Adaptive valuation of an annual cycle budget for rolling day-ahead battery energy storage arbitrage

Details

Supervisors

Legat, Benoît ; Glineur, François ; Bianchin, Gianluca

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en science des données, orientation technologies de l'information, à finalité spécialisée: analyse de données

Abstract

enBattery energy storage systems are increasingly used to capture short-term price differences in electricity markets. In practice, however, profitable operation is not only a question of charging at low prices and discharging at high prices. For an industrial battery, each dispatch decision also consumes part of a limited cycling capability. This thesis studies this trade-off in a rolling day-ahead battery dispatch problem developed in an industrial context. The asset is operated under an annual equivalent-full-cycle (EFC) budget, so using the battery today may reduce the flexibility available for arbitrage opportunities later in the year. The question is whether longer-horizon price scenarios can help value this remaining annual cycle budget more adaptively than a fixed throughput penalty. The industrial benchmark is a rolling 48-hour dispatch model with a fixed throughput penalty lambda, expressed in EUR/MWh of throughput. This fixed rule is simple and transparent, and when well calibrated, it is hard to beat. The thesis compares it with adaptive policies that estimate a continuation value for the remaining EFC budget using longer-horizon price scenarios that are updated monthly. Two ways of translating this value into the operational model are tested: a marginal-value penalty, computed from finite differences or from shadow prices of the linear programming continuation model, and a value-target rule, which searches for the penalty that reproduces a continuation-implied daily cycle target. DiffOpt sensitivities and a stylized multistage example are used as diagnostic extensions rather than as main empirical competitors. Evaluation is based on a chronological replay: on each decision day, deployable policies may use only information available at that time, while realized future prices are kept for settlement and non-deployable diagnostics. The results confirm that the annual cycle budget has an intertemporal value, but also show that this value is difficult to exploit with the available scenario information. Perfect-foresight diagnostics indicate that the marginal value of an additional EFC is not constant over the year. However, the scenario-based adaptive policies do not outperform the ex-post same-year tuned fixed benchmark in the two main evaluation years. In 2024, the best adaptive marginal strategy is almost indistinguishable from fixed. In 2025, the best adaptive strategy remains below the same benchmark by less than one percent. In both years, the marginal-penalty variants perform better than the value-target rules. The historical fixed benchmarks add an important practical perspective. A recent fixed calibration transfers very well to the following year, whereas a broader calibration that includes the exceptional 2022 market regime is too conservative. The adaptive policies outperform this broader historical fixed rule, but not the recent one. The main lesson is therefore that adaptive valuation is economically meaningful, but useful in practice only when the longer-horizon scenarios capture the timing and ranking of future spread opportunities well enough. In the data available here, this condition is not met consistently enough to improve on the recent fixed calibration. A fixed lambda remains a strong industrial rule when it is well calibrated, while adaptive valuation provides a natural framework for moving beyond a single frozen annual scarcity parameter.