Woodson Lees Ferry WY flow forecast

Overview

Figure 1a. Random forest model forecast of Water Year 2026 naturalized streamflow for the Colorado River at Lees Ferry, AZ. (Source: David Woodson)

Figure 1b. Two measures of the 'variable importance' (added value) of the four predictors in the random forest forecast model of WY 2025 streamflow shown in Figure 1. (Source: David Woodson)

This experimental streamflow forecast procedure ('randomWoods') was developed by David Woodson as a part of his PhD research at the University of Colorado Boulder. He now produces the forecast as a side project. The forecast uses a machine learning (random forest) model trained on Reclamation annual natural flow for the Colorado River at Lees Ferry, AZ for water years 1921 through the most recent year as the predictand, and the following predictors:

• Pacific Decadal Oscillation (PDO) index: July-August-September average preceding the water year being forecast
• Atlantic Multidecadal Oscillation (AMO) index: July-August-September average preceding the water year being forecast
• Community Earth System Model - Large Ensemble (CESM-LE) forecasts of precipitation and minimum temperature: October through September average coinciding with the water year being forecast

For example, for the Water Year 2026 forecast (October 2025 - September 2026 total natural flow), the PDO and AMO predictors are the July 2025 through September 2025 averages, and the CESM-LE precipitation and temperature predictors are the forecasted October 2025 through September 2026 averages. The random forest forecast is a 600-member ensemble.

WY 2026 Forecast

Figure 2a shows the naturalized streamflow forecast for Water Year 2026, in green, compared to the historical streamflows, 1921-2025, on which the model is calibrated (black line). The uncertainty in the forecasted 2025 streamflow (i.e., the distribution of the 600 model ensemble members) is depicted in two ways: The green boxplot shows the extent of the interquartile range (25th-75th percentiles) and the median or most-probable forecast (50th percentile), which is 9.1 maf. The green semi-violin density plot shows that the main bulk of ensemble members are narrowly distributed around a peak at 8.0 maf. The long, skinny tail extending to 20 maf indicates that while a wet year is not precluded by the model, it is seen as a much less probable outcome.

Figure 2b shows the 'variable importance' for each predictor over the 1921-2025 training period; each variable does add value to the forecast, and AMO has the highest importance.

WY 2025 Forecast

Figure 2a. Random forest model forecast of Water Year 2025 naturalized streamflow for the Colorado River at Lees Ferry, AZ. (Source: David Woodson)

Figure 2b. Two measures of the 'variable importance' (added value) of the four predictors in the random forest forecast model of WY 2025 streamflow shown in Figure 1. (Source: David Woodson)

Figure 2a shows the naturalized streamflow forecast for Water Year 2025, in green, compared to the historical streamflows, 1921-2024, on which the model is calibrated (black line). The uncertainty in the forecasted 2025 streamflow (i.e., the distribution of the 600 model ensemble members) is depicted in two ways: The green boxplot shows the extent of the interquartile range (25th-75th percentiles) and the median or most-probable forecast (50th percentile), which is 10.2 maf. The green semi-violin density plot shows that within the main bulk of ensemble members below 12.5 maf, there is a peak around 9.0 maf, and a smaller peak around 11.5 maf. The long tail extending to 20 maf indicates there is some potential for a wet year, though less so than for the FY2024 forecast.

Figure 2b shows the 'variable importance' for each predictor over the 1921-2024 training period; each variable does add value to the forecast, and AMO has the highest importance.

WY 2024 Forecast

Figure 3a. Random forest model forecast of Water Year 2024 naturalized streamflow for the Colorado River at Lees Ferry, AZ. (Source: David Woodson)

Figure 3b. Two measures of the 'variable importance' (added value) of the four predictors in the random forest forecast model shown in Figure 3. (Source: David Woodson)

Figure 3a shows the naturalized streamflow forecast for Water Year 2024, in green, compared to the historical streamflows, 1921-2023, on which the model is calibrated (black line). The uncertainty in the forecasted 2024 streamflow (i.e., the distribution of the 600 model ensemble member) is depicted in two ways: The green boxplot shows the extent of the interquartile range (25th-75th percentiles) and the median or most-probable forecast (50th percentile), which is 10.8 maf. The green semi-violin density plot shows that the forecast has a bimodal distribution with many members clustered around 7 maf, indicating potential for a very dry year, and a smaller second mode around 18 maf, indicating there is some potential for another wet year like WY2023.

Figure 3b shows the 'variable importance' for each predictor over the 1921-2023 training period; each variable does add value to the forecast, and AMO and PDO have the highest importance.