Comparing machine learning methods for dynamically identifying future water supply vulnerabilities under climate change
Abstract
Adaptation to climate change will depend on the ability to identify signals in observed hydroclimatic data that reliably indicate future vulnerabilities. This study compares a variety of machine learning methods for this purpose, testing which of multiple indicator variables can predict changes in long-term average water supply reliability (regression) and binary vulnerability outcomes (classification) over lead times ranging from 0-20 years. Methods are tested using a multi-reservoir simulation model, Operation of Reservoirs in California (ORCA), to simulate the reliability of the northern California reservoir system under an ensemble of downscaled CMIP5 inflow scenarios. The indicator variables and reliability metrics are calculated dynamically over multiple rolling windows to identify trends in observations. Regression methods are compared using R-squared, while classification methods are compared using true positive and true negative rates. Both performance measures are compared using leave-one-out scenario validation to investigate the effect of overfitting. Results indicate that several machine learning methods can reliably predict long-term water supply reliability and the occurrence of vulnerabilities, even when substantially reducing the number of indicator variables. This suggests a promising tradeoff between the number of indicator variables and predictive skill. In general, shorter lead times are predicted more accurately than longer lead times, representing an additional tradeoff between lead time and error. The multivariate combination of indicator values that precede drops in system reliability contains signals that may indicate future vulnerability in this system based on observations. These signals can be used to improve monitoring of long-term water supply vulnerabilities, allowing planners to better design adaptation pathways under a changing climate.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.H23G..04R
- Keywords:
-
- 1807 Climate impacts;
- HYDROLOGY;
- 1880 Water management;
- HYDROLOGY;
- 1918 Decision analysis;
- INFORMATICS;
- 6309 Decision making under uncertainty;
- POLICY SCIENCES & PUBLIC ISSUES