Deep Quantile Regression-Based Two-Stage Stochastic Optimization Framework for Ancillary Services Provision

This article presents a framework combining deep quantile regression (DQR) and two-stage stochastic optimization for ancillary service provision by an aggregate of prosumers. The purpose of the study is to support bidding decisions by considering uncertainties in electricity demand and photovoltaic generation together with the operation of electricity and hydrogen equipment. The aggregate includes battery storage, hydrogen storage, and an electrolyzer. DQR predicts the 5th, 50th, and 95th demand and generation quantiles, which are used to define three weighted operating scenarios. The service bid is determined in the first stage, while equipment schedules and grid exchanges are adjusted for each scenario in the second stage. Three optimization policies were investigated, which include deterministic median-forecast, expected-cost, and conditional value at risk (CVaR) optimization. These were evaluated across 24 operating conditions and different service prices. It was observed that, at service prices of 125, 130, and 135 €/MWh, the expected-cost approach reduced expected operating costs in ten upward-service conditions per price, with savings of up to €15 per one-hour case compared with median-only bidding. Moreover, at the baseline price of 200 €/MWh, CVaR reduced worst retained-scenario costs by declining ten upward-service offers, although expected operating costs increased. The results demonstrate how considering forecast uncertainty can help avoid economically unfavorable commitments and support the selection of bids according to expected returns and adverse-scenario costs.