Variations in the streamflow of the Nierji Reservoir Basin and quantification of the influencing factors

In recent decades, there have been significant changes in river flow patterns reported in many basins and regions around the world (Schindler 2001; Xie et al. 2018). For instance, the flooding process of the Niger River has transitioned from a single-peak pattern to double peaks since 1970, resulting in a more pronounced local flood during the rainy season (Descroix et al. 2012). However, the runoff of many rivers has shown a significant downward trend, posing a serious threat to global water security (Wu et al. 2017; Zhang et al. 2018). Understanding the characteristics of runoff changes is of utmost importance for the rational utilization of water resources. China, as the second-largest economy in the world and the world's largest exporter, is also facing a shortage of water. China's per capita available water resources are only a quarter of the world average (Guan et al. 2014). With the implementation of China's sustainable development plan for society, the economy, and the environment, water resource issues will become more severe in the future (Yue et al. 2017). Hydrologists have recently focused on studying the spatial and temporal variability of runoff in various catchments. However, most of these studies have primarily focused on runoff at a basin scale, while publications on changes in the inflow of reservoirs, which is crucial for the rational utilization of water resources, are relatively limited.

Changes in runoff are a response to the comprehensive impact of climate change and human activities. Hydrological processes within a watershed undergo significant changes due to the combined impacts of climate change and human activities (Zhang et al. 2008). Among various meteorological factors, rainfall holds the utmost importance (Zheng et al. 2009). Research has demonstrated that the natural peak runoff is consistent with the peak rainfall. For instance, in addition to changes in rainfall, in certain high-latitude regions, climate warming accelerates snow melting and augments the amount of snowmelt, resulting in earlier winter/spring floods (Wang et al. 2019). Additionally, alterations in water consumption and underlying surface, particularly the intensity and magnitude of land-use and land-cover changes, also play a crucial role in modifying runoff patterns (Scanlon et al. 2007; Yang & Tian 2009). Previous studies have indicated that changes in the underlying surface can account for up to 73% of runoff changes in certain locations (Ni et al. 2022). Thus, it is scientifically imperative to analyze these responses to understand watershed hydrology and enhance water resource and land management practices. Previous studies have extensively examined the extent to which changes in runoff can be attributed to climate variability and human activities. The methodologies employed to analyze the impacts of climate variability and human activities on runoff change encompass statistical methods and hydrological models (Wu et al. 2015). Commonly used statistical methods include the correlation coefficient, regression model, Budyko equation (Zheng et al. 2009), and climate elasticity of streamflow (Nenjiang Nierji Water Resources and Hydropower Co., Ltd. 2014), among others. The correlation coefficient can be used to qualitatively explain the influence of factors on variables. However, it is difficult to quantitatively analyze the influence using this method. A regression model can be established using multiple independent variables and dependent variables. The regression coefficients of each factor can help provide a qualitative assessment of their effects. However, this model does not consider the cross-correlation among the factors and, therefore, cannot quantitatively analyze the contribution of each factor to the runoff. The Budyko-type model is a simple and effective method for analyzing runoff changes and precipitation and potential evaporation with high reliability (Sun et al. 2023). Nevertheless, it is worth noting that when the Budyko model is applied at a long-term scale, the variations of water storage in the basin can be neglected. When the model is applied at the annual and intra-annual scales, ignorance can lead to huge errors (Gan et al. 2021). Precipitation elasticity is a fundamental estimation of the sensitivity of long-term streamflow to long-term rainfall and is particularly useful as an initial assessment of the impact of climate change on land and water resources projects (Khan et al. 2022).

Nierji Reservoir is the largest and most important water control structure in the Nenjiang River Basin, northeast China. It plays a crucial role in guaranteeing food production and wetland protection in the downstream areas, as well as in the entire Songhua River Basin. However, with the background of global climate change and increasing human activities such as deforestation and expansion of cultivated areas, numerous studies have indicated an obvious decreasing trend in the streamflow in the Nenjiang River Basin (Feng et al. 2011; Li et al. 2019). To effectively manage the Nierji Reservoir and mitigate flooding risks in downstream cities, it is essential to gain a comprehensive understanding of the changing characteristics of streamflow from the Nierji Reservoir Basin (NERB) and its response to environmental changes. Accurate information regarding these factors is crucial for optimizing reservoir utilization and reducing flood risks. Therefore, the objectives of this study are to (1) detect changing trend and abrupt changes in the streamflow of the NERB, (2) develop a streamflow simulation model in NERB based on the Back Propagation–Artificial Neural Network (BP-ANN) method, and (3) identify the factors influencing streamflow variations and further quantify the relative contribution of each factor to the streamflow change. The results of this research will be helpful to improve the management and exploitation of the NERB and the whole Nenjiang River Basin.

The inflow flow data of Nierji Reservoir were collected for 1925 to 2013 (Nenjiang Nierji Water Resources & Hydropower Co., Ltd. 2014). Climate change is a significant concern in the field of water resources. Research has shown that for every 1% increase in ground temperature, global runoff increased by 4% over the past century (Wu et al. 2015). Additionally, it has been established that precipitation plays an important role in runoff generation. As a result, daily precipitation and temperature data were collected from four meteorological stations located in the upper reaches of the NERB, including the Nenjiang station and the Shihuiyao station located on the mainstream of Nenjiang River, as well as the Kehou station and the Liujiatun station located on the two tributaries of Nenjiang River. Table 1 provides basic information about each station.

Human activity data, including GDP, forest area, and cultivated area, were collected from the Resources and Environment Science and Data Center (http://www.resdc.cn/) for 1988 to 2013.

To determine statistically significant trends in the hydrometeorological data, the non-parametric Mann–Kendall (M–K) trend analysis test was adopted in this study. The M–K method is effective in distinguishing whether a natural process is a natural fluctuation or a certain trend. Therefore, it has been widely used in trend-detecting in hydrological and meteorological time series (Joshi & Makhasana 2020; Wang et al. 2022).

The statistic series UF is the standard normal distribution, which is in the order of time series X.

The process is then repeated with the reverse order of series X, denoted as {x_n, x_n−1, …, x₁}, resulting in a new sequence UB = {UB₁, UB₂, …, UB_n}, which also follows a standard normal distribution. By plotting the time relation diagram of UF and UB, the point of intersection within the interval [−U_α, U_α], where α is the significance level, indicates the time of abrupt change.

The concentration coefficient (C_d) is used to measure the concentration level of the inflow within a year. The range of C_d is between 0 and 1, and the closer C_d is to 1, the higher the concentration level the runoff has.

To calculate C_d, the discharge of the runoff for each month is represented as a vector length, with values ranging from 0° to 330° in 30-degree intervals. Monthly runoff is decomposed into two vectors, R_x and R_y, in the X and Y directions.

In recent years, the artificial neural network has emerged as a valuable method for modeling nonlinear phenomena (Farsi & Mahjouri 2019). In this study, we employed a BP-ANN model to analyze the complex relationships among runoff, meteorological data, and human activity data in the NERB. The multi-layer BP-ANN model is a layered parallel processing system that consists of input, output, and hidden layers. It could have multiple processing layers and nodes in each layer, which are interconnected by links (Agarwal & Singh 2004). These connections between the nodes are represented by weighted values, which determine whether to pass or block the signal. BP-ANN is a kind of artificial neural network characterized by error backpropagation, where the error signal is distributed to the neurons of each layer for weight correction (Farajzadeh et al. 2014). Through continuous iterative correction training, the error is reduced to a reasonable range. In this model, the steepest gradient descent method is used as the learning rule to minimize the sum of square errors.

To evaluate the impact of selected climate and human activity factors on the inflow of Nierji Reservoir, a quantitative identification method proposed by Liu et al. (2014) was employed. The method in this study is based on the fixing–changing theory, but it has two inherent defects: first, the quantification of each factor is not unique; and second, the sum of the contributions of each factor is not equal to the total.

The improved method can address the mentioned problems by following these basic steps:

• (1)

  Assess the state of one factor and modify the state of other factors to create different scenarios. These scenarios can be categorized into two types based on the two states.

• (2)

  Execute the model and calculate the difference in simulation results of each factor between the two types of scenarios.

• (3)

  Determine the relative impacts of various factors on streamflow by calculating the proportion of differences from the simulation results (Chen et al. 2019).

To achieve this, we created an alternative state for each factor by incrementing each group of variables. Assuming there are m factors, we can generate 2^m input combinations for the trained BP-ANN model, resulting in 2^m simulation results. By calculating the changes in simulation results compared with the original results, we can determine the relative contribution rate of each factor. The specific steps involved are as follows:

• (1)
  For each series, X_k= {x_k,1, x_k,2…, x_k,i…, x_k,n}, k= {1, 2, …, m}, where n represents the series length, and m represents the total number of factors, the formula for the increment ΔX_k is

  

  (18)

The new variable series is X_k+ ΔX_k, where β_k represents the M–K inclination of the X_k series.

• (2)

  The current status of each influencing factor, denoted as X_k series, is recorded as state 0. State 1 is recorded as X_k+ ΔX_k, resulting in the generation of 2^m scenarios. These scenarios can be divided into two categories for each factor. In the first category, the state of the factor is 0, and there is a total of 2^m−1 scenarios in this category, which are respectively recorded as {S₁, S₂, …, S₂^m−1}. In the second category, the status of the factor is 1, and there are also 2^m−1 scenarios, which are recorded as {⁠,…,}.

• (3)
  The above scenarios’ data series are standardized and input into the trained BP-ANN model, resulting in 2^m simulation results. For each factor, the 2^m−1 results of scenarios in the first category are recorded as ,= {⁠,…,⁠,…,}, where j = {1, 2,…,2^m−1}. Similarly, the 2^m−1 results of the scenarios in the second category are recorded as={⁠,…,_,…,}, where j = {2^m−1 + 1, 2^m−1 + 2, …, 2^m}. It should be noted that both and have a sequence length of n, and the contribution of each factor to runoff is determined by:

  

  (19)

  where ΔR_k = {Δr_k,1, Δr_k,2,…,Δr_k,i,…,Δr_k,n}.

The mentioned steps above allow for the quantification of the relative contribution of each factor to changes in runoff. By calculating the influence rates of meteorological factors and human activities separately, it becomes possible to determine the relative impact on inflow in a quantitative manner. The improved method utilizes the average of all potential contributions of each individual factor as its impact rate on runoff.

As shown in Table 2, referring to different decades, the highest average streamflow occurred in the 1950s, while the lowest was observed between 1898 and 1909. Non-uniformity of the inter-annual streamflow variation was high in the NERB. The C_v value varied from 0.799 to 0.973, with the smallest and largest values occurring in the 1970s and 1980s, respectively. The value of C_d in the NERB ranges from 0.539 to 0.656, indicating that the annual distribution of runoff is not highly concentrated. The maximum values of both S_r and S_a were recorded in the 1980s. During this period, the maximum monthly streamflow was 320 times the minimum, and the value of S_a was 1,280.7 m³/s, which highlights the largest extreme runoff gap and significant changes in streamflow. The minimum value of S_r was observed in the 1940s, reaching 50. The smallest absolute change occurred from 1898 to 1909, with a difference of 471.8 m³/s, indicating relatively minor annual variation in internal flow during this period.

Table 2

Statistical indicators of streamflow variations in the NERB for different decades

Period .	(m3/s) .	Cv .	Cd .	Sr .	Sa (m3/s) .
1898–1909	2,171.4	0.812	0.564	137	471.8
1910–1919	3,091.7	0.832	0.591	114	598.5
1920–1929	2,555.4	0.897	0.606	75	611.4
1930–1939	4,734.7	0.927	0.646	71	1,041.2
1940–1949	3,813.5	0.822	0.569	50	714.0
1950–1959	4,858.5	0.923	0.637	163	1,105.4
1960–1969	4,074.1	0.876	0.618	122	789.0
1970–1979	2,644.3	0.799	0.576	157	481.1
1980–1989	4,744.4	0.973	0.656	320	1,280.7
1990–1999	4,723.2	0.937	0.656	167	1,138.3
2000–2013	2,926.0	0.927	0.539	69	668.7

Period .	(m3/s) .	Cv .	Cd .	Sr .	Sa (m3/s) .
1898–1909	2,171.4	0.812	0.564	137	471.8
1910–1919	3,091.7	0.832	0.591	114	598.5
1920–1929	2,555.4	0.897	0.606	75	611.4
1930–1939	4,734.7	0.927	0.646	71	1,041.2
1940–1949	3,813.5	0.822	0.569	50	714.0
1950–1959	4,858.5	0.923	0.637	163	1,105.4
1960–1969	4,074.1	0.876	0.618	122	789.0
1970–1979	2,644.3	0.799	0.576	157	481.1
1980–1989	4,744.4	0.973	0.656	320	1,280.7
1990–1999	4,723.2	0.937	0.656	167	1,138.3
2000–2013	2,926.0	0.927	0.539	69	668.7

Note: C_v is the coefficient of variation in streamflow, and C_d is the concentration coefficient of streamflow. S_a and S_r represent the absolute and relative change of annual streamflow, respectively.

View Large

Global warming has been reported in various countries (Shi et al. 2018; El Kenawy et al. 2019). In this study, a declining trend of temperature was witnessed from 1988 to 2013 (Figure 6(c)). However, the trend of temperature changes varied at different timescales. For instance, if the data length was set from 1951 to 2013, the annual average temperature in the NERB did increase significantly (z = 5.71). Additionally, changes in precipitation serve as crucial indicators of climate change (Hynčica & Huth 2019); precipitation was observed to decrease both from 1951 to 2013 and from 1988 to 2013 in the NERB. These findings align with the overall trends observed in the northeast region of China (Liang et al. 2011).

The linear correlation between precipitation and streamflow was found to be strongest, followed by forest area. This can be explained by the fact that precipitation in the upper basin of Nierji Reservoir directly contributes to the surface runoff in the same basin through the process of streamflow generation and concentration (Wei et al. 2018). A larger forest area leads to increased soil permeability (Lopes et al. 2020; Hemr et al. 2023), resulting in a lower surface runoff generation rate (Abou Rafee et al. 2021; Yi et al. 2023) and less inflow runoff. On the other hand, there was a weak linear correlation observed between other factors (i.e., temperature, GDP, and cultivated area) and streamflow. This suggests that quantifying the impact of human activities on streamflow change (Ling et al. 2014; Jiang et al. 2015) is relatively challenging in the NERB via a linear function alone. Therefore, it is particularly important to consider a model (e.g., BP-ANN model) that can accurately simulate the nonlinear relations (Longyang 2019) to further study and explore the complex effects of each factor on streamflow.

In this study, the BP-ANN model was used with five input factors, including precipitation, temperature, GDP, cultivated area, and forest area, while streamflow was the output. The model was built based on the data from 1988 to 2013 due to human activities. This period was further divided into a calibration period (1988–2008) and a validation period (2009–2013). The performance of the model was evaluated using the Nash–Sutcliffe efficiency coefficient (NSE) and R². After multiple simulations and comparisons, it was found that the model with eight nodes in the hidden layer performed the best.

Two series were created to analyze the changing trend of each factor, including the original series X_i (marked as status ‘0’) and the incremental series X_i+ ΔX_i (marked as status ‘1’). These series were established for each factor, and then the statuses of all factors were combined to create changing scenarios for the Nierji Reservoir. In total, 32 different scenarios were generated. The details of these scenarios are provided in Table 3.

The trained BP-ANN model is used to simulate the runoff under all scenarios. Based on the method introduced in section 3.5, the relative contribution of each factor was then calculated. The quantitative evaluation results revealed that the contributions rates of GDP and cultivated land area, reflecting the level of social and economic development, to the change of water inflow from Nierji Reservoir, were 32% and 18%, respectively. Forest is a major land-use type in the NERB, and over the course of the selected 25 years, forest area accounted for approximately 25% of streamflow change. Precipitation, as a primary source of river runoff in the NERB, was closely correlated with the streamflow, contributing about 23% to the overall changes. While temperature serves as an indicator of climate change, it contributed only 2% to the total variations. Based on the above findings, the changes in the inflow of the Nierji Reservoir are believed to be the response of both climate change and human activities, and human activities had a greater impact on the streamflow than chronic climate change in the NERB.

The results of this study indicate a significant decrease in precipitation, which is consistent with the findings of previous studies (Feng et al. 2011). However, it is important to note that due to limited data availability, the precipitation and temperature data from the Nenjiang station were used as representative of the entire study area in this study. This may introduce some inaccuracies when calculating the results of precipitation and temperature. Nevertheless, the research findings presented in this study still hold important practical implications. In future studies, it would be beneficial to conduct more accurate meteorological analysis if more comprehensive data become available. Additionally, previous studies have highlighted the significant role of precipitation in driving changes in runoff, with global warming also impacting runoff through increased evapotranspiration (Zhai & Tao 2017). However, in the context of this study, the contribution of temperature in the NERB to the inter-annual streamflow changes was found to be relatively low compared with the other factors.

Overall, precipitation, forest area, GDP, temperature, and cultivated area are identified as the five important factors that influence the streamflow of Nierji Reservoir. Interestingly, both GDP and forest area had a greater contribution than precipitation and temperature, indicating that the impact of human activities on annual runoff change is greater than that of climate change from 1988 to 2008. Although temperature did not exhibit a significant influence on the changes in the annual streamflow in the NERB, it is worth noting that temperature plays a significant role in influencing the streamflow of rivers in cold region areas during snowmelt periods (Hathaway et al. 2016; Jepsen et al. 2016; Wu et al. 2020). In addition to precipitation and temperature, there may be other climate factors that influence streamflow. However, these additional factors were not considered in this study. Therefore, future studies should investigate the specific influence of each climate factor and their interaction effects on streamflow at various timescales.

In this study, we successfully utilized the non-parametric M–K statistical method to analyze the temporal trend of streamflow in the NERB. We specifically investigated the variability of the inflow of Nierji Reservoir from 1898 to 2013 and quantitatively assessed the impacts of both climatic variability and human activities on the changes in streamflow during the period of 1988–2013.

We observed a statistically significant (α = 0.05) upward trend in the span of 116 years from 1898 to 2013, accompanied by a sudden increase in 1925 when the streamflow rose sharply. We used the BP-ANN model to simulate the nonlinear function of streamflow and the influencing factors, and the simulations yielded satisfactory results. In order to further understand the reasons behind the streamflow changes of Nierji Reservoir, we established an influencing contribution assessment framework. It was discovered that human activities in the NERB had a greater impact on the streamflow changes between 1988 and 2013 compared with climate change. Concretely, GDP, forest area, and cultivated area accounted for 32%, 25%, and 18%, respectively, of the total inflow changes. Precipitation, as the primary source of streamflow in the NERB, contributed 23%, while temperature contributed only 2%.

The findings of this study will be valuable for the improved management and development of water resources both in the NERB and the entire Nenjiang River Basin. We anticipate that the method employed in this study will be extensively utilized in future research.

This research was supported by the Strategic Priority Research Program of the Chinese Academy of Sciences, China (XDA28020501).

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.