|
DMSTA Calibration to Datasets North of Lake Okeechobee Lakes, Reservoirs, & Treatment Wetlands prepared for Wetland Solutions, Inc., October 30, 2003 |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
The Dynamic Model for Stormwater Treatment Areas (DMSTA) has been calibrated and tested against data from approximately 70 data sets derived from experimental platforms, field-scale test facilities, and full-scale treatment wetlands primarily located south of Lake Okeechobee. These data represent a variety of spatial scales, vegetation types, hydraulic regimes, and concentration regimes. This report summarizes describes calibrations to additional data from lakes and treatment wetlands north of the Lake (map). This effort has been undertaken because of the interest in applying the model for designing treatment areas in that region and in applying it to reservoirs with depths that extend over wider ranges, as compared with previous DMSTA calibration platforms. It is also consistent with the current focus on data from full-scale systems, as opposed to experimental mesocosms or test cells, as a basis for calibrating and refining the model. Ten of the datasets have been assembled from information compiled by Wetland Solutions from sources identified in the attached table. These include additional data for the Iron Bridge/OEW treatment area, which was also included in the original DMSTA datasets. The remaining three (Okeechobee, Istokpoga, and Boney Marsh) are derived from existing DMSTA datasets. Dataset features are summarized in the attached index and displayed on the attached summary charts. Detailed calibration results are given below. The index also contains notes on data limitations and platform features that may limit the accuracy of the data or calibration and the utility of the calibration as a basis for design. Depth time series were extrapolated from stage/discharge regressions for some of the lakes. Depth information is generally spotty for the treatment wetlands. Errors in the depth information could affect the calibrations, especially in systems with depths below 60 cm (upper limit of depth-dependence in the DMSTA P cycling model). Some of these additional datasets are substantially different from previous ones with respect to depth regimes, water loading, vegetation types, and/or inflow concentration regimes, all of which may influence the calibrations. While there is no specific reason to expect that model calibrations would vary between these two regions, geographic factors (climate, soil types) may influence the calibrations. The treatment wetland datasets include periods with relatively high inflow concentrations (200-300 for Titusville, 100-600 ppb for Iron Bridge and 5000 ppb for Lakeland), as compared with previous DMSTA datasets that had average inflow P concentrations less than ~150 ppb. In general, the lake/reservoir datasets tend to be relatively deep and dominated by phytoplankton and/or submersed vegetation. While none of the lake/reservoir datasets are from systems that routinely experience dryout (as expected for many CERP reservoirs), they may be representative of periods when CERP storage reservoirs are flooded for extended periods. All of the lake/reservoir datasets have mean hydraulic loads that are either above or below the range expected for treatment area cells (3 - 10 cm/day). This limits the accuracy of the calibrations and/or the utility of the datasets as design prototypes. Two (Saw grass, Hellnblazes) have extremely short hydraulic residence times (< 2 days) and hydraulic loads > 100 cm/day ( + in table below). The sensitivity of the predicted outflow concentration to the P uptake calibration is low in these cases. Additional data collection efforts are needed to develop a design basis for reservoirs that dry out frequently and have hydraulic loading ranges that are more typical of treatment areas and CERP reservoirs. Daily inflow and outflow concentration time series have been generated by interpolating measured values between sampling dates. A regression algorithm thought to provide more accurate estimates for sparse inflow concentration datasets is used for Lake Istokpoga (Walker & Havens, 2003). Platform morphometry (area, width, depth), rainfall, and evapotranspiration rates were compiled by Wetland Solutions. The hydraulic coefficients (minimum depth of flow and stage/discharge parameters) have been adjusted to match the observed depth and/or outflow volume time series. Simulated depths are constrained to measured values in Lake Okeechobee because water levels are regulated rather than discharge-controlled. Seepage is ignored. Atmospheric deposition rates (insensitive) are assigned values used in other DMSTA calibrations. Two Tanks-in-Series (TIS) are assumed for lakes/reservoirs (one for Lake Okeechobee) and six for treatment marshes (typical of measured TIS for emergent marshes in the existing DMSTA platforms). It is desirable to calibrate the model to time periods when the simulation is insensitive to the assumed initial conditions. For a period after the start of each simulation, predicted values for outflow concentration and storage are influenced by the initially assumed value for the biomass P storage, which is not measured. The duration of this influence varies with dataset. Initialization of the storage term is a particular problem in systems with high concentrations and low net P removal rates (extreme case = Lakeland). One approach to removing this influence (taken in many of the previous DMSTA calibrations) is to cycle through the input time series several times. The procedure assumes that P storage levels at the beginning and end of the simulated period are in the same range. This assumption is inappropriate if there is a significant increasing or decreasing trend in the inflow concentration or load time series. Such trends are present in several of the time series. An alternative approach is to run the model for a sufficient time period to flush out the initial conditions before using the results for calibration. The first N years of the input time series is simulated to initialize the storage term and then the entire simulation is restarted. In most cases, sensitivity testing indicates that N=5 is sufficient to flush out the initial conditions. Shorter duration are used in some cases, depending upon the total length of the dataset, presence of concentration or load trends in the time series, and results of sensitivity testing. The calibration period is further adjusted to exclude periods with sparse data and to provide an additional buffer against initial conditions. The initialization procedure assumes that the lake or treatment area was in "stable operation" at the start of the data period (i.e., the storage and concentration regimes reflected the current loading regimes rather than antecedent conditions). Since DMSTA is not designed to simulate treatment area startup periods, the calibrations may be inaccurate for systems that are still in startup mode or with significant trends in loading prior to the dataset period. The absence of trends in model residuals (observed - predicted concentrations) during the model calibration period is one indication that the simulation is not affected by initial conditions. Orlando Easterly Wetland (Iron Bridge) calibrations have been developed for the entire system (Cells 1-17 + Lake) and for Cells 1-15. The latter is more homogeneous and representative of an emergent marsh. The mean and variance of the observed OEW outflow concentration increased dramatically around 1998, despite the fact that inflow concentrations did not change. While the reasons for these increases are unknown, experimental manipulations involving shutdown of the southern flow path were initiated in 1997 (R.Kadlec, pers. comm.). Because of this pattern, OEW calibrations focus on the 1988-1997 period. During some periods, the accuracy of the outflow concentration data is also constrained by a high analytical detection limit (20 ppb). Calibration procedures follow those previously applied to other DMSTA datasets. An optimal estimate of the "settling rate" (K, meters/year) is derived for each platform. The K value is selected to minimize the residual sum of squares of the log-transformed concentration time series. The remaining P cycling parameters are fixed at those used in the existing emergent (Boney Marsh) calibration (C0 = 4 ppb, C1 = 22 ppb), as initially derived from biomass P and water column P data. Calibration results and dataset features are summarized in the attached figures. which include bar charts of dataset parameters and scatter plots of calibrated K against depth, inflow concentration, and hydraulic loading. While detailed interpretations are beyond the scope of this report, a few general observations are made below:
The parameters C0 and C1 parameters determine the amount of stored P at a given water column concentration, as calibrated to biomass and water column P data from DMSTA platforms. The steady-state solution of the P cycling model indicates that predicted outflow concentrations are independent of C1 in a steady-flow system. Because it influences the P storage or "memory" of the system, sensitivity to C1 is greater in datasets with increasing or decreasing trends in inflow concentration or load, such as OEW and Titusville. Most of the previous DMSTA calibration platforms did not have concentration or load trends. Superior fits of the OEW & Titusville data are obtained with a C1 value higher than that originally calibrated to biomass P data (22 ppb). With C1= 22 ppb, the calibrated K values for OEW vary from 8 m/yr in 1989-1991 to 13 m/yr in 1993-1997. Neither calibration gives an acceptable fit of the entire 1989-1997 period. If we fit both parameters to the entire 1989-1997 time series, least-squares parameter estimates are C1 = 66 ± 6 ppb and K = 9.6 ± 0.2 m/yr. With the higher C1 value, a single calibration fits the entire 1989-1997 series. A similar pattern is observed for the OEW Cell 1-15 and Titusville datasets, using the C1 value calibrated to OEW (see table below). Simulations of datasets without concentration trends (including the Boney Marsh prototype) are insensitive to C1. While alternative C1 calibrations are worth exploring, it is possible these patterns may reflect startup phenomena in the treatment wetlands or limitations of the procedures used to initialize the P storage term. C1 values greater than the originally calibrated value of 22 ppb are defensible in the context of the measured P storage in macrophyte systems if we assume that a portion of the measured storage is "inactive", i.e. dead biomass or stable inorganic/organic residuals that would not be involved in P uptake. Higher C1 values (predicting less storage) may be also appropriate for lake/reservoirs that are phytoplankton dominated. Predicted P storage may be too high for treatment areas operating at high inflow concentration (Lakeland, in particular). These limitations would generally influence the temporal dynamics of the simulations, but not the long-term-average predicted outflow concentrations. Improvements in the calibrations would require compilation of biomass P storage data (fractionated), recalibration of K & C1 to datasets with trends and/or strong loading pulses, and possible changes in model structure (e.g., placing a cap on storage or building a P saturation term into the uptake function). Such modifications are beyond the scope of this report, but will be considered in future DMSTA development. Among the lake datasets, Poinsett comes closest to matching the depth regimes of CERP reservoirs (see attached depth time series and depth frequency distributions). The depth range (0-6 feet) spans that observed in DMSTA emergent platforms and CERP reservoirs, although the dryout frequency is lower. It's K value (15 m/yr) is much higher than the calibrations to other other lake & reservoirs (0.5 - 5.1 m/yr), which were generally deeper and had more stable water levels. Phosphorus mobilization following dryout/rewetting is evident in the time series (1980-1982). When the model is calibrated to the load time series (vs. log-concentration), however, the K value decreases to 0.7 m/yr, largely as a consequence of the dryout/rewetting pulse. The loading calibration is also strongly affected by 2 extremely high measured P values (potential outliers). Further study of this system is recommended, including verification of depth data and compilation of vegetation & P storage data. The code for least-squares calibration of DMSTA's reservoir P uptake model has not been developed. Preliminary simulations of each dataset using the reservoir P uptake function with K2 = 0.1 1/yr/ppb are attached (concentrations, loads). One general pattern is that the predictions agree better for the lake/reservoir datasets than for the treatment wetland datasets. This is to be expected, since the model was originally developed based upon data from reservoirs and detention ponds. The treatment wetlands (macrophyte dominated) are generally more efficient at removing P than predicted by the reservoir model. Simulations of the lakes with high hydraulic loads (other than Istopoga & Okeechobee) are relatively insensitive to the reservoir model calibration. With the possible exception of Poinsett, the lake/reservoir datasets are of limited use for calibrating the reservoir P uptake model because of their hydraulic loads (too high or too low) and/or depth regimes.
C1 = 22 ppb unless otherwise noted * calibration accuracy possibly limited by sensitivity to unknown initial conditions for biomass P storage and/or treatment area startup phenomena + calibration sensitivity limited by short hydraulic residence time & high hydraulic load
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
