|Thursday, June 03|
Practical guidance on split sample decision-making in hydrological model calibration and validation based on a large sample study of 473 CAMELS catchments
* Hongren Shen, University of Waterloo, Canada
Bryan A. Tolson, Canada
Juliane Mai, Canada
"Model calibration and validation processes are generally required to assess performance and robustness of hydrological models. Validation specifically functions to test model performance under independent conditions not used in calibration. The available data record in a given catchment is conventionally split into different sub-periods for independent model calibration and validation based on the classic Klemes split sample test (SST) and/or differential split sample test (DSST) framework. Various versions of SST and DSST have been proposed recently in terms of different methodologies for sub-periods selection. Unfortunately, these methods typically require modelers to make subjective decisions during model calibration/validation. Which sub-period do I calibrate the model to? What minimum level of calibration performance is needed? What is the minimum level of validation period performance? This large sample calibration study employing two different hydrological models aims to empirically assess the answers to these age-old subjective model calibration and validation decisions. This study applies two conceptual hydrological models, i.e. the GR4J-CemaNeige (6 parameters) and HMETS (23 parameters) models, and 473 catchments from the CAMELS dataset in the contiguous United States, which consists of catchments with minimal human interference and long-term meteorological forcing and streamflow series. We consider a period of 35 years (1980-2014) in our experiments enabling us to split the record into three unique subperiods: calibration, validation and model testing. Model testing period(s) always follows calibration/validation periods to emulate what actually being applied in practical model development processes. For a given model testing period, numerous continuous sub-periods (CSPs) for calibration are defined by employing sliding-windows with various lengths. A full-period window, e.g. calibrating to all available data and skipping validation, is also evaluated. The remaining years prior to the model testing period are treated as the (historical) validation period. During the period of 1980-2014, 50 different CSPs are proposed for each catchment. The two models are calibrated over each CSP with KGE performance metric using the DDS optimization algorithm. Each CSP calibration is optimized using five independent optimization trials to ensure one of them finds a solution that approximates the global optimum. In total, there are 236,500 (473 × 50 × 5 × 2) model calibrations conducted. The KGE in the calibration CSPs and the corresponding validation and testing periods are analyzed to identify the impacts of different sub-periods on model prediction during the model testing period. Our results provide unique and robust general guidance for modelers deciding how exactly they should split the observation data between calibration and validation. For example, results show that: (1) When data is sufficient, testing period performance is the best when the most-recent, shorter year CSPs are utilized; however, data insufficiency tends to make full-period CSPs results more robust. (2) Calibrating full-period CSPs (i.e., no validation) generally obtains higher model success ratio in model testing compared to almost all other shorter year CSPs. (3) Model calibration failure is a much stronger predictor of model failure during the testing period than model validation failure. These findings have critically important implications regarding how hydrological models are calibrated and validated. "
A novel calibration strategy for a routing-based forecasting model of the North and South Saskatchewan Rivers
* Frezer Seid Awol, University of Waterloo, Canada
James Craig, Canada
Bryan Tolson, Canada
Curtis Hallborg, Canada
"The timely and accurate forecasting of flows on North and South Saskatchewan river systems is vital for water supply, hydropower generation, ferry operation, irrigation, and flood protection works in the Province of Saskatchewan. A routing-based forecasting model was built using the Raven hydrologic modeling framework (Craig et al., 2020) to support the Flow Forecasting and Operation Planning unit at the Water Security Agency (WSA). This project describes the calibration strategy employed in the Raven model to route observed streamflows from headwater gauges in Alberta down to the location of interests in the Province of Saskatchewan. The river-lake network routing structure for the catchment was initially prepared using QGIS and GRASS-based toolbox BasinMaker. Two main in-channel routing parameters of the Raven model were calibrated to control the timing and dissipation of the flood wave simulated with the Diffusive Wave routing technique. Observed flow data from multiple gauging stations were used as a forcing inflow to downstream stations and as the routing input to the model. In this routing-only model, the only source of water used as input to the model was the upstream gauging stations since the alpine headwater areas for these rivers account for the bulk of the observed flows, even very far downstream, during high flow events. Also, because the travel times from these headwater areas to locations of interest in Saskatchewan are several days, the routing of observed flows can provide an accurate and timely forecast. The calibration was performed starting from the most downstream gauge moving successively upwards to the next upstream gauge(s) along the same river. This approach is different from the traditional calibration method which is usually conducted from upstream to downstream. An iterative calibration approach was used 1) to adjust the routing parameters at all available gauges driven with observed flows from their closest upstream gauges, and 2) to refine the parameters with inflows from the next upstream gauges to improve the forecast performance at different lead times ranging from three to eight days. By balancing the performance of the flood hydrographs at multiple gauges using different upstream forcing locations, a successful routing-based flow forecasting model was built that can predict flows several days ahead without the need for simulating runoff across the contributing watershed. The next step for the WSA is to integrate this model into Delft-FEWS for operational deployment."
Calibrating the Raven hydrological model to a dense network of continuously monitored natural lake levels on the Petawawa River Watershed
* Bryan Tolson, University of Waterloo, Canada
Ming Han, Canada
Robert A. Metcalfe, Canada
"The 4100 km2 Petawawa River watershed is unique, having more than 5 years of continuously monitored lake levels for 19 natural lakes across the watershed ranging in area from 0.1 km2 to 26 km2. A spatially distributed hydrological model of the watershed was developed that explicitly simulates the water levels of the almost 400 natural lakes in this system (covering 8% of the watershed). To our knowledge, few studies have developed a daily or sub-daily time step hydrological model fit to such a dense continuous data set of natural lake levels. The models applied here were built in the Raven hydrological modelling framework and all model calibration was conducted using the Ostrich parameter estimation software (using single and multi-objective calibration formulations). In addition to the 19 lake level time series, the model was calibrated to one streamflow station (Water Survey of Canada gauge 02KB001, which defines the watershed outlet). Results show our most basic calibrated model fed with climate data from only a single Environment Canada and Climate Change (ECCC) weather station shows excellent model performance for streamflow (e.g., KGE above 0.9) and good quality model performance for lake levels (e.g., KGE of nearly 0.5 for large lakes). We also evaluated how the model performance changes when we: 1) changed model structure; 2) used more refined climate forcing data; and 3) modified the calibration formulation to fit increasingly more observed system response. For model structure, we simulated rainfall-runoff response in each subbasin using GR4J and HBV-EC rainfall runoff model formulations and these subbasin level runoff fluxes are then routed through a complex lake and river routing structure involving roughly 1000 subbasins. We considered three sets of climate forcing inputs: a) one daily climate station; b) four climate stations combining all available meteorological stations run by multiple agencies; and c) available high resolution (in both time and space) gridded climate data from ECCC. In particular, we show how the model parameter uncertainty is appropriately reduced when we add lake level calibration data to constrain the calibration problem. "
Influence of Wastewater Temperature on Treatment Processes under Climate Change
* Vaibhav Bhate, Carleton University, Canada
Hamidreza Shirkhani, Canada
Banu Ormeci, Canada
Shuang Liang, Canada
Yehuda Kleiner, Canada
Andrew Colombo, Canada
"Wastewater treatment is important for health and preventing the contamination of aquatic systems and the environment. Primary, secondary, and tertiary treatment processes are the typical elements of a wastewater treatment system. Primary treatment involves physical and mechanical processes, which remove inorganic and settleable matter. Secondary treatment, such as activated sludge systems, involves the biological processes occurring in wastewater treatment. Biological nutrient removal is an emerging issue, which includes the assimilation of nitrogen and phosphorous in water bodies. Increased concentrations of nitrogen and phosphorous in receiving environments pose a threat to aquatic flora and fauna, by significantly competing for dissolved oxygen. The biological nutrient removal systems including the Wuhrman, Ludzack-Ettinger, Bardenpho, and modified Bardenpho process, are effective methods of reducing organic carbon and nutrient loads from wastewater influent. Climate change can affect wastewater flow, temperature, and concentrations, as well as their patterns. These changes, in turn, can impact various wastewater treatment processes. This study investigates the impacts of temperature change on primary treatment processes as well as various processes of biological carbon and nutrient removal systems using the BioWin numerical modeling tool. Preliminary results showed that effluent COD and cBOD concentrations were higher at colder wastewater temperatures. In the conventional activated sludge system, increasing SRT up to 11 days resulted in lower cBOD concentration in the effluent for wastewater temperatures above 13oC, while a slight increase in cBOD was observed at colder wastewater temperatures for SRT longer than 9 days. The TSS concentrations were also higher for colder wastewater temperature and decreased with increasing wastewater temperatures. Due to the sensitivity of the nitrifying bacteria to cold temperatures, it was shown that for low wastewater temperatures and low SRT, the nitrification rate was very low or negligible. The outcome of this study could provide insight into the performance of wastewater treatment processes under changing climate conditions and wastewater temperatures. In addition, it could help with the efficient operation of wastewater treatment facilities under the potential impacts of climate change. "
Assessment of adaptation solutions to floods with PCSWMM and a multicriteria analysis for a very small watershed
* Audrey Coulombe, École de technologie supérieure, Canada
Jean-Luc Martel, Canada
Annie Poulin, Canada
Mathias Glaus, Canada
Geneviève Audet, Canada
Steve Girard, Canada
Floods are of major concerns as they constitute a hazard for inhabitants of riparian areas. In Quebec, at least 80 % of riparian municipalities undergo flooding and must find adequate solutions to diminish their impacts on both agricultural and urban sectors. Small watersheds are not spared by this problematic and often they do not have the resources to address it. Therefore, the aim of this study is to develop a methodology to evaluate different adaptation solutions for a municipality combining agricultural and urban areas in an ungauged small watershed subject to flooding. To do so, a detailed hydrologic and a hydraulic model that suits the territory was designed using the PCSWMM software. This model was calibrated and validated using selected events from a one-year monitoring of water level. The model was then used to test the adaptation solutions in a climate change context to better assess their resilience. With all these solutions, a multi-criteria decision analysis (MCDA) with the PROMETHEE method considering socio-economic, environmental and technical criteria was performed. The MCDA approach aims to prioritize the solutions according to the priorities of the onsite decision makers This study allowed to elaborate a robust methodology and target the needed data to conduct a flood study in a small ungauged watershed to prioritize adaptation solutions.