Seasonal forecasts for reservoir systems operation with an over-year carryover capacity - what is their value?

Abstract

Recent advances in weather and climate forecasts have brought fresh energies to studies on operation of reservoir systems with the prospect of improving their performance by assisting decisions through forecasts of water availability. Most of the studies available in the literature focus on stylized schemes with one, single- purpose, reservoir with constant demand and analyse parametrically the effect of forecasts using more or less simplified operation models. While these studies play a fundamental role in creating a general framework for analysis, there is the feeling that the diversity of the boundary conditions (climatic contexts and purposes) to the problem of reservoir management calls for more detailed analyses on specific types of reservoir systems. In this paper, we concentrate on systems with over-year behaviour that are widespread in arid and semiarid areas of the world. The question is if and in what sense streamflow forecasts can help improve reservoir operation in this type of systems, the ones that probably would most benefit from them, given their exposure to droughts. To answer this question, we set up a real time management model of a multi-reservoir, multi-purpose system based on the rolling horizon technique (RHT). The model uses current reservoir levels and inflow forecasts over a forecasting horizon of 24 months to schedule allocations to the system's demand centres with the objective of minimizing the sum of standardized squared monthly municipal deficits and of standardized squared yearly irrigation deficits. Of this schedule, only decisions at the first time step are implemented because actual inflows become available, so that the new actual reservoir levels are available and the decision-making process can continue for the next month based on updated forecasts. We simulate the RHT over a period of 480 months in a system of two reservoirs with side irrigation demands and municipal demands in parallel. We use three types of forecasts: the two extremes of forecasting quality (average inflows and real inflows) and a simple data-driven univariate technique, denoted as Quantile Generation (QG). QG generates future inflows to the end of the current water year by equalling the quantile of cumulated inflows from the current month to the end of the water year to the quantile of the observed cumulated inflows from the beginning of the water year to the current month. We simulate the system with different demand levels, thus obtaining scenarios with different drift indexes covering a wide range of over-year storage behaviours. We show that, perhaps not surprisingly, the RHT using forecasts of the best possible quality (real inflows) provides very similar performances, from the standpoint of total standardized squared deficits, to those obtained using worst quality forecasts (average inflows) and that the RHT with QG performs even worse than with average inflows. This would confirm that, especially in systems where the value of water is levelled out across the various demand centres, as is the case when allocation criterion is the minimization of squared standardized departures from targets, the largest value of forecasts does not reside in their ability to improve system's performances. We then explore another dimension of forecasts, by concentrating on the different ability of the RHT to provide suitable predictions of annual deficits with one-year lead when associated to the three forecast types. As expected, we find that RHT in association with real inflows is able to provide the best forecasts of actual deficits and that there is a value of the drift index above which the QG outperforms average inflows in providing forecasts of annual deficits. We conclude that the real value of forecasts in the management of this type of systems relies in their ability to provide information on future water restrictions with an advance suitable to enforce mitigation measures. If no mitigation measure is possible, then there is little scope for using forecasts to manage this type of systems. On the contrary, mitigation measures based on a timely recognition of future possible use restrictions are likely to benefit from accurate inflow forecasts. From this standpoint the QG proves a simple, promising technique to be further assessed.