An Application of Neural Networks To Improve Water Quality Forecasting In The Colorado-Big Thompson Project

The Bureau of Reclamation (Reclamation) manages the Colorado-Big Thompson Project (C-BT), which collects water in the headwaters of the Colorado River on the Western Slope of the Continental Divide and delivers it to the Big Thompson River on the Eastern Slope. Water primarily originates as snowmelt in Rocky Mountain National Park and Grand County, before flowing from Grand Lake through Shadow Mountain Reservoir and into Granby Reservoir, a group of water bodies collectively referred to as the Three Lakes. Water is primarily stored in Granby Reservoir, with additional inputs pumped from nearby Willow Creek and Windy Gap Reservoirs. To meet demands, flows are reversed when the Farr Pump Plant lifts water to the Granby Pump Canal into Shadow Mountain Reservoir and the connected Grand Lake. The Alva B. Adams Tunnel carries water from Grand Lake to the Eastern Slope.

The unique interconnection and characteristics of the Three Lakes creates a complex physical, chemical, and biological system that ultimately controls the water quality in Grand Lake. Further, water quality predictions are needed to understand how operational alternatives and future summer weather conditions will impact the system. Historically, water quality in Grand Lake was monitored by measuring Secchi depth, or simply the depth below the water surface at which a Secchi disk is no longer visible. Secchi depth is a measure of the clarity of the water but is influenced by optical properties, dissolved constituents, and total suspended solids (TSS) including inorganic suspended sediment (ISS), particulate organic matter (POM), and algae.

An improved approach to estimating Secchi depth should consist of multiple operations extending from an initial water quality model – clustering, bias correction, and regression – to account for the stochastic nature of the system and uncertainty within estimation. Each of these operations builds upon the previous step to maximize the predictive skill. While these operations can be implemented separately within traditional regression-based approaches to achieve reasonable results, neural networks (NNs) can simultaneously handle all three operations within a single architecture to produce a more user-friendly product with potentially greater accuracy. This work describes the process for calibrating the initial water quality model and using the predicted water quality values from it as inputs to a NN which was constructed and trained to improve Secchi depth estimates within the C-BT.