QUBO-enhanced SARIMA scenario generation for day-ahead market bidding of hybrid energy storage aggregators

Abstract

Day-ahead joint bidding exposes a hybrid energy storage aggregator (HESA) to uncertainty in the locational marginal price, regulation market clearing capacity price, and regulation market performance clearing price. Conventional seasonal autoregressive integrated moving average (SARIMA) modeling calibrates forecast residuals separately for each price and hour, without using their co-movement within complete historical days. This paper proposes Day-SARIMA, a quadratic unconstrained binary optimization (QUBO)-enhanced SARIMA method for uncertainty calibration. Each historical day is represented by a 24-h joint residual matrix. QUBO selects calibration days according to their relevance to the target day and the co-movement of the three prices, while penalizing similar residual patterns. The selected subset determines the price- and hour-specific deviation radii and adaptive budgets of a budgeted uncertainty set. Complete-day information therefore affects calibration rather than being imposed as explicit cross-price correlation constraints. The resulting set is embedded in a max-min-max two-stage robust bidding model solved by column-and-constraint generation (C&CG). Experiments using PJM Interconnection (PJM) market data from October 2024 to February 2025 yield 58 eligible out-of-sample operating days, with 56 matched days used for the primary comparison. QUBO reproduces the correlations among the three prices more accurately than fast forward selection and K-medoids and achieves the highest mean net revenue. Block-SARIMA uses approximately 13 times as many QUBO variables as Day-SARIMA but yields lower scenario similarity and coherence, indicating that more variables cannot replace complete-day information. Relative to conventional SARIMA, Day-SARIMA increases mean net revenue by $53.21/day, supported by paired statistical testing. It also reduces mean modeled cycling-degradation cost by 16.0% and the supercapacitor equivalent-full-cycle proxy by 20.5%, although these reductions should be interpreted as average trends. The results apply to the winter test period and reference HESA configuration used in this study.