Studies have been conducted consistently to increase the prediction and accuracy of the SWAT model and improve the uncertainty of simulation results. Common techniques for calibration of parameters are MCMC (Vrugt et al. 2008), GLUE (Generalized Likelihood Uncertainty Estimation, Beven & Binley 1992), PSO (Particle swarm optimization, Abbaspour 2011), ParaSol (Parameter Solution, van Griensven and Meixner 2007), and SUFI-2 (Sequential Uncertainty Fitting, Abbaspour et al. 2004), and they can be easily applied in conjunction with the SWAT model using SWAT-CUP. SWAT-CUP Premium is a program that calibrates, validates, and improves SWAT analysis results through algorithms, such as SPE (Swat Parameter Estimator), which can be used to perform validation and sensitivity analysis (Fig. 7).
In this study, SWAT-CUP Premium was used to perform calibration and validation, and the number of 1,000 simulations was performed during the calibration in consideration of the efficiency. In addition, the simulated flow was optimized with the SPE algorithm of SWAT-CUP Premium for the observed flow; the optimized parameter results are shown in Table 3.
Table 3
Maximum and minimum ranges and optimal parameter values for SWAT input parameters
Input | Parameters | Initial range | Fitted value |
Gw | ALPHA_BF | −0.048 to 0.952 | 0.316 |
Hru | CANMX | 0–100 | 58.950 |
Rte | CH_K2 | −0.01 to 500 | 294.746 |
Sub | CH_N1 | −0.004 to 29.986 | 10.058 |
Rte | CH_N2 | −0.024 to 0.286 | 0.065 |
Mgt | CN2 | −0.25 to 0.25 | −0.115 |
Hru | EPCO | −1 to 0 | −0.350 |
Hru | ESCO | −0.95 to 0.05 | 0.008 |
Gw | GW_DELAY | 0–500 | 382.250 |
Gw | GW_REVAP | 0.02–0.2 | 0.074 |
Gw | GWQMN | 0–5000 | 1327.500 |
Gw | RCHRG_DP | −0.05 to 0.95 | 0.844 |
Gw | REVAPMN | 0–500 | 209.250 |
Bsn | SFTMP | −6 to 0 | −1.377 |
Hru | SLSUBBSN | −0.25 to 0.25 | −0.043 |
Bsn | SMFMN | −4.5 to 5.5 | −4.355 |
Bsn | SMFMX | −4.5 to 5.5 | 0.445 |
Bsn | SMTMP | −0.25 to 0.25 | 0.122 |
Sol | SOL_AWC | −0.25 to 0.25 | −0.004 |
Sol | SOL_K | −0.25 to 0.25 | 0.248 |
Sol | SOL_Z | −0.25 to 0.25 | −0.227 |
Bsn | SURLAG | −3.95 to 20 | 3.175 |
Bsn | TIMP | −1 to 0 | −0.650 |
R2 (Coefficient of determination), NS (Nash-Sutcliffe Efficiency), MNS (Modified NS), and PBIAS (Percent Bias) were used to evaluate the simulation results with calibrated parameters. Where MNS, R2, and NS were well matched as they were closer to 1. The closer PBIAS was to zero, the better the observed and simulated values were matched (Gupta et al. 1999). Eq. 1 shows the calculation of PBIAS, and in the case of NS values, if it is 0.50 or more, it can be determined that the improvement effect on uncertainty is high, and can be expressed as shown in Eq. 2 (Nash and Sutcliffe 1970).
$$PBIAS =\frac{[{\sum }_{i=1}^{n}\left({O}_{i}-{S}_{i}\right)*100]}{[\sqrt{{\sum }_{i=1}^{n}\left({O}_{i}\right)]}}$$
1
Where: \({O}_{i}\) is observed values
\({S}_{i}\) is simulated values
$$NS=1- \frac{\sum {({O}_{o}-{O}_{s})}^{2}}{\sum {({O}_{o}-{\stackrel{-}{O}}_{o})}^{2}}$$
2
Where: \({O}_{o}\) is observed values
\({O}_{s}\) is simulated values
\({\stackrel{-}{O}}_{o}\) is average observed values
In SPE algorithm, uncertainty of the simulation results is quantified by 95 PPU (95% prediction uncertainty band). The 95 PPU is calculated with parameters, including a measured 95% prediction interval, based on this, P- and R-factors are measured. The P-factor is an indicator of the percentage at which the observed values are included in 95 PPU, and when the P-factor is 1, it means that the observed value is 100% included in the 95% prediction interval. On the contrary, the R-factor represents an average band width of 95 PPU, the closer the value is to zero, the better the calibration results coincide with the observed values. In general, for the simulation of runoff, if the P-factor is between 0.70 and 0.75, it is considered suitable.
Figures 8–10 show graphs comparing the simulated results and observed flow using optimal parameters obtained by the SPE algorithm. The graphs showed that the tendency of daily and seasonal variability was well reproduced, and the flow level corresponding to each day was similar to others. Table 4 shows a comparison of R2 values before and after calibration and validation. The R2 value for SWAT analysis before calibration and validation is 0.40, but after calibration and validation using SWAT-CUP Premium, the simulation result was improved to 0.71. Table 5 shows the model performance evaluation, with NS 0.51, PBIAS 37.1, P-factor 0.73, and R-factor 0.33, indicating that the relationship between the observed and simulated flows has improved, which is considered to be well matched.
Figure 10 Comparison of simulated and observed flow for the entire study period (2013–2016) (after calibration and validation) a Comparison of observed and simulated flows b Comparison of R2 comparison of observed and simulated flows
Table 4
Comparison of R2 values by year before and after calibration and validation
Year | 2013 | 2014 | 2015 | 2016 | 2013–2016 |
Before calibration and validation (SWAT) | 0.21 | 0.52 | 0.17 | 0.49 | 0.40 |
After calibration and validation (SWAT-CUP Premium) | 0.55 | 0.80 | 0.56 | 0.74 | 0.71 |
Table 5
Simulation results using SWAT-CUP Premium after calibration and validation
Evaluation factor | R2 | NS | PBIAS | MNS | P-factor | R-factor |
Simulation results | 0.71 | 0.51 | 37.1 | 0.52 | 0.73 | 0.33 |