Exploring climate-change impacts on streamflow and hydropower potential: insights from CMIP6 multi-GCM analysis

The vulnerability of hydropower refers to its sensitivity and exposure to changing climatic conditions, which can subsequently affect its efficiency, reliability, and overall performance. Climate change is a globally acknowledged phenomenon with significant impacts on water resources worldwide. It induces shifts in precipitation patterns, glacier melt rates, and the timing of snowmelt in the Himalayas, profoundly affecting downstream river flows (Moors et al. 2009; Ali et al. 2018; Bolch et al. 2019). The rise in temperatures is anticipated to extensively influence rainfall patterns, leading to changes in intensity and volume and potentially altering hydrological cycles (Oki & Kanae 2006; Haddeland et al. 2014; Sarker 2022). Projections based on global climate models (GCMs) suggest a possible increase in the average annual global temperature by 2–5 °C by 2100, primarily due to the escalation of greenhouse gas (GHG) emissions (Herring 2012; Gao et al. 2017; Sarker 2022).

The Intergovernmental Panel on Climate Change (IPCC), in its Sixth Assessment Report, has projected an intensification of extreme climate events, underscoring the critical need to predict future streamflow variations (Singh et al. 2023). Mountainous regions, such as the Sutlej River basin, which depend heavily on snow and glacier-fed rivers, are especially susceptible to climate-change impacts (Sarker et al. 2019; Sarker 2021). The accelerated retreat of glaciers and altered runoff patterns threaten the availability and timing of water resources, posing significant challenges for agricultural irrigation, hydropower generation, and overall water security (Jeelani et al. 2012). To address the vulnerability of hydropower to climate change, it is essential to develop a comprehensive understanding of the hydrological processes that govern rainfall–runoff generation within a basin. This understanding can be achieved through the use of hydrological models, which simulate and predict basin responses to climatic inputs using mathematical equations and theoretical principles (Kirchner 2006; Sood & Smakhtin 2015; Singh et al. 2019). Lotfirad et al. (2023) evaluated CMIP6 GCM outputs for the Navrood Watershed, Iran, using bias correction and Bayesian model averaging to improve projection accuracy. Their findings indicate significant increases in T_min, T_max, and precipitation, leading to higher annual runoff under shared socioeconomic pathway (SSP) scenarios SSP245 and SSP585. This study highlights the importance of ensemble techniques in reducing climate projection uncertainties. In hydrological modeling, data-driven approaches have often demonstrated superior predictive accuracy across various applications (Shortridge et al. 2016; Adnan et al. 2020; Kabir et al. 2020; Herath et al. 2021; Singhal et al. 2024). Madhusudana Rao et al. (2020) highlighted the effectiveness of general circulation models over hypothetical scenarios for predicting climate-change impacts for the near future (2014–2040) under representative concentration pathway (RCP) 4.5 and RCP8.5 in the Subarnarekha River basin in eastern India.

However, the field has witnessed remarkable advancements in computational intelligence, particularly in machine-learning (ML) and deep-learning (DL) techniques (Adnan et al. 2020; Fu et al. 2020; Rahimzad et al. 2021; Ghobadi & Kang 2022). A notable example of this progress is seen in the study by Yang et al. (2019), where artificial neural networks, support vector regression (SVR), and random forests were employed to predict monthly streamflow in the Qingliu River basin, China, under varying environmental conditions. This study demonstrated that ML models consistently outperformed traditional process-based models in terms of predictive accuracy and adaptability. Shortridge et al. (2016) stated that there were lower streamflow prediction errors with ML models compared with physical models in Lake Tana and adjacent rivers in Ethiopia. Studies have integrated these models with GCM scenarios to project streamflow patterns under future scenarios. Das & Nanduri (2018) integrated relevance vector machine and support vector machine (SVM) models with the Coupled Model Intercomparison Project (CMIP) 5 GCMs to project monsoon streamflow within the Wainganga basin, India. Thapa et al. (2021) used a combination of long short-term memory (LSTM) models and CMIP5 GCM scenarios to assess streamflow patterns in the Langtang basin, situated in the central Himalaya region, and showed a significant projected increase in streamflow, primarily attributed to anticipated rises in precipitation.

However, applicability of these models for streamflow prediction in mountainous basins under future scenarios remains limited due to constraints associated with data availability (Xenarios et al. 2019; Adnan et al. 2020). To overcome the limitations of CMIP3 and CMIP5 GCMs in simulating extreme precipitation, CMIP6 GCM scenarios are suggested to offer a more realistic representation of future rainfall and temperature patterns (Gusain et al. 2020; Kim et al. 2020). Hence, it becomes crucial to investigate the efficacy of ML methodologies in forecasting streamflow within mountainous river basins, utilizing hydrometeorological data in conjunction with CMIP6 GCM scenarios as input.

The study of GCM evaluation and ranking contributes to enhancing the accuracy and reliability of climate predictions (Lotfirad et al. 2023). GCM output acts as a valuable dataset to address the challenges built up by climate change and its impacts on hydrological processes. These advancements enable us to practically plan for water resource management, ensuring resilience and adaptation in the face of an uncertain future (Xu et al. 2005). Properly evaluating GCMs using a range of statistical performance metrics, such as skill score (SS), correlation coefficient (CC), Nash–Sutcliffe efficiency (NSE), normalized root mean square deviation (NRMSD), coefficient of determination (R²), and absolute normalized mean biased deviation (ANMBD), is necessary to identify reliable models (Kheireldin et al. 2020; Shiru & Chung 2021). Weighting these metrics and employing distance measure techniques, such as the combined compromise solution (CoCoSo), ensures a comprehensive assessment of a GCM's performance and its suitability for climate impact assessments and future climate projections.

The Bhakra Dam and the Sutlej River basin have been subject to intensive study regarding climate–water interactions and hydrological responses to climate change. Ali et al. (2018) developed a hydrological model to assess streamflow sensitivity to climate change scenarios and identified potential shifts in the timing of the monsoon season using RCP2.6 and RCP8.5 in various reservoirs including the Bhakra Nangal reservoir. In the Bhakra Nangal reservoir, a 15.30% ± 20.16% increase in mean annual precipitation by the end of the 21st century and a subsequent increase in streamflow and development of hydropower potential were predicted. Gupta & Sharif (2021) evaluated CMIP5 GCMs for simulating climate over South Asia, which includes the Upper Sutlej River basin (USRB) and highlighted the significance of GCM performance in simulating the regional climate. Singh et al. (2015a) evaluated streamflow sensitivity to glacier mass balance in the Bhagirathi River basin, which is geographically close to the Sutlej River basin, indicating the potential impacts of glacier withdrawal on water availability in the region. Massive hydropower dams in the Mekong River basin impact ecology and biodiversity, particularly due to monsoon season shifts and glacier retreat. This study identifies critical nodes and harmful dams, emphasizing the need for accurate streamflow prediction to understand and mitigate these climate–water interaction effects on the MRB's ecological integrity (Gao et al. 2022). These previous studies have focused on specific aspects of climate–water interactions, such as monsoon season shifts and glacier retreat impacts, and emphasized the need for accurate streamflow prediction.

However, there is a research gap in understanding the integrated and cumulative effects of climate change on hydropower potential in the region using the latest CMIP6 multi-GCM. The current study has provided valuable insights that consider various climatic factors and their interconnected impacts on streamflow and water availability for hydropower potential. This paper aims to evaluate the performance of GCMs by considering atmospheric patterns from CMIP6 models in modeling rainfall, maximum temperature (T_max), and minimum temperature (T_min) in the context of climate change assessment for the study area. However, certain challenges, particularly concerning data assimilation, arise when working with coarse-resolution scenario data from GCMs, limiting their direct applicability in regional impact assessments (Hagen et al. 2021; Adib & Harun 2022). Consequently, to address these issues, this study employs the National Aeronautics and Space Administration (NASA) Earth Exchange Global Daily Downscaled Projections (NEX-GDDP-CMIP6) dataset. The dataset is generated using the Bias-Correction Spatial Disaggregation (BCSD) method, a statistical downscaling technique specifically designed to overcome the current limitations associated with global GCM outputs (Raghavan et al. 2018; Dross 2023; Mohammed et al. 2023). The NEX-GDDP-CMIP6 climate projection is downscaled at a spatial resolution of 0.25° × 0.25° (approximately 25 × 25 km²).

This study uses an innovative approach to examine how climate change affects hydropower generation by combining DL techniques with advanced climate models. Specifically, it employs the LSTM model integrated with the latest CMIP6 multi-GCM datasets at a high resolution of 0.25° × 0.25°, chosen using entropy and CoCoSo methods. This technique offers a detailed projection of streamflow changes and potential increases in hydropower generation at the Bhakra Dam site from 2015 to 2050. The research is conducted under two SSP scenarios, moderate (SSP245) and severe (SSP585) climate change, using the mean multimodel ensemble of CMIP6. This thorough analysis provides new insights into future hydropower potential in the USRB, adding valuable knowledge to water resource management under climate change conditions. The primary aim of this study is to identify an appropriate model using combinations of statistical performance metrics and to investigate how these metrics affect the ranking of each GCM. Six statistical performance indices, namely, NSE, CC, NRMSD, ANMBD, SS, and R², are utilized to assess the performance of 14 CMIP6 multi-GCMs within the study area. Another objective of this research is to identify suitable GCMs to serve as references for hydroclimate impact studies in the USRB. The selection of appropriate GCMs is based on simulations from the 14 CMIP6 multi-GCMs.

The average rainfall across the study area varies with a yearly range of 696–1,651 mm. Rainfall–runoff contributes largely to the lower part of the watershed (Gogineni & Chintalacheruvu 2024a, 2024b). The average annual highest and lowest temperatures in the area are 31 and 18 °C, respectively. The mean annual discharge (averaged throughout 1985–2014) of the river gauged at Bhakra was 885.6958 m³/s. There is a significant interdiurnal and monthly variation in the river discharge pattern. The minimum and maximum daily discharge recorded at Bhakra was 87.82 and 3,393.93 m³/s, respectively.

The power generation facilities at the Bhakra Dam station are distributed across two powerhouses, one on the right bank and the other on the left bank, each with five turbines or power units (Rao & Ramaseshan 1985). Five units, initially having a capacity of 120 MW each, were upgraded to 157 MW each, while the remaining five units with 90 MW each were upgraded to a capacity of 108 MW each. Hence, the installed power capacity of 1,050 MW was increased to 1,325 MW (Rao & Ramaseshan 1985; CEA 2019). The study area is experiencing challenges due to climate change and anthropogenic activities (Singh et al. 2015b; 2015c). This could change the streamflow in future and eventually affect hydropower production (Singh et al. 2014). In the current study, the future period (2015–2050) of hydropower generation is predicted for the Bhakra Dam site of the USRB under two GHG trajectories, SSP245 and SSP585, using a mean ensemble of selected models.

Figure 2 illustrates the methodology employed in this study for the selection of an appropriate individual model and ensemble model tailored to estimate the projected hydropower potential of the study area.

Daily data for 30 years (1985–2014) at the Bhakra Dam station is used to examine the statistical performance metrics of GCMs and to simulate streamflow through the ML model. Observed values of rainfall, T_max, and T_min are obtained from the Indian Meteorological Department (IMD); discharge data are obtained from the Central Water Commission (CWC), New Delhi, India; wind speed, solar radiation, and relative humidity data are extracted from the Climate Forecast System Reanalysis (CFSR) Global Weather Data (http://globalweather.tamu.edu/).

The most recent generation of climate models, known as CMIP6 GCMs, is provided by the NASA NEX-GDDP-CMIP6. The NEX-GDDP-CMIP6 dataset enhances accuracy through two key steps: statistical downscaling and bias correction. Statistical downscaling employs the BCSD method, which corrects biases in GCM outputs by aligning them with historical observations and spatially disaggregates data to a finer 0.25° × 0.25° grid. This ensures that local-scale projections maintain consistency with large-scale climate patterns. The quantile mapping (QM) technique is used for bias correction, where the cumulative distribution functions (CDFs) of modeled and observed variables are matched. This process adjusts the modeled data to correct systematic biases in mean, variance, and higher-order moments, ensuring that the corrected outputs statistically align with observed data. By combining BCSD for downscaling and QM for bias correction, the NEX-GDDP-CMIP6 dataset provides spatially detailed and statistically consistent climate projections, improving the reliability of future climate assessments (Wood et al. 2002, 2004; Maurer & Hidalgo 2008; Thrasher et al. 2012). These methods provide high-resolution climate data that are crucial for reliable impact assessments, such as the evaluation of streamflows and hydropower potential in this study. They are being employed to project streamflow for the near future, specifically the period from 2015 to 2050. The framework of CMIP6 multi-GCM is specifically designed to simulate future climate scenarios based on four GHG trajectories, known as the SSPs (SSP126, SSP245, SSP370, and SSP585). These scenarios aim to explore potential GHG emissions in the coming decades, taking into account various global socioeconomic changes that may occur by the year 2100 (Riahi et al. 2017; Karan et al. 2022). The NASA-NEX-GDDP dataset comprises climate predictions of finer scale from simulations developed by CMIP6 GCMs. NASA NEX-GDDP-CMIP6 was accessed on 3 June 2023 from https://registry.opendata.aws/nex-gddp-cmip6.

The dataset was obtained for 14 CMIP6 multi-GCMs (1985–2014) to examine their performance by simulating with observed data. Furthermore, selected GCM datasets were obtained to generate climate projections under the SSP245 and SSP585 scenarios (2015–2050). SSP245, characterized as a medium scenario, represents the typical trajectory for future GHG emissions, resulting in a radiative forcing of 4.5 W and m⁻² by the year 2100. In contrast, SSP585 stands as an upper-limit scenario within the spectrum, projecting a radiative forcing of 8.5 Wm⁻² by the year 2100 (O'Neill et al. 2016). These datasets are available at a daily time resolution and a horizontal spatial resolution of 0.25° × 0.25°.

The study area is covered by ten grids of downscaled CMIP6 multi-GCM data. Rainfall and temperature values at each grid were subsequently averaged across the catchment area using the Thiessen polygon method. This process generated daily rainfall data integrated at the catchment scale, facilitating the assessment of climate changes in comparison with the observed period from 1985 to 2014. In addition, a ranking process was conducted on the CMIP6 multi-GCM to identify the most suitable models for generating realistic future climate scenarios within the catchment area. These selected models will play a crucial role in projecting streamflow and hydropower potential.

The performances of GCMs were examined by simulating with observed data through the statistical performance metrics mentioned. The six parameters used are listed in Table 1 (SS, CC, NSE, R², NRMSD, and ANMBD); to depict the relation between GCMs and observed datasets, statistical performance metrics were chosen based on the literature.

Table 1

Statistical performance metrics used for the study

S. no. .	Statistical performance metrics .	Equations .
1	SS	fo and fs are frequencies of the historical and simulated values, ci are class intervals, z is the number of metrics, and x is the number of datasets
2	CC	So and Ss are standard deviations of observed and simulated values;
3	NSE	Qo and Qs are observed and simulated values of streamflow; and are mean observed and simulated values
4	R2
5	NRMSD	Oi and Si are the observed and simulated values
6	ANMBD	i is the dataset number varying as 1, 2, …, x

S. no. .	Statistical performance metrics .	Equations .
1	SS	fo and fs are frequencies of the historical and simulated values, ci are class intervals, z is the number of metrics, and x is the number of datasets
2	CC	So and Ss are standard deviations of observed and simulated values;
3	NSE	Qo and Qs are observed and simulated values of streamflow; and are mean observed and simulated values
4	R2
5	NRMSD	Oi and Si are the observed and simulated values
6	ANMBD	i is the dataset number varying as 1, 2, …, x

View Large

A new generation-developed technique of multicriteria decision-making (MCDM), CoCoSo, developed by Yazdani et al. (2019), is used to rank the GCMs. CoCoSo is an integrated, simple additive weighting and exponentially weighted product-based method leading to the execution of the compromise solution. This method has demonstrated its effectiveness by offering a flexible solution that accommodates multiple alternatives assessed across various conflicting scenarios, resulting in corresponding assessment scores. It influences distance measures, thereby enhancing its flexibility. The ranking performance of the CoCoSo method has been shown to be highly competitive with well-established MCDM techniques such as multiobjective optimization based on ratio analysis and complex proportional assessment. Due to its consistent and reliable ranking performance, CoCoSo has gained widespread acceptance in the field (Adar et al. 2022). Primarily, the initial decision-making matrix is developed for six statistical performance metrics to execute CoCoSo. Among the metrics considered for the analysis, CC, R², NSE and SS, a value of 1 is ideal, whereas for NRMSD and ANMBD, a minimum value of 0 is desirable (Nash & Sutcliffe 1970; Gupta et al. 1999; Adnan et al. 2020; Sreelatha & Anand Raj 2021). Therefore, in the study, CC, R², NSE, and SS are the beneficial criteria, and NRMSD and ANMBD are the non-beneficial criteria. For CoCoSo, the following steps are validated based on the alternative and their related criteria.

Step 1: Development of initial decision-making matrix

X = [X_ij]_m×n, where X_ij denotes the performance score of the ith alternative with respect to the jth criterion.

Step 2: Application of linear normalization technique

Step 3: Estimation of corresponding performance indexes S_i and P_i for each of the alternatives

Step 4: Estimation of three different appraisal scores

LSTM is a new generation version of the recurrent neural network. It has the advantage of learning from long-term dependency in sequences, making it successful in many fields, such as rainfall–runoff modeling in hydrology. The application of ML models based on DL like LSTM, deep neural networks, and convolutional neural networks are being extensively utilized in the prediction of streamflow due to capabilities of handling large and complex stochastic datasets and abstracting the internal physical mechanism (Fu et al. 2020; Ghobadi & Kang 2022; Roy & Rao 2024). In the daily streamflow prediction of the Kentucky River basin in the United States of America, a study based on a statistical performance estimation criteria showed that LSTM models outperformed the linear regression, SVR, and multilayer perceptron models (Rahimzad et al. 2021). This study exclusively utilized LSTM, a data-driven model, for streamflow and hydropower estimation, bypassing traditional physical models. LSTM's strength lies in capturing complex patterns from historical data, essential for accurate predictions. Table 2 shows the information regarding hyperparameters used for estimating model parameters in this study.

Window size is a hyperparameter that defines the number of time steps in the input sequence used for prediction. For this study, 280 window size is taken. Sixty-four dense layer units are taken as the number of neurons in the dense layer. A rectified linear activation function (ReLU) is used in the neural network layers, and linear activation is used in the final output. The learning rate which controls the step size during optimization is defined as 0.0001 for the Adam optimizer (Ghimire et al. 2021; Shu et al. 2021). The Adam optimizer is a widely used optimization algorithm in DL that takes care of updating model weights during training. In the study model, the epoch number is 30. Epochs are the number of training iterations or passes through the entire dataset. The number of LSTM units taken is 64, which determines the number of LSTM units in the LSTM layer of the study model. These are the primary hyperparameters used in the LSTM code for time-series forecasting.

Alteration of hyperparameters can significantly impact the model's performance, and tuning of these parameters is required to optimize the model for specific datasets and tasks. The suitability of a hydrological model for a specific application depends on how accurately it can match simulated flow with observed flow during the training (calibration) and testing (validation) period (Refsgaard 1997). LSTM's utilization in this study allows for the incorporation of multiple GCM data, enhancing the strength and comprehensiveness of the analysis conducted at the Bhakra Dam site. It learns patterns and relationships directly from historical data without incorporating physical laws. To assess model predictions, various methods have been developed, covering quantitative statistics and graphical techniques (Legates & McCabe 1999). Moriasi et al. (2007) categorized these methods into three groups: standard regression, dimensionless, and error index. Each category serves a distinct purpose, such as explaining the relationship between observed and simulated values, comparing model performance, and quantifying deviations in the data units. It is well recognized that a single metric is inadequate to comprehensively evaluate a model's performance; hence, multiple metrics must be evaluated (Adnan et al. 2020).

In this study, the competency of the LSTM model was estimated using five statistical parameters: Table 3 shows the R², NSE, mean absolute error (MAE), percent bias (PBIAS), and root mean square error (RMSE). Moriasi et al. (2007) established guidelines for the assessment of hydrological models and ranking them based on several metrics. Model performance is categorized as very good, good, satisfactory, or unsatisfactory when the NSE value is between 0.75 and 1, 0.65 and 0.75, 0.50 and 0.65, or less than 0.50, respectively. Similarly, R² is used to classify models as satisfactory, very good, or unsatisfactory if a value is between 0.6 and 0.7, 0.85 and 1, and below 0.5, respectively. Lower values are preferred, whereas 0 is ideal for PBIAS, RMSE, and MAE. Highly acceptable PBIAS values are below ±10%, while values exceeding ±25% are considered unsatisfactory. Negative PBIAS indicates overestimated bias, while positive suggests underestimated bias (Gupta et al. 1999; Ritter & Muñoz-Carpena 2013).

Table 3

Statistical parameters used to check LSTM model performance

S. no. .	Parameters .	Equation .
1	MAE
2	PBIAS
3	RMSE
4	NSE
5	R2
	Pi and Qi are predicted and observed values respectively; n is the number of samples; and and denote the mean of the observed value and predicted value, respectively.

S. no. .	Parameters .	Equation .
1	MAE
2	PBIAS
3	RMSE
4	NSE
5	R2
	Pi and Qi are predicted and observed values respectively; n is the number of samples; and and denote the mean of the observed value and predicted value, respectively.

View Large

The hydropower potential at the Bhakra Dam site has been evaluated by applying the discharge values in Equation (12); ƞ = 85%, and a net head for the generator (H) of 116 m (Rao & Ramaseshan 1985).

Tables 4–6 show the statistical performance metric values for 14 CMIP6 multi-GCMs and ranking for varying weight scenarios of rainfall, T_max, and T_min, respectively. Using the six performance metrics, the performance of 14 CMIP6 multi-GCMs in modeling three climate variables (rainfall, T_max, T_min) in the USRB was compared with the observed data (1985–2014) of IMD for validating the GCM dataset. Weights for the metrics are calculated using the entropy method, and following the steps of CoCoSo, ranking has been done as a result of the performance comparison of the 14 CMIP6 multi-GCMs. For this study, only CMIP6 GCMs with NSE > 0.50 and R² > 0.50 are considered, ensuring the condition defined by Moriasi et al. (2007). In rainfall metrics, CanESM5 and NESM3 exhibited negative NSE, and only six showed relatively good performance with NSE and R² > 0.50. Results showed that while some GCMs exhibited good agreement with observed data, others had notable biases, particularly in capturing the magnitude of extremes. This thorough validation ensures the strength of the climate projections used in the study, enhancing the credibility of the findings related to changes in streamflows and hydropower potential at the Bhakra Dam site under future climate scenarios.

Further, in T_max and T_min, negative NSE is presented by CanESM5 and NESM3, respectively, and 11 and ten CMIP6 multi-GCMs show R² and NSE >0.50. Considering the three climate variable metrics (rainfall, T_max, T_min) and their CMIP6 multi-GCM ranks, the final ranking has been done (using Equation 11); hence, the selected six models as per their rank order are (1) Centro Euro-Mediterraneo sui Cambiamenti Climatici Climate Model 2, Scenarios RCP 5 (CMCC-CM2-SR5); (2) Max Planck Institute for Meteorology Earth System Model version 1.2 with higher resolution (MPI-ESM1–2-HR); (3) Norwegian Earth System Model version 2 – Medium Resolution (NorESM2-MM); (4) Meteorological Research Institute Earth System Model version 2.0 (MRI-ESM2-0); (5) Centro Euro-Mediterraneo sui Cambiamenti Climatici – Earth System Model Version 2 (CMCC-ESM2); and (6) Beijing Climate Center Climate System Model 2, Version MR (BCC-CSM2-MR). Within the selected models, NSE and R² range from 0.50 to 0.73 and 0.51 to 0.80, respectively, for rainfall; 0.67 to 0.78 and 0.76 to 0.88, respectively, for T_max; and 0.71 to 0.85 and 0.82 to 0.9, respectively, for T_min. These six models were studied to examine the streamflow and hydropower potential pattern for the future period (2015–2050) in the USRB.

LSTM modeling provides predicted values by capturing and learning patterns in sequential data of historical time-series (Ghimire et al. 2021). To train the model, daily observed data from 1985 to 2014 at the Bhakra gauging station is utilized, and CMIP6 multi-GCMs generate historical projections. The implementation of LSTM simulations was executed using Python programming. The models utilized seven variables as input: rainfall, T_max, T_min, discharge, wind speed, solar radiation, and relative humidity; other meteorological parameters remained constant throughout the modeling processes. The entire dataset is divided into training, validation, and testing subsets. The training subset comprises 80% of the total data facilitating model fitting, while the testing and validation subset consists of 10% each of the total data to assess model accuracy.

The predicted range shows an increase in mean monthly streamflow; the mean ensemble of the model shows an increase in streamflow throughout the year from 2.06% to 15.90% under SSP245 and from 4.43% to 17.36% under SSP585. However, in August under SSP245, a decrease in streamflow is predicted by 5.9%; and in July and August, it decreases by 3.9% and 4.6%, respectively, under SSP58. Under both scenarios, the maximum rise in streamflow is predicted in February and November, which is about 15% under SSP245 and about 17% under SSP585 compared with the reference period. The projected decrease in streamflow under a warming climate and increase in precipitation is consistent with previous studies, for instance, Ali et al. (2018) and Singh et al. (2023). A clear pattern (seasonal increase or decrease) is not observed in the predicted range of mean monthly streamflow. The deviation aligns with the findings of other researchers (Singh et al. 2015a; Gusain et al. 2020; Kim et al. 2020). Variations might influence inconsistencies in streamflow estimates in climate-variable projections arising from differences in spatial resolutions and parameterization levels within climate models (Sperna Weiland et al. 2010; Singh et al. 2015a). The averaging of an ensemble of GCMs mitigates the errors inherent in individual models, and as the number of models included in the ensemble increases, the overall uncertainty in the ensemble decreases (Murphy et al. 2004). Consequently, to minimize the uncertainty associated with streamflow projections linked to individual CMIP6 multi-GCMs, the study examined the streamflow pattern of the USRB by incorporating the mean ensemble data from all six GCMs. Flow duration curves (FDCs) describe the relationship between the probability of exceedance of time and flow magnitude, and are used to assess hydropower potential (Post 2004).

The performance of 14 GCMs within the CMIP6 framework was analyzed using six performance statistics: SS, NSE, CC, R², NRMSD, and ANMBD for predicting rainfall, T_max, and T_min of USRB. The entropy technique was used to assign weights to each performance metric and the CoCoSo technique ranked the multi-GCM based on performance. BCC-CSM2-MR, CMCC-CM2-SR5, CMCC-ESM2, MPI-ESM1-2-HR, MRI-ESM2-0, and NorESM2-MM are the six selected models out of the 14 CMIP6 multi-GCMs. Bias-corrected statistically downscaled data at 0.25° × 0.25° grid resolution of the six selected CMIP6 multi-GCMs were used in LSTM, a DL model which shows competence in predicting future streamflow at the Bhakra Dam site, during 2015–2050 under two GHG trajectories SSP245 and SSP585. The performance statistics of the LSTM model show MAE = 56.36, NSE = 0.93, RMSE = 131.87, R² = 0.93, and PBIAS = −3.01% in training. They are 59.58, 0.91, 151.73, 0.91, and −4.39% in validation; and 52.92, 0.95, 107.61, 0.95, and −1.68% in testing, respectively. The mean ensemble of models shows an increase in mean monthly rainfall, T_max, and T_min in most of the months of the future period (2015–2050) under both scenarios SSP245 and SSP585 (Gupta & Sharif 2021; Singh et al. 2023). Due to the projected increase in rainfall, mean monthly streamflow and hydropower potential are projected to rise under the future climate scenario. However, predicted rise in T_max and T_min shows significant warming to result in a decline in streamflow and hydropower production in July and August at the Bhakra Dam site hydropower projects (Ali et al. 2018). Also, significant variations in the streamflow pattern as well as hydropower potential were observed throughout the periods of months, seasons, and years for the CMIP6 GCMs. The FDC developed to assess the dependability flows at Q50, Q75 and Q90 show the rise of streamflow by 13% and 21%, 16% and 17%, 17% and 18% compared with the reference period (1985–2014) under SSP245 and SSP585, respectively.

The distinct variations in spatial resolution and parameterization levels in CMIP6 GCMs make these advantageous aspects to predict the streamflow and hydropower potential for future periods. These latest advancements in GCMs are contributory in generating valuable insights of projected changes in temperature and rainfall during the study period. The results provide insights into the development and planning of hydropower projects in India under the currently projected future climate. This study provides valuable insights for strategic planning and effective decision-making in the face of changing climatic conditions, ensuring sustainable water resource management.

While this research provides valuable insights into the potential impacts of climate change on streamflow patterns and hydropower generation at the Bhakra Dam site, several limitations must be noted. First, the study's reliance on only six out of the 14 bias-corrected, statistically downscaled datasets from CMIP6 multi-GCMs may not fully capture the complete range of uncertainties inherent in climate models, potentially biasing the results. Second, the data resolution at a 0.25° × 0.25° grid may not sufficiently represent local-scale variations that are critical for accurate hydrological predictions. In addition, while the LSTM model is a powerful tool for time-series forecasting, its performance can be sensitive to the length and quality of training data and the choice of hyperparameters. Moreover, estimating hydropower potential involves complex interactions beyond streamflow patterns, such as infrastructure and operational constraints, which may not be fully captured in the analysis. Addressing these limitations could enhance the robustness and reliability of the study's findings regarding climate-change impacts on streamflows and hydropower generation.

The authors acknowledge the Indian Meteorological Department (IMD) for rainfall and temperature data and the Central Water Commission (CWC), New Delhi, for discharge data used in this study.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.