The IHACRES_GW model was calibrated to daily streamflow data from gauging station 419032 obtained from the Department of Natural Resources streamflow database (DIPNR, 2004). The period selected for calibration was 1/6/1965 to 30/6/1980, spanning a period of 5 508 days (or 15 years) with a continuous record of daily streamflow data which includes some periods of data-infilling undertaken by the Department of Natural Resources. The river flows during this period of time were considered to be representative of pre-groundwater extraction conditions. Annual groundwater extraction data is available from 1985 onwards, and it is understood that
prior to around 1980 there were relatively small amounts of groundwater extraction taking place within the catchment.
The method of calibration used was one of trial and error where parameters were varied within a sensible range. This proved to be practicable, as well as valuable, because it facilitated parameter value selection by the inspection of visual behaviour of the model together with the use of multiple objective performance criteria (Section 6.4) in a flexible way. It also resolved the problem of how to weight different parts of the fit to the record e.g. where data were known to be in error due to infilling.
The calibration parameters providing the best model fit over the calibration period based on visual inspection to observed streamflows and flow duration curves are given in Table 6-1, along with the associated objective function fits.
Table 6-1 IHACRES_GW calibration period (1/6/1965 to 30/6/1980) parameter values and objective function fits
Parameter Calibrated Value
vs 0.09
τs 15 days
τq 0.9 days
Loss 6 ML/day
Objective Function Value
R2 0.89
R2_Baseflow 0.62
R2inv 0.79
RB 0.28
RB_Baseflow 0.48
The calibrated parameters listed in Table 6-1 resulted in derived values of αq= -0.329 (Equation 5-4), αs= -0.936 (Equation 5-5), vq= 0.91 (Equation 5-6), βq=0.61 (Equation 5-7), βs=0.006 (Equation 5-8) and a= 0.07 (Equation 5-16).
The fits for R2 and RB are good, as expected, because the effective rainfall was calculated using the quick flow signal in the observed streamflow data (refer to Section 5.6). Hence the fit biased towards higher flows is good. The R2_Baseflow and R2inv , which are biased towards the low flows, suggest that the model is also capturing the recession volumes of baseflows rather well on a daily time step. The RB_Baseflow
statistics, however, indicate that the overall volume of baseflow predicted over the 15- year calibration period is about half the volume “measured” in the filtered observed flow. This is in large part due to the fact that low flow events below 8 ML/day were not
reliably recorded at the gauging station, resulting in either no data or data infilling over much of the early record. These low flow events were unreliably recorded due to the bias in the rating curve towards medium and high flows – which is not unexpected for an ephemeral river system- and because of the flat concrete causeway in the base of the river at the gauging station site making low flow readings difficult. The reader is referred to Sauer and Meyer (1992) for further discussion regarding uncertainty in stream discharge measurements. The implications of data-infilling on modelled streamflows are discussed in more detail later in this section.
A confusion matrix for the prediction of baseflows greater than 0.01 ML/day is shown below in Table 6-2. The data from this matrix indicates that the proportion of modelled flows with incorrect recession timings for each of the four possible combinations was 6.6% (4.5 plus 2.1%). The proportion of time where baseflow was “measured” (from the filtered observed streamflow using the minimum filter described in Chapter 5.6) but not modelled was 2.1%. These percentages suggest that the model is performing well in terms of capturing the baseflow switching behaviour within streamflow. By contrast, if no groundwater losses are included in the model run, as per the original IHACRES model formulation, 49% of the total proportion of flows is incorrectly modelled. This is because the exponential formulation does not allow for zero flows to occur. Hence, the IHACRES_GW model captured a substantial amount of baseflow recession behaviour not captured by the original IHACRES model.
Table 6-2 Confusion matrix for calibration period to assess performance of IHACRES_GW model
“Measured” Baseflow <0.01 ML/day Modelled Baseflow <0.01 ML/day
“Measured” Baseflow <0.01 ML/day Modelled Baseflow >0.01 ML/day
69.1% 4.5%
“Measured” Baseflow >0.01 ML/day Modelled Baseflow <0.01 ML/day
“Measured” Baseflow >0.01 ML/day Modelled Baseflow >0.01 ML/day
2.1% 24.3%
As indicated earlier, visual inspection of the observed versus the modelled streamflow data was a key component in the selection of appropriate parameter values over the calibration period. One can see from Figures 6-5 to 6-10 that the modelled streamflow
output shows a good match to the observed streamflow record. Observed flows below 8 ML appear to have been unreliably recorded at the stream gauging station since these flows tend to be reported as zero flow values in the streamflow record. The general underprediction of modelled baseflows is also evident in the statistical measure of
RB_Baseflow, with a value of 0.48. The largest residual differences between observed and modelled flows are associated with very high flow events, which are at times underpredicted by the model due to an overly rapid decay of modelled flows.
0.1 1 10 100 1000 10000 100000 6/1965 6/1966 6/1967 6/1968 6/1969 6/1970 Streamflow (ML/day) Observed Modelled
Figure 6-5 Observed and modelled streamflow at gauging station 419032, Cox’s Creek at Boggabri, for the calibration period (1/6/1965 to 1/6/1970)
-1000 -500 0 500 1000 1500 6/1965 6/1966 6/1967 6/1968 6/1969 6/1970 Re s idua ls ( M L/da y )
Figure 6-6 Residual difference between observed and modelled streamflow for the calibration period (1/6/1965 to 1/6/1970)
0.1 1 10 100 1000 10000 100000 6/1970 6/1971 6/1972 6/1973 6/1974 6/1975 Streamflow (ML/day) Observed Modelled
Figure 6-7 Observed and modelled streamflow at gauging station 419032, Cox’s Creek at Boggabri, for the calibration period (1/6/1970 to 1/6/1975)
-1000 -500 0 500 1000 1500 6/1970 6/1971 6/1972 6/1973 6/1974 6/1975 Residuals (ML/day)
Figure 6-8 Residual difference between observed and modelled streamflow for the calibration period (1/6/1970 to 1/6/1975)
0.1 1 10 100 1000 10000 100000 6/1975 6/1976 6/1977 6/1978 6/1979 6/1980 Streamflow (ML/day) Observed Modelled
Figure 6-9 Observed and modelled streamflow at Gauging Station 419032, Cox’s Creek at Boggabri, for the calibration period (1/6/1975 to 30/6/1980)
-1000 -500 0 500 1000 1500 6/1975 6/1976 6/1977 6/1978 6/1979 6/1980 Residuals (ML/day)
Figure 6-10 Residual difference between observed and modelled streamflow for the calibration period (1/6/1975 to 30/6/1980)
A more detailed portion of the calibration record has been plotted in order to better illustrate the performance of the modelled streamflow recession behaviour (Figure 6-11). From this figure it is evident that the modelled streamflow recessions are at times too rapid, too slow or about right. This suggests that the partitioning of effective rainfall into quick and slow flow components may not be a relationship that remains constant in time (an assumption in the IHACRES_GW linear module and IHACRES generally). For example, different sizes or intensities of rainfall events may result in variable volumes of groundwater recharge and so the partitioning of effective rainfall between the quick and slow pathways would also vary.
0.1 1 10 100 1000 10000 100000 9/1973 9/1974 9/1975 9/1976 St reamf low ( M L/day) Observed Modelled
Figure 6-11 Detailed record of observed and modelled streamflows (22/9/1973- 22/9/1976)
Because of missing data records, some infilling of data was undertaken by the data provider, the Department of Natural Resources. These periods are evident in the record when streamflow data remains constant for a period of time (e.g. up to several weeks or longer), which adds some difficulty in attempting to appropriately fit model parameters. The infilling becomes particularly apparent in the detailed plot of streamflow (Figure 6-12) and the flow duration curve (Figure 6-13) where sections of the observed data look blocky/angular. Because of the data infilling, the modelled flow exceedence percentages are less than those recorded as having been observed. The formulation for generation of effective rainfall (Equation 5-17) requires increasing streamflow volumes
in order for effective rainfall to be generated (e.g. if streamflow values remain constant no effective rainfall is generated). Data infilling has resulted in constant streamflow values over large parts of the record, and consequently the effective rainfall contribution will have been calculated as zero over those periods of the data record. Without effective rainfall input to the model, the modelled streamflow will decline exponentially and result in modelled streamflows that are less than those recorded as having been ‘observed’. Despite data infilling, the IHACRES_GW model applied to the calibration period has performed well in terms of both visual inspection and goodness of fit statistics over the calibration period. In particular, it has captured the switch between baseflow and no flow periods remarkably well.
0.1 1 10 100 1000 10000 100000 6/1970 6/1971 6/1972 St reamf low ( M L/day) Observed Modelled
Figure 6-12 Examples of data infilling in observed streamflow record (1/6/1970- 2/6/1972) at Gauging Station 419032, Cox’s Creek at Boggabri
0.1 1 10 100 1000 10000 100000 0 10 20 30 40 50 60 70 80 90 100
Flow exceedence percentages (%)
Streamflow (ML/day)
Observed Modelled
Figure 6-13 Flow exceedence percentages for observed and modelled streamflow for the 1/6/1965 to 30/6/1980 calibration period