Li et al AI-Driven Virtual Power Plants Review (2026)
Source details
- Type
- Paper
- Publisher
- MDPI (Energies)
- Author
- Jian Li, Chenxi Wang, Yonghe Liu
- Published
- 2026-02-20
- Pages
- 28
- Link
- doi.org/10.3390/en19041084
“AI-Driven Virtual Power Plants: A Comprehensive Review.” Jian Li, Chenxi Wang, Yonghe Liu — Department of Computer Science and Engineering, University of Texas at Arlington. Energies 2026, 19(4), 1084. DOI: 10.3390/en19041084. Open access (CC BY), published 20 February 2026.
Not Sweden- or EU-specific. Global academic literature review (28 pages, ~100 references) — a taxonomy and method-comparison reference, cited here sparingly per this wiki’s Nordic-depth-over-global-breadth scope.
What it is
A systematic review mapping AI method families (machine learning, deep learning, reinforcement learning, hybrid/collaborative frameworks) onto Virtual Power Plant functions (forecasting, scheduling, market bidding, aggregation/coordination, ancillary services), plus a comparison of centralized-cloud, edge/distributed, and hierarchical cloud-edge deployment architectures. Global aggregated VPP capacity is cited at ~35 GW, with major deployments in Germany, the US, Australia and China; the concept traces to Awerbuch’s 1997 “virtual utility” and the EU-funded FENIX project (2005–2009, Spain/UK).
Task-based method comparison (the review’s central table)
No single AI method dominates across all VPP functions; the review’s assessment by function:
| VPP function | Best-fit AI approach | Note |
|---|---|---|
| Forecasting | Hybrid deep learning (Transformer + graph neural network) | Strong spatiotemporal generalisation across multi-energy sources |
| Scheduling/dispatch | Reinforcement learning, esp. hybrid RL + model predictive control | Adapts to dynamic environments; more resilient than pure RL |
| Market bidding | Cooperative/risk-aware learning (safe RL, game theory) | Suited to real-time bidding under coupled, fluctuating markets |
| Aggregation/coordination | Multi-agent RL, federated learning, graph neural networks | Scales to large DER clusters while preserving data privacy |
| Ancillary services/resilience | Deep RL, digital twins, fault-diagnosis algorithms | Supports predictive maintenance, autonomous frequency/voltage regulation |
Six evaluation dimensions are used throughout: predictive accuracy, operational performance, robustness/uncertainty handling, scalability, computational/deployment cost, interpretability. Across AI method families generically: machine learning is lightweight and interpretable but only moderately accurate; deep learning is highly accurate but computationally expensive and a “black box”; reinforcement learning adapts well but trains expensively and isn’t interpretable; federated/hybrid approaches score highest on scalability (“very high”), with machine learning, multi-agent RL, and graph neural networks close behind (“high”).
Architecture: the shift to hierarchical cloud-edge
Centralized (cloud) architectures give full data visibility and strong global optimisation but suffer latency and bandwidth dependency, unsuitable for sub-second control loops. Edge/distributed architectures cut latency and improve resilience under intermittent connectivity but are constrained by limited local compute. The review frames the emerging consensus as hierarchical cloud-edge collaboration — cloud for long-term planning/training, edge for real-time inference and local correction — citing commercial platforms (AutoGrid Flex, Siemens Grid Edge) achieving <100ms end-to-end scheduling latency this way. 5G and LEO satellite links (Starlink cited) are named as extending this pattern to remote/off-grid VPP assets.
Challenges and future directions (the review’s own conclusions)
Named limitations of the current “narrow AI” VPP paradigm: data fragmentation/inconsistency across heterogeneous sources (solar inverters, batteries, building loads, market prices); the DL/RL “black-box” problem undermining regulatory/operator trust; poor cross-region generalisation of models trained in one regulatory/geographic context. Six proposed research directions (the paper’s own text labels them “five” but its list runs to six): physics-informed AI (embedding grid physical laws into model architecture); explainable/trustworthy AI; collaborative edge-cloud intelligence; privacy-preserving federated learning; foundation models/LLMs — including time-series foundation models (TimesFM, Chronos named) for zero-shot forecasting that reduces new-site data dependency, and LLM-based multi-agent “grid agents” for interpreting shifting market rules; and multi-energy complementarity/cross-domain integration — extending the electricity-centric VPP paradigm to hydrogen, heat, and gas for deeper decarbonization.
Relevance to this wiki
- A reference-quality taxonomy for Virtual Power Plant‘s currently thin treatment of how AI is applied to VPP operation, used to add a compact methods-and-architecture note rather than exhaustively catalogued (this wiki’s scope is Sweden/Nordic depth, not global AI-methods literature).
- The federated-learning-for-shared-forecasting theme independently recurs in Source - Digitala Monster — AI och Svenska Energisystemet (2026)‘s Energiforsk recommendations — arrived at there via a separate citation (RISE’s federated-learning explainer, plus a practitioner interview) rather than this review, corroborating it as a genuinely emerging cross-cutting direction rather than one author’s idea.
- The time-series foundation model point (TimesFM/Chronos, zero-shot cross-region forecasting) is relevant background for Source - Energiforsk 2026-1168 AI-modeller Prognostisering Efterfrågan El (2026)‘s DESGRID scoping work, though DESGRID does not currently plan to use this approach.