Calibrated uncertainty quantification for prosumer flexibility aggregation in ancillary service markets

TL;DR

This work tackles uncertainty in forecasting aggregator-provided prosumer flexibility for frequency-restoration markets under the $P90$ reliability constraint. It introduces a hybrid uncertainty quantification framework that merges Monte Carlo dropout with conformal prediction to produce calibrated, finite-sample prediction intervals for multivariate, multi-time-step forecasts of aggregated flexibility. The authors apply the method to a Danish mFRR context with a large synthetic dataset, showing that standalone MCD is miscalibrated while MCD-CP achieves reliable coverage and reduces overbidding, enabling profits up to about 70% of the perfect-information benchmark. They also develop and compare several multivariate CP strategies within the aggregator bidding optimization, and highlight practical implications for market design and prosumer engagement. The work demonstrates a scalable, market-ready approach to uncertainty-aware bidding under data scarcity.

Abstract

Reliable forecasting of prosumer flexibility is critical for demand response aggregators participating in frequency controlled ancillary services market, where strict reliability requirements such as the P90 standard are enforced. Limited historical data, dependence on exogeneous factors, and heterogenous prosumer behaviour introduce significant epistemic uncertainty, making deterministic or poorly calibrated probabilistic models unsuitable for market bidding. This paper proposes the use of scalable uncertainty quantification framework that integrates Monte Carlo dropout (MCD) with conformal prediction (CP) to produce calibrated, finite sample prediction intervals for aggregated prosumer flexibility. The proposed framework is applied to a behind-the-meter aggregator participating in the Danish manual frequency restoration reserve capacity market. A large-scale synthetic dataset is generated using a modified industry-grade home energy management system, combined with publicly available load, solar, price, activation and device-level data. The resulting machine learning surrogate model captures aggregate prosumer price responsiveness and provides uncertainty-aware estimates suitable for market bidding. Multiple multivariate CP strategies are evaluated and benchmarked against conventional MCD-based methods. Results show that standalone MCD systematically overestimates available flexibility and violates P90 compliance, whereas the proposed MCD-CP framework achieves reliable coverage with controlled conservatism. When embedded in aggregator bidding model, conformalised methods substantially reduce overbidding risk and achieve upto 70% of perfect-information profit while satisfying regulatory reliability constraints, providing practical, computationally efficient, and market-compliant solution for aggregator flexibility forecasting under uncertainty.

Figures (10)

• Figure 1: Aggregator business model
• Figure 2: Example conformal prediction region $\xi$ for $\boldsymbol{y} \in \mathbb{R}^{2}$ with bounding rectangle $\mathcal{B}_{\xi}$
• Figure 3: Illustration of prediction regions $\xi$ and corresponding bounding hyperrectangles for $\boldsymbol{y} \in \mathbb{R}^{2}$, covering all methods evaluated.
• Figure 4: Synthetic data generation framework
• Figure 5: Error metrics for the NN trained with hyperparameters in Table \ref{['tab: ch5 nn hyperparams']}