Efficient dam-altered streamflow prediction and forecasting using VIC within CyberWater framework

Name: Sophea Rom Phy Affiliation: University of South Florida Email: sphy@usf.edu

1. Research Question or Modeling Problem

The main objective is to provide an efficient framework for dam-altered streamflow prediction and forecasting, using VIC and reservoir and flow routing. This study provides one scientific and one practical benefit on top of the scientific application of streamflow prediction and forecasting of any time scales. The two benefits are the estimation of dam impacts on streamflow and high efficiency of forecasting extended to forecasting. Potential modeling problems using CyberWater are reservoir routing, saving multiple state files simultaneously, and options for autocalibration and forecast data retrieval. The importance of these capabilities is centered around the significant improvement of efficiency of modeling exercise for prediction and forecasting.

The pilot study area is the Mekong River, heavily regulated by mainstream and tributary dams (totaling over 100 dams). The study region is not only one of the global large rivers, where dam constructions vary across space and time, amplifying the variability of water resources for agriculture and fisheries, which impact millions of people every year. From the technical standpoint, the space-time variation of dam constructions make streamflow estimation highly unpredictable and model initialization for forecasting challenging. The temporal resolution is daily with the coverage from 1994 to the present, which covers the period of mainstream dam constructions and operations.

2. Desired Workflow Outcome

The outcomes to contribute to science and application are threefold. One is the fundamental analysis of long-term streamflow variability, which partly justifies model performance and forecast skill (and forecast skill). Second is the evaluation of forecast skill of dam-altered streamflow, which ultimately implies the forecasting capability in real time. Third is an efficient forecasting framework that can be employed operationally and is transferable to other developers for other downstream use cases. Within the framework, users can carry out three analyses: dam impacts on streamflow, model performance, and forecast skill evaluation. Final outputs comprise of dam-altered and naturalized streamflow and streamflow forecasts.

Comparing the two sets of streamflow strengthen the understanding of how dams impact streamflow, model simulation, its performance, and justification for forecast skill. With streamflow forecasts, users can finally evaluate the skill and also investigate how the skill can be improved for operational application. This is valuable for operational hydrology services in the region.

To enable the process, users have to provide reservior operations like seasonal target reservoir levels or daily reservoir storage for reservior routing. These types of data have been increasingly made available due to satellite derivations.
Users also have to provide the static and meteorological inputs for the VIC model as well as other inputs for flow routing.
Within the reservoir module, users will have the option to use reservoir and flow routing or flow routing alone. This will aid the estimation of how dams impact streamflow, as the first step analysis .

3. Required Inputs

The required inputs are as follows: - Observed streamflow data (from the Mekong River Commission, MRC) - Meteorological forcing data (precipitation from GPM IMERG, temperature, wind speed from MERRA-2) - Meteorological forecast data (precipitation, temperature, wind speed from NASA GEOS-S2S) - Spatial datasets (Vegetation from SERVIR SEA, Soil from Harmonized World Soil Database or HWSD) - Initial conditions (required for forecasting) - Reservoir operations (static attributes, seasonal target reservoir levels from Reservoirs for Mainland Southeast Asia or MSEA-Res) - Model parameters (soil and routing parameters to be calibrated) - Model executables of VIC and flow and reservoir routing to be incorporated in CyberWater - Other information related to reservoir routing is dam location file (GIS Ascii file) which includes dam ID and reservoir cells.

4. CyberWater Components That Already Exist

• Existing Data Agents include GPMAgent, MERRA2Agent and DEMAgent.
• Existing Model Agents include MainGenerator, AreaWiseParamGenerator, ForcingDataFileGenerator, InitialStateFileGenerator, RunModuleAgent.
• Data-processing includes msmUnitConversion, msmDatasetOperation, Resample, msmComputeOutput, StaticDataSetToFile.
• Visualization components include msmShowChart.

For model setup and daily performance evaluation: - GPMAgent and MERRA2Agent will be used to extract precipitation, temperature, and wind speed. - msmUnitConversion is to convert the unit of temperature from K and C. - msmDatasetOperation is to calculate wind speed from the u and v components. - DEMAgent is to extract DEM for msmComputeOutput. - Resample is upscale the meteorological variables and DEM to modeling resolution (0.05deg). - StaticDataSetToFile is to save DEM for RoutingAgent_g. - MainGenerator is to control components for both the VIC model framework and routing framework. - AreaWiseParamGenerator outlines directories to the parameter files for both VIC and routing. - ForcingDataFileGenerator outlines forcing data for VIC simulation and flux outputs from VIC for routing. - msmShowChart is to display observed and simulated streamflow. - RunModuleAgent incorporates VIC and routing executables and generates final state files, which will be used as initial states for forecasting.

For model forecasting: - InitialStateFileGenerator locates the initial state file of VIC and routing for RunModuleAgent. - Other components will be used in the same manner as in previous model setup for historical simulation.

Overall, the workflow is similar to Figure 59 in the CyberWater manual.

5. Missing CyberWater Components or Development Needs for This Workflow

Components to be enhanced and incoporated in CyberWater include data agent for the NASA GEOS-S2S, how multiple final state files is saved, and an optional automated calibration module. The use of data agent for the GEOS-S2S aids forecast retrieval linked with preprocessing for optimal efficiency similar to historical simulation. The use of state files can be challenging for both VIC and routing, given that there are multiple forecast start dates per year. This requires state files on multiple dates to be saved in one go to avoid a large number of simulations for individual state generation. This is where the RunModuleAgent becomes useful once modifications are made to the executables of VIC and routing. Automated calibration for streamflow is also a crucial add-on to avoid intensive manual calibration.

The reservoir+flow routing is currently available as an executable for RunModuleAgent. However, it is uncertain if its incorporation works as intended. Current reservoir+flow routing version is not available to save the routing states for both flow and reservoir states, which are essential for proper initialization. Therefore, its incorporation in the workflow must be tested first.

6. Proposed End-to-End CyberWater Workflow

The general workflow is as follows:

Step 1: Model setup, performance evaluation, and saving final states

Inputs for VIC (historical forcings and static) → Data preparation → VIC model execution → Runoff & Baseflow Inputs for Routing + Runoff & Baseflow → Data preparation → Routing (flow+reservoirs) execution → Streamflow → Visualization and statistical evaluation with observations (Auto)calibration based on streamflow evaluation by changing soil and routing parameters → repeat the above processes until finishing all iterations or reaching the optimal performance → Save optimal soil and routing parameters Repeat the VIC and routing for once, using the optimal soil and routing parameters → Save Final States of VIC and Routing

• The routing here should have the options to activate/deactivate reservoir routing to assess the difference between natural and dam-altered flows. This can be done during either the first setup run or final run with optimal parameters.
• (Auto)Calibration is embedded in the framework for optimizing streamflow performance by modifying soil parameters and routing parameters.
• Best sets of soil and routing parameters are saved for setting up model for forecasting.
• Final state files for each forecast period are saved for forecasting.

Step 2: Forecasting and skill evaluation

Inputs for VIC (forecast forcings + optimal soil +other static) → Data preparation → VIC model execution using initial states → Runoff & Baseflow forecasts Inputs for Routing (optimal parameters) + Runoff & Baseflow forecasts → Data preparation → Routing (flow+reservoirs) execution using initial states → Streamflow forecasts → Visualization and skill evaluation with observations

• The completed workflow produced ensemble streamflow forecasts in different periods/initialization dates.
• More importantly, the final optimal model can be used operationally by updating forcings with forecast data and initial states.
• The overall workflow incorporating all existing and newly developed components in CyberWater significantly improves the efficiency of streaflow forecasting studies.

7. Risks and Uncertainty

The major uncertainty in this study is the reservoir data and forcings (both historical and forecasts), which are mainly derived from multiple satellites containing structural biases without any ground-based validation and correction. This is due to the lack of ground observations, which would otherwise be used for bias correction. The computation can also be challenging when HPCs are not available, as the workflow is computationally intensive considering the number of forecast runs and autocalibration. These are all built on top of existing VIC and routing models' structural uncertainties. Moreover, the workflow of incorporating reservior routing and saving flux, reservior and flow states remains uncertain until tested, as it requires additional modeling components and inputs. The extraction of multiple state files for efficiency requires some of the CyberWater components to be advanced or newly developed to enable the workflow.

8. Questions for the CyberWater Team

Two questions were raised to a team member: - Clarify whether the existing modules can save multiple state files in one run or this requires future modification. - Is RunModuleAgent the right component to be used for potential reservior+flow routing incorporation, given that the routing is available as an executable?

9. CyberWater Team Feedback

The feedback and answers to both questions are: - Changes to both VIC and routing to enable multiple state extraction for this work should be considered because the current capability is saving one state file per simulation. This means that saving tens of state files requires the same number of simulations rather than in one run. - The RunModuleAgent is appropriate to incorporate the available executable for reservoir+flow routing.