Three Lakes System decision support system (TLS DSS) repository for forecasting water temperature in Shadow Mountain Reservoir using forecasted weather and varying operational pumping regimes.
The code in this repository is covered by the MIT use license. We request that all downstream uses of this work be available to the public when possible. During development of the code in this repository, AI (ChatGPT) was used to help create the architecture for the Shiny App and was also used to help with coding the rollout of the forecast in Python, most code was modified in part or whole after creation by ChatGPT. The streamlit app was aided by Claude Code in the same manner as described here.
Repository contact: B Steele (b dot steele at colostate dot edu)
Water temperature is an indicator of water quality, as it governs much of the biological activity in freshwater systems. Northern Water, the municipal subdistrict that delivers drinking water to approximately 1 million people in northern Colorado and irrigation water for ~600,000 acres of land, has had recurring issues with water clarity in Grand Lake, the deepest natural lake in Colorado. They believe that the clarity issues in Grand Lake are primarily due to algal and diatom growth in Shadow Mountain reservoir which are pushed into Grand when they initiate pumping operations. Clarity in Grand is regulated by Senate Document 80 which dates back to 1937 and the inception of the Colorado Big-Thompson project, however in 2016 stakeholders and operators adopted a system of “goal qualifiers” for Grand. The goal qualifiers are defined through Secchi disc depth measurements (a measure of water clarity), requiring a 3.8-meter Secchi depth average and 2.5-meter Secchi depth daily minimum to be met throughout the July 1 to September 11 regulatory season.
Water in the Three Lakes System (TLS) naturally flows from Grand into Shadow Mountain into Granby, but pumping operations reverse that natural order by introducing hypolimnetic water (cold water) from Granby reservoir into Shadow Mountain. This process reverses natural flow from Shadow Mountain into Grand and finally into the Alva B Adams tunnel to serve the Front Range (Figure 1). Northern suspects there is a biological “sweet spot” for water temperature in Shadow Mountain Reservoir that may reduce algal and diatom growth and therefore mitigate clarity impacts during pumping operations. The optimal temperature for reducing algal growth is to keep the upper 1m of water less than 15°C in the Summer and Fall and to reduce diatom growth is to keep the average temperature of 0-5m (“integrated depth”) greater than 14°C in the Spring and early Summer, which is a bit of a “Goldilocks” problem. Currently, Northern Water uses simulations of a computationally-intensive physical model to estimate clarity in Grand Lake; however these models take days to run and it is not possible to continually run them to create daily estimates, much less forecasts of either water temperature or clarity.
Figure 1. Cartoon schematic of the three lakes system
We have created an auto-regressive neural network to predict water temperature at the two depth horizons (near surface 0-1m and integrated depth 0-5m) that incorporates the parameters of the physical model that Northern Water uses. This model is accurate and performs better than a persistence model (yesterday-is-today) and can make an estimate of temperature at the two depth horizons in seconds. The value of this model (and the decision support system) is the speed at which these estimates and forecasts can be made as well as the accuracy. The usefulness of a decision support system is not just for this estimate of tomorrow’s temperature, but the ability to use forecasted meteorological and pump operations to estimate lake temperature days into the future. This decision support system would allow for Northern Water and their partners to test augmented pumping operations to determine the impacts to water temperature (and therefore clarity, if the Goldilocks temperature hypothesis is true), since we already know that the model is sensitive to large changes in pumping operations. Currently, pumping operations are mostly defined by expert operators, meaning that operators have embedded knowledge of the system. The hope is that between their expert knowledge and this data driven model, we can provide additional context to the decisions these operators are making on daily basis.
Figure 2. Sketch of decision support submodules for the TLS DSS
Figure 3. Sketch of the TLS DSS dialog user interface
This repository is built using {targets} infrastructure. To run the workflow
and update the underlying data, use the command targets::tar_make() in the R
console.
conda create --name env_TLS_DSS python=3.10.10 conda activate env_TLS_DSS conda install -c apple -c conda-forge -c nodefaults tensorflow-deps conda install tensorflow python -m pip install tensorflow-macos==2.10.0 python -m pip install tensorflow-metal==0.6.0 pip install tensorflow-probability==0.15 silence-tensorflow conda install numpy scipy matplotlib scikit-learn jupyterlab pip install pandas statsmodels icecream palettable seaborn progressbar2 tabulate isort pip install tqdm pydot graphviz pip install -U scikit-image pip install shap conda install pandas pip install chardet