Advances in monitoring and forecasting water availability at various time and spatial scales offer a cost-effective opportunity to enhance water system flexibility and resilience. By enriching the basic information system traditionally used to design reservoir operating policies (i.e., time index and reservoir storage) with additional inputs regarding future water availability, operators can better anticipate and prepare for the onset of extreme hydrologic conditions (wet or dry years). Numerous candidate hydro-meteorological variables and forecasts may potentially be included in the operation design, however, and the best input set for a given problem is not always evident a priori. Additionally, for multi-purpose systems, the most appropriate information set and policy shape likely changes according to the objective tradeoff. In this work, we contribute a novel Machine Learning approach to link an Input Variable Selection routine with a multi-objective Direct Policy Search framework in order to retrieve the best policy input set online (i.e., while learning the policy) and Pareto-dynamically. The selected policy search routine is the Neuro-Evolutionary Multi-Objective Direct Policy Search (NEMODPS) which generates flexible policy shapes adaptive to changes in the policy input set. This approach is demonstrated for the lower Omo River basin, in southern Ethiopia, where regulation of the recently constructed Gibe III megadam must strike a balance between hydroelectricity generation, large scale irrigation, and ecosystem services downstream. We develop a dataset of candidate policy inputs comprising streamflow and precipitation forecasts at multiple spatial and temporal scales, from days to months ahead. Long term (season-ahead) forecasts are conditioned on well-recognized climatic oscillations in the region. Specifically, Artificial Intelligence tools are used to detect relevant anomalies in gridded global climatic datasets of sea-surface temperature, sea-level pressure and geopotential height, which are used as predictors for a multi-variate non-linear forecast model. Moreover, we analyze how varying objectives – and tradeoffs therein – benefit from different information. Results suggest that informing water system operations with appropriate information can reduce conflicts between water uses, especially in extreme years when a basic policy is particularly inefficient.
Dynamic retrieval of informative inputs for multi-sector reservoir policy design with diverse spatio-temporal objective scales
Zaniolo, Marta;Giuliani, Matteo;Castelletti, Andrea
2020-01-01
Abstract
Advances in monitoring and forecasting water availability at various time and spatial scales offer a cost-effective opportunity to enhance water system flexibility and resilience. By enriching the basic information system traditionally used to design reservoir operating policies (i.e., time index and reservoir storage) with additional inputs regarding future water availability, operators can better anticipate and prepare for the onset of extreme hydrologic conditions (wet or dry years). Numerous candidate hydro-meteorological variables and forecasts may potentially be included in the operation design, however, and the best input set for a given problem is not always evident a priori. Additionally, for multi-purpose systems, the most appropriate information set and policy shape likely changes according to the objective tradeoff. In this work, we contribute a novel Machine Learning approach to link an Input Variable Selection routine with a multi-objective Direct Policy Search framework in order to retrieve the best policy input set online (i.e., while learning the policy) and Pareto-dynamically. The selected policy search routine is the Neuro-Evolutionary Multi-Objective Direct Policy Search (NEMODPS) which generates flexible policy shapes adaptive to changes in the policy input set. This approach is demonstrated for the lower Omo River basin, in southern Ethiopia, where regulation of the recently constructed Gibe III megadam must strike a balance between hydroelectricity generation, large scale irrigation, and ecosystem services downstream. We develop a dataset of candidate policy inputs comprising streamflow and precipitation forecasts at multiple spatial and temporal scales, from days to months ahead. Long term (season-ahead) forecasts are conditioned on well-recognized climatic oscillations in the region. Specifically, Artificial Intelligence tools are used to detect relevant anomalies in gridded global climatic datasets of sea-surface temperature, sea-level pressure and geopotential height, which are used as predictors for a multi-variate non-linear forecast model. Moreover, we analyze how varying objectives – and tradeoffs therein – benefit from different information. Results suggest that informing water system operations with appropriate information can reduce conflicts between water uses, especially in extreme years when a basic policy is particularly inefficient.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.