Exploring the Use of Vulnerability-Based Analyses with Classification and Regression Trees in Reclamation's Mid-Term Operational Projections

The Bureau of Reclamation (Reclamation) projects 2 – 5-year Colorado River Basin (CRB) system conditions for stakeholders using their Colorado River Mid – Term Modeling System (CRMMS). As inputs, this model utilizes 30, 5 – year traces of simulated streamflow generated by the Colorado Basin River Forecasting Center using a historic record (1991 – 2020) of precipitation and temperature data. The 30 inflow traces result in 30 output traces of system conditions, which are together used as an ensemble to estimate probabilities of system outcomes (e.g. 30% of the traces showing Lake Powell falling below 3500 feet indicates a “30% chance” of this happening). The limited number of traces and probabilistic analysis in the 2 – 5 year forecasts of system conditions sometimes lack the degree of skill useful to stakeholders. We therefore propose providing stakeholders with a different form of analysis and the usage of a new inflow data set. Vulnerability analyses use factor – mapping algorithms to connect projected system conditions of interest to the characteristics of the model inputs that caused them; much of the prior work in vulnerability analysis has typically focused on a long-term planning context. This research seeks to adapt these methods to Reclamation’s mid-term operational CRB projections. First, we adapt an expanded set of inflow traces that represent a wide range of plausible conditions, derived from climate model outputs and other sources. We then define a critically low Lake Powell pool elevation threshold as our failure condition of interest and classify the CRMMS output traces as ‘failures’ or ‘successes’ based on whether that threshold is crossed in each output trace. We create multiple Classification and Regression Tree (CART) models that discover characteristics of the inflow dataset that are likely to cause failure conditions, defined over different time intervals (e.g., in the next 12 months, or over the entire modeling period). Each CART model identifies the inflow factors/characteristics that are the strongest indicators of the failure conditions for the specific forecasted time interval. Initial results indicate that the cumulative streamflow volumes ahead of and/or in the same year as the failure condition are strong indicators of system failures. Furthermore, annual inflow volumes below approximately 70 – 75% of the historic record (1991 – 2020) annual average are commonly linked to our failure condition. 

Authors:

Zachary Carpenter, Joseph Kasprzyk, Edith Zagona