Assessment of climate change impacts on rainfall and streamflow in the Alto Paranapanema Basin, Brazil

The latest Assessment Report (AR6) published by the Intergovernmental Panel on Climate Change (IPCC) in August 2021 unequivocally indicates that human influence is responsible for the rapid pace of climate change, exceeding what would – otherwise – occur naturally. The report confirms that greenhouse gasses emitted from human activities are the main culprit for climate change (IPCC, 2021). The impacts of climate change are widespread and diverse, ranging from rising sea levels and altered rainfall patterns to changes in evapotranspiration and energy supply in nations heavily reliant on hydroelectric power. In addition, climate change poses a significant threat to food security (Zhang et al. 2018).

The relationship between human activity and climate prediction dates to the origin of agriculture, which marked the first large-scale food supply for humanity and was essential for population growth and technological development (Commoner 2020). However, nowadays, the planet faces significant disruptions to its ecosystems, and there are important questions regarding the maximum population it can sustain. In the early days of agriculture, people viewed the seasons as milestones to determine when to plant different crops and how long food stores would last. Today, this concept of carefully observing the climate behavior is more relevant than ever, as human activity is heavily dependent on water resources, and climate change is altering the spatial and temporal distribution of rainfall. While modern climate prediction requires advanced supercomputers, it remains as essential as it was for our ancestors (Jackson et al. 2001).

The complexity of climate, including the multiple factors that influence weather patterns and climate change, makes accurate prediction remarkably difficult (Gueymard 2012). However, statistical models offer a means of predicting climate trends by analyzing historical data, despite inherent uncertainties. One useful tool for climate prediction is the stochastic climate generator, which can simulate hydrological series based on historical data. To assess the effectiveness of such simulations, several common methods are employed to determine the degree of statistical correlation with historical data, such as the Nash–Sutcliffe efficiency (NSE) model coefficient.

Global Climate Models (GCMs) are computer simulations used to predict how the climate will respond to future changes (Chou et al. 2014a, 2014b). The outputs of GCMs can be incorporated as inputs to stochastic models, which can generate projections based, e.g., on the IPCC Representative Concentration Pathways (RCPs), indicating how much greenhouse gas emissions are expected to increase (van Vuuren et al. 2011). GCMs are generated and provided by several institutions, including the Canadian Centre for Climate Modelling and Analysis (CCCma), the Center for Climate System Research (CCSR), the National Institute of Space Research (INEP), and the Met Office Hadley Centre. To improve their accuracy, GCMs can also be downscaled or regionalized by incorporating additional variables, such as topography, land use, or local climate data that are specific to the region under analysis. In Brazil, the regionalization process is performed by the National Institute for Space Research (INPE), which uses multiple GCMs to better represent and understand possible future climate scenarios in the Brazilian territory (Chou et al. 2014b).

Simulating long-term climate datasets for a specific region and altering its initial conditions is a feasible approach to understand the impact of climate change on future rainfall, evapotranspiration, and river flow (Maraun et al. 2010; Green et al. 2011). To this end, in Brazil, hydrologists widely employ the Soil Moisture Accounting Procedure (SMAP) model, which utilizes rainfall and evapotranspiration data to derive river flow. The SMAP model generates both unit runoff values, a measure of the watershed's water production per unit area, and total river flow discharge (Barros et al. 2009). Further details on SMAP can be found in Lopes (1999) and Tiezzi et al. (2018, 2019).

The relationship between rainfall and runoff is highly nonlinear and depends on specific watershed characteristics, such as land use and basin topography. The interconnection between different sub-basins has a significant influence on the river flow downstream, as the runoff from upstream watersheds affects the river flow in the watersheds downstream, a phenomenon named as incremental watersheds. It is well known that although headwater watersheds are located miles away upstream, they can significantly impact the quantity and quality of water downstream along the river path (Skaggs et al. 1994).

The aim of this study was to investigate the potential impacts of climate change on rainfall and river flow in two key watersheds, Jurumirim and Chanvantes. These watersheds play a critical role in supplying water to cities in the Avare region, besides their importance for the local economy, including tourism and property values around the reservoir. Climate change can significantly alter hydrologic processes in the watersheds, which in turn can affect communities along the river, disrupt electricity generation in hydroelectric plants, and impact agriculture and livestock practices in the region. Furthermore, these changes can have a ripple effect throughout all connected rivers. Despite the abundance of research on the impacts of climate change on hydrological systems, few studies have incorporated such many climate models and scenarios for basins with a major impact on society, as is the case in this study.

In Brazil, one of the major headwater watersheds for hydroelectric generation is the Alto Paranapanema (ALPA) Basin, in the State of São Paulo. The Rio Paranapanema, in which the ALPA is the main headwater sub-basin, has 11 hydroelectric power plants located in this basin with a total output of 2.4 GW. This basin is also one of the major tributaries of the Parana River, that leads to the Itaipu dam, the second biggest hydroelectric power plant in the world regarding installed capacity (14 GW).

The Chavantes power plant has a production capacity of 414 MW and a reservoir area of 428.34 km², while the Jurumirim power plant has a production capacity of 100.96 MW and a reservoir of 449 km². It is worth noting that the water flowing through this basin comes from the Paranapanema River and the Itararé River, and it plays a crucial role in regulating water resources in the region.

GCMs are simulated by several research institutions around the world, each using their own models. GCMs used in this study were run with greenhouse gas emissions projected by the IPCC Fifth Assessment Report. The grid size for these models typically ranges from 400 × 400 km to 200 × 200 km. The Regional Climate Models (RCMs), like the Eta model, takes the GCM and focuses computational efforts to improve the output data. Table 1 summarizes the main features of the GCM models used in this study (BESM, CanESM2, HadGEM2-EM, Miroc5).

The Eta model has been adapted to facilitate downscaling, which involves regionalizing GCMs on a more precise scale, tailored to the Central and South American regions. It has been utilized by CPTEC/INPE for climatological and meteorological forecasts since 1997. Starting from 2010, this model was also employed for studying climate change driven by global models. The Eta model incorporates the annual dynamics of the microphysical vegetation cycle of clouds, a convective cloud scheme, and short- and long-wave balance at a constant CO₂ concentration (Chou et al. 2014a, 2014b).

The GCMs used in this study (see Table 1) were selected as these were the only four GCMs regionalized using the Eta model. These models were coupled resulting in four variants: Eta–BESM, Eta–CanESM, Eta–HadGEM2-ES, and Eta–Miroc5. Precipitation and evapotranspiration data spanning the period from 1960 to 1989, as well as projections up to 2,100 based on RCPs, were obtained using these models. The Eta model employs a vertical resolution of 38 levels and a spatial resolution of 20 km (Chou et al. 2014a). It is important to clarify that although the Eta models generate a significant amount of data, only precipitation and evapotranspiration were considered for modeling the rainfall–runoff process in this study.

RCP 4.5 and 8.5 were the scenarios employed for the different regionalized Eta models, as they represent the best- and worst-case scenarios projected by the IPCC. The RCP 4.5 scenario represents a radiative forcing level, or energy balance in the atmosphere, caused by human activity of 4.5 W/m², and is a stabilization scenario. The RCP 8.5 scenario represents a high emissions scenario, with a radiative forcing level of 8.5 W/m² (IPCC, 2021). The RCP 2.6 scenario was omitted from this study due to its projected imminent occurrence. The primary objective of this study is to examine the potential ramifications if the scenarios deteriorate. Likewise, the RCP 6.0 scenario was not employed to concentrate solely on the extreme cases and streamline the analysis process.

To enhance the accuracy of the models employed in this study, a meticulous calibration process was undertaken using historical data obtained from sound sources such as the Brazilian National Water Agency (ANA) and the Brazilian National System Operator (ONS). Typically, historical series spanning a period of five to ten years are used to convert rainfall into streamflow through the application of a numerical simulation model. In this study, the SMAP model was specifically adopted for this purpose. This model has been extensively used in previous research and has undergone validation in multiple studies (e.g., Barros et al. 2009; O'Neill et al. 2010; Zhang & Zhou 2016; Wigneron et al. 2017; Peng et al. 2021). The SMAP model operates by calculating flow based on precipitation and evaporation and is classified as a lumped parametric model and is based on the Standard Watershed IV model and on the Mero model. SMAP was originally developed for daily data, and later updated to handle hourly and monthly data (Lopes 1999). This study uses the monthly version of the SMAP model. SMAP was developed by Lopes et al. (1982) and was presented at the International Symposium on Rainfall–Runoff Modeling that took place at Mississippi State University (USA). To optimize the model's performance, the NSE coefficient (Nash & Sutcliffe 1970) was selected as the optimization criterion (employing Excel's SOLVER function).

To ensure an accurate interpretation of the dimensionless average analysis, it is fundamental to consider the potential for misinterpretation. For instance, if we consider a dry period with a historical average rainfall of 10 mm and a model indicating an additional 10 mm of rainfall, the total rainfall would be 20 mm. Using dimensionless average analysis, this increase would be calculated as 100%, with the resulting line reaching 200%. This dramatic increase has a strong visual impact and can bias the interpretation of the data. However, in a wet period with a rainfall depth of 200 mm, the same model showing a 10 mm increase would only result in a dimensionless average increase of 5%, i.e., a minimal change. In both hypothetical scenarios, the increase in (absolute) rainfall volume is the same, but the historical (relative) data differs. To avoid misinterpretation, special attention should be paid to the dry season from May through August, where base precipitation values can be very low.

The data presented can be further disaggregated by counties. To minimize the variability in climate models and identify trends in rainfall and its impacts for the next 100 years, the analysis was conducted on distinct scenarios and models. Figure 3(a) and 3(b) displays the results of Eta–BESM RCP 4.5 and RCP 8.5 models, respectively. Figure 4(a) and 4(b) shows the outcomes of Eta–CanESM RCP 4.5 and RCP 8.5 models. Similarly, Figure 5 illustrates the results of Eta–HadGEM RCP 4.5 and RCP 8.5 models, while Figure 6 displays the findings of Eta–Miroc RCP 4.5 and RCP 8.5 models. In these figures, warm colors such as red and yellow indicate a reduction in rainfall, while cold colors, such as blue and green, represent an increase in rainfall.

According to Figure 4, there is a noticeable increase of hydrological stress in the region under analysis. This is clear from the reduction in rainfall during the months of January to April. Additionally, the results indicate an increase in rainfall during June and July, which are months of the dry season, and during October and November, which are months of the wet season. This pattern emphasizes the shift in rainfall regime due to climate change.

A comparison between (a) and (b) of Figures 3–6 illustrates the impact of climate change, considering the differences between both RCP scenarios. Figure 3(b), which depicts the Eta–BESM climate model and RCP 8.5 scenario, demonstrates a more extreme scenario than that shown in Figure 3(a) (RCP 4.5 scenario). The reduction of rainfall during January through April is greater in Figure 3(b) than in the previous scenario, and the increase in rainfall during the wet season, mainly October and November, is smaller than in the RCP 4.5 scenario. These findings indicate that extreme events have a higher chance of occurring in the RCP 8.5 scenario. However, there is no significant change in rainfall during June and July.

When comparing the Eta–BESM and Eta–CanESM models, a slight increase in rainfall during the dry season and a slight decrease in rainfall during the wet season can be observed.

Comparison of the Eta–HadGEM model with the other models reveals that it combines the results of the previous two models (Eta–BESM and Eta–CanESM). Specifically, during the dry season, the Eta–HadGEM model indicates less rainfall compared to the historical data, like the Eta–BESM model. Additionally, during the wet season, the Eta–HadGEM model shows a decrease in rainfall, although not as much as in the Eta–CanESM model.

To enhance visualization and facilitate comparison, Table 2 presents a concise summary of the findings depicted in the maps.

One of the reasons behind this contrast lies in the structural variances of the GCMs, as illustrated in Table 1, which lead to the generation of the distinct outcomes. The GCMs and the RCMs can better represent some climate zones, such as the relationship found between Eta–Miroc and the Intertropical Convergence Zone (Chou et al. 2014a). Furthermore, the seasonality affects the models, and the precipitation can correlate to a region in one season but not the other. An example is Eta–HadGEM, where in the December-January period it is highly correlated with the whole territory, but during the June-August period, it loses its power and is outperformed by other models (Chou et al. 2014a). Another remark is the different RCP scenarios. Reduction in precipitation caused by the more severe RCP 8.5 scenario can only be visualized in some models and seasons, like Eta–Miroc and Eta–HadGEM (Chou et al. 2014b). While previous studies have not delved into this aspect, it is highly probable that Eta–CanESM exhibits analogous behavior to the other models and demonstrates a stronger correlation with a particular region during a specific period.

The Eta–Miroc model shares a similar trend to that of other models. However, the change in rainfall when compared to the historical data is more subtle than the other models, both in the wet and dry seasons.

Three primary inferences regarding the trends in rainfall for the predicted years emerge. Firstly, there will be a general reduction in rainfall, with precipitation concentrated in fewer months during the wet season. This will result in longer and more severe droughts, which will have a significant impact on sectors such as agriculture and power generation. Furthermore, the concentration of rainfall in fewer months may lead to extreme weather events occurring more frequently. Secondly, during the dry season, there will be an increase in rainfall, which could affect the agricultural sector by altering the optimal time to begin planting crops. This change could result in crop losses and have a negative impact on the economy. Finally, the third trend predicts an increase in rainfall during the wet season, accompanied by a reduction in rainfall during the dry season. This trend may lead to frequent storms that could, e.g., damage stormwater drainage systems in cities, cause soil erosion in rural areas, and create dangerous droughts for water supply and energy generation systems.

The calibration and validation results were obtained for both sub-basins of the SMAP model. The model was calibrated using the optimal time series data from 1979 to 1984, while validation was performed using data from 1968 to 1973. The NSE coefficient for the calibration and validation of both sub-basin is represented in Table 3. Chavantes was calibrated using data from the Jurumirim sub-basin, as the available data from the Chavantes sub-basin alone proved insufficient for the SMAP model calibration. The weather stations used in this step were SAG, E5-030, F5-041, E5-027, E5-019, E6-013 and E6-022

The river flow was subjected to the SMAP model using the parameters obtained during the previous calibration step, and the resulting NSE coefficients are presented in Table 4.

Interestingly, the RCP 4.5 yields a suboptimal NSE coefficient, whereas the RCP 8.5 exhibits two scenarios with positive scores. Negative values of the NSE coefficient suggest that the mean values of the observed series serve as a more effective predictor.

Located in the southwest of Brazil, both watersheds exhibit a known transient behavior that mimics a sine curve. This is a result of the watersheds location in a transition zone between the atmospheric convergence zones of the north and south (Chou et al. 2014a, 2014b; Utida et al. 2019; Wong et al. 2021). During the wet season, there are losses from January to March and gains from October to December, with a steep drop in the latter month. In the dry season, losses occur from April to June and gains from July to September. Overall, losses account for approximately 60% of the total, while gains can exceed 200%. As with the rainfall analysis, caution must be taken when analyzing these values to avoid misinterpretation of the river flow data.

The RCP 8.5 scenario presents a less optimistic outlook, with greater losses and less accentuated gains when compared to RCP 4.5. The Eta–Miroc model, on the other hand, shows an interesting trend in the two different scenarios, with a smaller difference between them regarding the other models. This trend can be better visualized when the data are broken down by climate normal, as shown in Figure 8(b)–8(d).

All trend lines point to river flow reduction, with two exceptions. First, the model Eta–HadGEM RCP 4.5, presents a slight increase in river flow. This can be attributed to a bias of this model in the study area that evidences a bigger rainfall response than the other models. Secondly, the Eta–Miroc in RCP 8.5 scenario reinforces the previously discussed about trends on climate normal.

Climate modeling data plays an indispensable role in providing essential insights for water managers, enabling them to effectively address and mitigate the adverse consequences of climate change on societal well-being. Efficient water resource management is achieved by combining predictive models and promoting strong information exchange among basin management units. This approach safeguards these resources for both present and future generations. This is particularly significant within the context of the ALPA Basin in the state of São Paulo, Brazil, which assumes a pivotal role as a fundamental water source across various socio-economic sectors.

The importance of proficient water resource management highlights the need for cohesive collaboration and effective communication among the entities responsible for overseeing basin dynamics. This collaborative synergy ensures a comprehensive approach to the sustainable preservation of water resources. Simultaneously, the development and implementation of predictive models assumes paramount significance. These models provide us with the foresight needed to anticipate shifts in water availability and demand, thereby empowering us to proactively adapt to evolving circumstances.

The exhaustive analysis conducted within the Chavantes and Jurumirim watersheds underscores the urgency for proactive interventions. The projected decline in water flow attributed to a multitude of factors, including growing population, escalating water consumption, and evolving land utilization patterns – notably exemplified by the proliferation of center-pivot irrigation systems – foreshadows a concerning alteration in the hydrological cycles of these basins. The ramifications extend beyond the boundaries of the studied headwater sub-basin, cascading downstream and causing disruptions across interconnected basins.

The hydrological stress in the studied region becomes evident in the marked reduction in rainfall of up to −40% during the traditionally wet months of January through April. Furthermore, the anomalous increase of rainfall of up to 40% during the dry season months of June and July, coupled with the wet season months of October and November, underscore the climate-induced modulation of historical precipitation patterns. According to the calculations, in terms of river flow, the Chavantes River witnessed its lowest historical discharge of 73 m³/s, which would be modified to 20 m³/s, while its highest historical flow of 2,266 m³/s would be modified to 907 m³/s. Similarly, for the Jurumirim sub-basin, the lowest historical river flow of 55 m³/s would be modified to 39 m³/s, and the highest historical flow of 1,552 m³/s would be modified to 1,304 m³/s. These manifestations underscore the urgency of a concerted and informed response. Amidst the predominant downward trends in river flow evidenced by this study, notable exceptions warrant discernment. The Eta–HadGEM RCP 4.5 model presents a marginal increment in river flow, ascribed to its distinctive capture of localized rainfall responses. Moreover, the Eta–Miroc model within the RCP 8.5 scenario reiterates the pivotal influence of climate normal in shaping these discernible trends.

In addition to its impacts on water flow patterns, climate change can also shift the distribution of precipitation, potentially causing difficulties for farming and other land uses within the ALPA Basin. Addressing this scenario requires stakeholders to come up with intelligent and creative adaptation plans. To tackle these issues effectively, it is important to look carefully into forecasts on rainfall and, generally, on local climate trends. With a deeper understanding, stakeholders in charge of water management can adopt and implement strategies to cope with the anticipated challenges.

The authors thank CPTEC/INPE for providing the data and collaboration and the Coordination for the Improvement of Higher Education Personnel (CAPES) of Brazil and Foundations for Supporting Research in the States of Minas Gerais (Fapemig) for supporting the Post-graduate Program. This study also had the support of Portuguese funds through Science and Technology Foundation (FCT), under the projects [grant numbers UIDB/04292/2020, UIDP/04292/2020], granted to MARE, and [grant number LA/P/0069/2020], granted to the Associate Laboratory ARNET.

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

The authors declare there is no conflict.