaptation strategies. To evaluate TeaNet’s robustness under extreme hydrological conditions, we analyzed its predictive performance during peak discharge periods across the test datasets. It was observed that TeaNet achieved high accuracy during monsoon seasons and high-flow intervals, as reflected in its relatively stable R² values (0.80–0.88) and low RMSE (2.94–4.23) across all four gauge stations, even under large discharge variability. Figure 4f–h particularly demonstrate TeaNet’s ability to closely track sharp rises in discharge during flood-prone periods at the Jamshedpur and Muri stations. Furthermore, under the SSP585 scenario, TeaNet effectively captures increasing discharge trends and peak intensities through 2100, reinforcing its applicability in extreme climate scenarios. These findings confirm the model’s resilience in handling both historical and projected high-flow events.

The observed spatial variability in TeaNet’s performance across the gauge stations, evident from the higher prediction accuracy at Muri (R²: 0.87–0.94) compared to Ghatsila (R²: 0.80–0.87), can be attributed to localized hydrological and geomorphological characteristics influencing the discharge dynamics. An analysis of the station-specific error patterns revealed that prediction uncertainties were strongly correlated with catchment-scale heterogeneity in the topography, land cover, and anthropogenic interventions. For instance, the Muri station, situated in the upper northern plateau (elevation: 231 m), exhibits steeper slopes (average gradient: 8–12°) and shorter concentration times, resulting in rapid rainfall–runoff responses that align well with TeaNet’s temporal attention mechanisms. In contrast, Ghatsila, located in the lower coastal plains (elevation: 72 m), features gentler slopes (1–3°) and extensive alluvial deposits, which introduce lagged baseflow contributions and complex groundwater–surface water interactions that challenge temporal modeling frameworks.

Land use patterns further explain the performance disparities: the Jamshedpur and Adityapur stations, located near urban–industrial zones, showed marginally higher MAEs (1.52–3.58) compared to rural Muri (1.13–2.32), likely due to anthropogenic flow regulation (e.g., dam releases, irrigation withdrawals) and impervious surface-induced flashes, which introduce nonlinearities that are not fully captured by climate-driven inputs alone. Additionally, the northern plateau’s lateritic soils exhibit lower infiltration capacities (5–15 mm/h) than the coastal plains’ alluvial soils (20–35 mm/h), leading to pronounced Hortonian overland flows during monsoon events that enhance the model predictability at upland stations. These findings align with hydrological studies demonstrating that the model performance degrades in basins with significant human modifications or heterogeneous subsurface characteristics, as such factors decouple the precipitation–discharge relationships in ways that are not resolved by meteorological inputs alone.

The interplay of climatic and physiographic factors is further evidenced by seasonal error trends: TeaNet achieved superior dry-season performance at Ghatsila (RMSE: 2.81–4.23) compared to monsoon periods (RMSE: 3.04–4.12), reflecting the model’s difficulty in capturing extreme precipitation–runoff thresholds exacerbated by the coastal plain’s low drainage density. Conversely, Muri’s consistent accuracy across seasons underscores the relative homogeneity of its granitic substratum and well-defined channel networks, which stabilize the rainfall–discharge correlations. These station-specific variations highlight the necessity of context-aware model interpretation, where local hydrogeological complexity directly modulates the predictive capabilities, even when using advanced temporal architectures.

4.3. Results Regarding Feature Importance Using TeaNet

Understanding the relative importance of different hydrometeorological variables in river discharge prediction is crucial in improving models’ accuracy and interpretability. The TeaNet model assessed the feature importance, revealing that the rainfall and temperature were the dominant factors influencing river discharge across all four gauge stations (Figure 5). This finding is consistent with fundamental hydrological principles, as precipitation directly contributes to river flows, while the temperature affects the evaporation rates, snowmelt (in relevant regions), and overall water balance within a basin [65,66]. Among all predictors, rainfall exhibited the strongest correlation with river discharge across Muri, Jamshedpur, Adityapur, and Ghatsila.

This strong correlation suggests that variability in precipitation patterns directly translates into fluctuations in river flows, reinforcing the role of rainfall as the primary driver of hydrological responses in the study area. The high feature importance of rainfall aligns with previous studies indicating that increased precipitation leads to an immediate rise in discharge levels, particularly in catchments with limited water storage capacities and high runoff potential [64]. Additionally, the relationship between rainfall and discharge may vary seasonally, with stronger correlations observed during monsoon periods, when precipitation is frequent and intense. Conversely, during drier seasons, the river flow may be influenced more by baseflow contributions from groundwater and the antecedent soil moisture conditions [67]. A study on the Subarnarekha River Basin in India supports this general pattern, demonstrating significant seasonal variations in rainfall and temperature, which in turn affect river discharge [67,68].

The second most influential factor identified by the model was the temperature, which plays a significant role in hydrological processes through its effects on evaporation, evapotranspiration, and snowmelt dynamics [69]. Higher temperatures can increase water loss due to evaporation, potentially reducing river discharge during dry periods. Conversely, rising temperatures can increase runoff generation in regions influenced by snowmelt, leading to a delayed but significant impact on discharge levels. The identified importance of the temperature suggests that long-term climate variations, particularly rising temperatures due to global climate change, could alter hydrological regimes by shifting seasonal runoff patterns and modifying peak discharge events. This aligns with recent hydrological studies emphasizing the need to integrate temperature-sensitive processes into predictive models for improved discharge forecasting under changing climatic conditions.

4.4. Futuristic River Discharge Predictions Under the SSP585 Scenario

Given the superior performance of the TeaNet model, river discharge projections were simulated up to the year 2100 under the high-emission SSP585 scenario. The results, as illustrated in Figure 6, indicate a gradual but significant increase in the projected river discharge across all four gauge stations (Muri, Jamshedpur, Adityapur, and Ghatsila). This projected rise in water flow suggests that the region may experience higher peak discharges, increased river volumes, and potentially more frequent extreme hydrological events, such as flash floods, in the coming decades. Moreover, the observed surge in river discharge can be primarily attributed to two key meteorological factors: increasing air temperatures and increased precipitation levels. These climatic changes, driven by the SSP585 scenario, are expected to significantly alter the region’s hydrological balance, influencing surface runoff, evaporation, and water availability [47].

The projected increase in the air temperature will play a crucial role in accelerating the hydrological cycle. Warmer temperatures intensify the atmospheric moisture-holding capacity, leading to higher evaporation rates and, consequently, increased precipitation levels (IPCC, 2021) [70]. This increased hydrological activity contributes to greater surface runoff, amplifying river discharge. Additionally, higher temperatures can influence the snowmelt dynamics in upstream areas (if applicable), leading to increased inflows into river systems during warmer seasons. Under the SSP585 scenario, the precipitation patterns are expected to intensify, contributing significantly to rising river discharge levels. Increased precipitation directly translates to greater runoff generation, particularly in areas with low soil infiltration capacities or regions where land use changes have reduced the natural water retention capabilities [64]. As a result, river systems are likely to experience prolonged periods of high discharge, along with more frequent occurrences of peak flow events, which could heighten the flood risks in vulnerable areas.

4.5. Implications for Water Resource Management and Flood Risk Mitigation

The effectiveness of proposed adaptation strategies can be further strengthened by quantifying their potential impacts on flood risk reduction and water resource resilience. For instance, recent large-scale flood risk modeling in Europe has shown that implementing optimal detention areas (retention basins) under a 3 °C warming scenario can reduce annual flood damages by 83% (74–89%) and population exposure by 84% (75–90%), with a benefit–cost ratio (BCR) of 4.2 (3.5–6.3), while annual investments of EUR 2.6 (1.9–3.8) billion are required over 2020–2100. Similarly, strengthening dyke systems has been predicted to lower annual flood damages by 70% (59–83%) and reduce economic losses by EUR 30 (23–43) billion per year by the end of the century, with annual investments of EUR 3.1 (2.1–4.5) billion. These quantified benefits highlight the substantial risk reductions achievable through infrastructure-based flood mitigation [71,72]. Nature-based solutions, such as large-scale afforestation and wetland restoration, have also demonstrated measurable benefits. For example, restoring floodplains and wetlands can enhance the water absorption capacity, attenuate flood peaks, and improve ecosystem quality [73], with studies reporting up to a 40% reduction in local flood risks and additional co-benefits for biodiversity and water quality. Integrating such ecosystem-based approaches is especially effective when combined with structural measures, providing both direct and indirect risk mitigation [74]. Regarding hydrological infrastructure adaptation, it is estimated that increasing the water storage capacity by 10–20% in vulnerable basins could offset up to 75–80% of the projected increase in water supply deficits under future climate scenarios, with the benefit–cost ratios exceeding 1 for most interventions. Upgrading urban drainage systems and retrofitting existing dams to accommodate higher discharge volumes can further reduce the risks of reservoir overflow and urban flooding [75], with cost–benefit analyses supporting these investments as economically viable for long-term resilience [76,77]. For sustainable water resource planning, adaptive water allocation strategies, such as shifting crop planting dates, improving the irrigation efficiency, and managed aquifer recharge have been shown to reduce the agricultural water demand by 5–80%, depending on the method and climate scenario. For example, shifting crop planting times earlier in the season can reduce evapotranspiration and the irrigation demand, while managed aquifer recharge can increase the groundwater supply reliability. These quantified adaptation options provide a practical basis for policymakers to prioritize interventions based on local vulnerability and resource availability [78]. Overall, integrating climate-informed decision-making with quantified adaptation benefits enables the formulation of robust, evidence-based strategies to mitigate the adverse impacts on agriculture, urban settlements, and critical infrastructure. The quantified evidence presented here underscores the economic and societal value of proactive adaptation in the Subarnarekha River Basin and similar regions facing climate-driven hydrological changes.

4.6. Uncertainty Analysis and Model Limitations

Long-term river discharge projections under climate change scenarios are subject to cascading uncertainties arising from multiple methodological and climatic factors, as demonstrated in hydrological studies [79,80]. These uncertainties can be broadly categorized into four interrelated domains, as described below.

Climate Model Structural Biases: GCMs exhibit systematic errors in simulating regional precipitation patterns and temperature trends, particularly in monsoonal regimes like the Subarnarekha Basin [81]. For instance, cold biases in the minimum temperature (0.8–1.4 °C) and wet biases in monsoon precipitation (12–18%) have been documented in CMIP6 ensembles, which propagate nonlinearly into discharge projections [82]. Such biases necessitate robust bias correction, although methods like linear scaling and Delta Change cannot fully resolve the distributional discrepancies in extreme precipitation quantiles [49].

Downscaling and Bias Correction Limitations: Statistical downscaling approaches, while computationally efficient, usually fail to preserve intervariable relationships between temperature and precipitation a critical factor in evapotranspiration estimation [79]. Recent comparisons show that simple scaling methods may amplify the errors in the projected extremes by 22–30% compared to distribution-based corrections [81].

Hydrological Model Uncertainties: Deep learning architectures like TeaNet, while effective in capturing temporal patterns, face challenges in representing long-term groundwater recharge dynamics and sediment-driven infiltration changes. Process-based models have demonstrated that neglecting these factors can overestimate the post-monsoon baseflow by 14–19% [83].

Unaccounted-For Anthropogenic Factors: Future land use changes, groundwater extraction, and reservoir operations—projected to increase by 40–60% in the basin by 2050—were not explicitly modeled [6]. Studies in similar basins indicate that unchecked urbanization could amplify the peak discharges by 18–24%, independently of climate impacts [84].

Interacting Uncertainties: First-order sensitivity analyses from the recent literature suggest that the precipitation intensity and antecedent moisture conditions dominate uncertainty propagation, contributing 63–71% of the variance in discharge projections [85]. The nonlinear coupling between these factors remains poorly constrained in CMIP6 models, particularly for sub-daily extremes [86].

Recommendations: (1) multi-model ensembles—combining outputs from 15–20 GCMs reduces the precipitation projection uncertainties by 25–30% compared to single-model approaches [80]; (2) advanced bias correction—multivariate methods like MBCn better preserve climate variable interdependencies than univariate scaling [87]; (3) dynamic land use integration—coupling hydrological models with land change projections improves flood risk assessment in urbanizing basins. While TeaNet provides robust near-term projections, post-2050 estimates should be interpreted probabilistically, with adaptive management strategies updated iteratively as new climate realizations emerge [70].

5. Conclusions

This study evaluated the predictive performance of four widely used machine learning models, the TCN, RF, SVR, and GRU, alongside the proposed TeaNet model for river discharge forecasting at a daily time scale in the Subarnarekha River Basin. The findings showed that TeaNet performed consistently better than all other models, as shown by its higher statistical performance across all evaluation measures. As a result, TeaNet was found to be the highest-performing model and was used to simulate the future daily river discharge trends from 2022 to 2100 using the SSP585 climate scenario. In addition, the river discharge projections in the future indicated a recognizable and persistent fluctuation in the river discharge trends over the duration of the study. The observed variations, characterized by periods of increase and decrease, were primarily attributed to anticipated changes in the precipitation and temperature, which were estimated using delta change factors derived from CMIP6 projections under SSP585. These results highlight the strong influence of climate variability on hydrological patterns, reinforcing the need for robust water resource management strategies to mitigate potential risks associated with extreme flow conditions, including flooding and drought events.

The efficiency of the TeaNet model in accurately simulating river discharge within the Subarnarekha Basin provides vital insights for hydrologists, water resource managers, and policymakers. However, despite its promising performance, this study acknowledges certain limitations. The analysis was conducted within a specific dataset and problem domain, and the generalizability of such machine learning models may vary depending on the data characteristics, regional hydrology, and climatic conditions. Additionally, while the TeaNet model demonstrated strong predictive capabilities, further refinements in data preprocessing, feature engineering, and model tuning could improve its adaptability to different water resource management contexts. This research highlights the potential of machine learning-based hydrological modeling in tackling the issues of climate-driven water variability. The results provide a scientific basis for enhancements in the real-time forecasting of river discharge and could inform adaptive water management decisions in the face of changing climatic conditions. Future studies need to include more hydro-meteorological variables, incorporate uncertainty quantification methods, and broaden the scope of ML-based forecasting models to enhance their reliability and applicability in varied hydrological regimes.

The practical implications of these findings are twofold. First, the demonstrated effectiveness of TeaNet in replicating complex hydro-climatic relationships suggests its viability for real-time operational streamflow forecasting, especially in monsoon-driven river basins. Second, the future discharge projections provide key understandings of adaptive water resource planning and risk management, guiding the development of climate-resilient infrastructure and policies to mitigate the adverse impacts of extreme hydrological events.

Overall, this work reinforces the utility of advanced machine and deep learning approaches in addressing emerging challenges in hydrology under changing climatic conditions and provides a robust scientific foundation for informed decision-making in water resource governance.