In the evolving energy landscape, the integration of prosumers—consumers who also produce energy—has gained significant attention. Researchers from the Technical University of Denmark, including Yogesh Pipada Sunil Kumar, S. Ali Pourmousavi, Jon A. R. Liisberg, and Julian Lesmos-Vinasco, have been exploring ways to better harness the flexibility of these prosumers to support grid stability. Their recent work, published in the journal “Applied Energy,” focuses on improving the reliability of forecasting prosumer flexibility, which is crucial for demand response aggregators participating in ancillary service markets.
The researchers highlight that accurate forecasting of prosumer flexibility is essential for demand response aggregators, especially in markets like the Danish manual frequency restoration reserve capacity market, where strict reliability standards such as the P90 standard are enforced. The P90 standard requires that forecasts are accurate 90% of the time, ensuring that the grid can rely on the promised flexibility. However, forecasting prosumer flexibility is challenging due to limited historical data, dependence on external factors like weather and market prices, and the diverse behaviors of individual prosumers. These challenges introduce significant uncertainty, making traditional deterministic or poorly calibrated probabilistic models unsuitable for market bidding.
To address these issues, the researchers propose a scalable uncertainty quantification framework that combines Monte Carlo dropout (MCD) with conformal prediction (CP). Monte Carlo dropout is a technique used to estimate the uncertainty in machine learning models by randomly dropping out neurons during prediction. Conformal prediction, on the other hand, provides a way to produce calibrated prediction intervals that account for the uncertainty in the data. By integrating these two methods, the researchers aim to produce reliable, finite sample prediction intervals for aggregated prosumer flexibility.
The proposed framework was tested using a large-scale synthetic dataset generated from a modified industry-grade home energy management system. This dataset included publicly available data on load, solar generation, prices, activation signals, and device-level information. The resulting machine learning surrogate model captured the aggregate price responsiveness of prosumers and provided uncertainty-aware estimates suitable for market bidding.
The researchers evaluated multiple multivariate conformal prediction strategies and benchmarked them against conventional MCD-based methods. Their results showed that standalone MCD systematically overestimated available flexibility and violated the P90 compliance, leading to potential overbidding and regulatory non-compliance. In contrast, the proposed MCD-CP framework achieved reliable coverage with controlled conservatism, significantly reducing the risk of overbidding.
When embedded in an aggregator’s bidding model, the conformalized methods substantially improved bidding performance, achieving up to 70% of the profit that would be possible with perfect information while satisfying regulatory reliability constraints. This approach provides a practical, computationally efficient, and market-compliant solution for aggregator flexibility forecasting under uncertainty.
The research offers valuable insights for the energy sector, particularly for demand response aggregators and grid operators. By improving the reliability of prosumer flexibility forecasts, this framework can enhance the participation of prosumers in ancillary service markets, contributing to grid stability and the integration of renewable energy sources. The practical applications of this research are significant, as it provides a robust method for managing uncertainty in a rapidly changing energy landscape.
This article is based on research available at arXiv.