Distributionally Robust Model Predictive Control for Virtual Power Plants This work was supported by FME Solar, funded by the Research Council of Norway and industry partners, under project number 350244.

Nikolas Recke, Mathias Hudoba de Badyn · 2026

Read Distributionally Robust Model Predictive Control for Virtual Power Plants This work was supported by FME Solar, funded by the Research Council of Norway and industry partners, under project number 350244. by Nikolas Recke, Mathias Hudoba de Badyn (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-14642.txt.

Introduction

Recke and Hudoba de Badyn present a distributionally robust model predictive control framework for virtual power plants under electricity price uncertainty. One model links buildings, battery storage and renewable generation under weather and market signals. Data-driven forecasts feed quantile-based uncertainty sets that adapt to forecast spread, yielding a tractable controller. On Nordic weather and market data, gains up to about eight tenths of a percent over standard predictive control are claimed for small ambiguity radii.

Catalog record

CREATORNikolas Recke, Mathias Hudoba de Badyn
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30792
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.14642

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Distributionally Robust Model Predictive Control for Virtual Power Plants This work was supported by FME Solar, funded by the Research Council of Norway and industry partners, under project number 350244. Nikolas Recke, Mathias Hudoba de Badyn Abstract This paper presents a distributionally robust model predictive control (DR-MPC) framework for the optimal Virtual Power Plant (VPP) operation under electricity price uncertainty. A unified VPP model is formulated that captures the interaction between buildings, battery storage, and renewable generation, all influenced by exogenous weather and market signals. The proposed approach integrates data-driven forecasting with quantile-based uncertainty quantification to construct time-varying Wasserstein ambiguity sets that adapt to forecast dispersion and distributional shifts. This yields a tractable DR-MPC formulation that incorporates predictive distribution information directly into real-time decision making. The method is evaluated using real weather and market data from a Nordic case study across two seasonal scenarios. The results show that DR-MPC improves economic performance relative to standard forecast-based MPC when the ambiguity radius is chosen appropriately, with consistent gains of up to 0.8 % for small radii across both seasonal scenarios. Larger radii become overly conservative and reduce revenue, underscoring the importance of proper radius selection. These findings demonstrate the practical value of distributionally robust optimization for uncertainty-aware VPP operation. keywords Economic Model Predictive Control, Virtual Power Plant, Distributionally Robust Model Predictive Control, Uncertainty Quantification, Data-Driven Forecasting 1 Introduction Renewable energy sources (RES) are playing an increasingly important role in global electricity generation, but their inherent variability introduces uncertainty in power generation that complicates grid stability and system operation ( 13 ; 9 ). In energy systems that additionally face uncertain electricity market prices as well as uncertainties in renewable generation and thermal demand, these uncertainties pose major challenges for optimal operational decision-making ( 12 ). Forecast errors in these variables can propagate through the control system, leading to suboptimal scheduling, increased costs, or constraint violations, therefore modern control schemes must explicitly account for uncertainty. Virtual Power Plants (VPPs) aggregate dispatchable and non-dispatchable sources, residential buildings, and battery storage into a single controllable entity capable of providing grid support and participating in electricity markets ( 9 ). However, effective VPP operation requires predictive, uncertainty-aware control strategies that handle the stochastic nature of both physical and market environments. The operation of VPPs under uncertain forecasts constitutes a decision-making under uncertainty problem for which several control paradigms have been proposed. Model Predictive Control (MPC) provides some inherent robustness through its receding-horizon structure, but its performance deteriorates under large, unmodeled disturbances ( 4 ). Robust MPC (RMPC) ensures constraint satisfaction for all admissible disturbances yet is often overly conservative, while Stochastic MPC (SMPC) reduces conservatism but depends on precise probabilistic knowledge rarely available in practice ( 10 ). Distributionally Robust Optimization (DRO) offers a data-driven alternative that optimizes against all distributions within a data-informed ambiguity set, providing robustness to distributional shift and model misspecification ( 11 ). DR-MPC bridges the gap between SMPC and RMPC by combining robustness, interpretability, and adaptability ( 10 ). The main contributions of this paper are as follows: • formulation of a distributionally robust optimal control problem for real-time VPP operation, jointly optimizing economic performance and thermal comfort while explicitly robustifying with respect to electricity price uncertainty, • construction of time-varying Wasserstein ambiguity sets from empirical quantile forecasts of the market price, • integration within a receding-horizon DR-MPC framework using causal probabilistic forecasts from the TiREx forecaster, • validation on a realistic Nordic case study using real weather and market data, demonstrating consistent revenue improvements over forecast-based MPC. 2 Virtual Power Plant Model The VPP considered in this work aggregates residential buildings, battery storage systems, renewable generation through wind and solar plants, and a grid interface into a single controllable entity, see Figure 1 . Figure 1: Conceptual illustration of a Virtual Power Plant. At each time step k , the VPP must satisfy an instantaneous power balance that links renewable production, building demand, battery operation, and grid exchange. The power balance at each time step is \displaystyle P_{\text{ren},k}-P_{\text{load},k}-P_{\text{bat},k}=P_{\text{grid},k}, (1) We represent the battery by a net electrical power P_{\text{bat},k}=P_{\text{ch},k}-P_{\text{dis},k} , where P_{\text{ch},k}\geq 0 and P_{\text{dis},k}\geq 0 denote the charging and discharging powers respectively, and positive values of P_{\text{bat},k} correspond to charging. Similarly, the grid exchange is expressed as P_{\text{grid},k}=P_{\text{exp},k}-P_{\text{imp},k} , where P_{\text{exp},k}\geq 0 and P_{\text{imp},k}\geq 0 denote the power exported to and imported from the grid. For tractability, import and export are represented as separate non-negative variables, without explicitly enforcing mutual exclusivity. 2.1 Energy Storage System The dynamics of the battery storage system are modeled by the discrete-time equation \displaystyle S_{k+1} \displaystyle=S_{k}+\eta^{\text{ch}}P^{\text{ch}}_{k}\tau-\frac{1}{\eta^{\text{dis}}}P^{\text{dis}}_{k}\tau, (2) subject to \displaystyle S^{\min}\leq S_{k}\leq S^{\max},\quad 0\leq P^{\text{ch}}_{k}\leq\bar{P}^{\text{ch}},\quad 0\leq P^{\text{dis}}_{k}\leq\bar{P}^{\text{dis}}, (3) where S_{k} denotes the state of charge (SoC), P^{\text{ch}}_{k} and P^{\text{dis}}_{k} are the charging and discharging powers, \eta^{\text{ch}} , \eta^{\text{dis}} the efficiencies, and \tau the time step in hours. 2.2 Renewable Energy Generation The VPP has access to different renewable assets, such as solar and wind plants. The power generated by the photovoltaic (PV) modules at each discrete time step k is modeled as \displaystyle P_{\text{solar},k} \displaystyle=\eta_{\text{PV}}\,A_{\text{PV}}\,I_{\text{sol},k}\big[1-\gamma_{T}\big(T_{\text{cell},k}-T_{\text{ref}}\big)\big], (4) where \eta_{\text{PV}} is the nominal efficiency, A_{\text{PV}} the panel area, I_{\text{sol},k} the solar irradiance, and \gamma_{T} the temperature coefficient. The cell temperature is estimated as T_{\text{cell},k}=T_{\text{air},k}+k_{T}I_{\text{sol},k} . The electrical power output of a wind turbine is given by \displaystyle P_{\text{wind},k} \displaystyle=\tfrac{1}{2}\,\rho_{\text{air}}\,A_{\text{rot}}\,\eta_{\text{wind}}\,v_{k}^{3}, (5) where \rho_{\text{air}} is the air density, A_{\text{rot}}=\pi r_{\text{b}}^{2} the rotor-swept area, \eta_{\text{wind}} the overall efficiency, and v_{k} the wind speed. For the aggregated VPP, the total renewable generation is expressed as \displaystyle P_{\text{ren},k} \displaystyle=\sum_{i=1}^{N_{S}}P_{\text{solar},i,k}+\sum_{j=1}^{N_{W}}P_{\text{wind},j,k}. (6) 2.3 Energy Demand of Residential Buildings The thermal dynamics of a residential building is represented by a three-resistance, two-capacitance (3R2C) model ( 2 ). The model describes the heat exchange between the outer and inner walls and the indoor air temperature. The continuous-time dynamics are given by \displaystyle\dot{T}_{1} \displaystyle=-\tfrac{hA+1/R}{C_{1}}T_{1}+\tfrac{1}{RC_{1}}T_{2}+\tfrac{hA}{C_{1}}T_{0}+\tfrac{1}{C_{1}}q_{\text{sol}}A, \displaystyle\dot{T}_{2} \displaystyle=\tfrac{1}{RC_{2}}T_{1}-\tfrac{hA+1/R}{C_{2}}T_{2}+\tfrac{hA}{C_{2}}T_{z}, \displaystyle\dot{T}_{z} \displaystyle=\tfrac{hA}{C_{z}}(T_{2}-T_{z})+\tfrac{1}{C_{z}}\!\left(Q_{\text{ihg}}+Q_{\text{HVAC}}\right), (7) where T_{1} , T_{2} , and T_{z} denote the wall and indoor air temperatures, and T_{0} is the outdoor temperature. The inputs to the 3R2C model are the outdoor air temperature T_{0} , the effective solar gains q_{\text{sol}}A , the internal heat gains Q_{\text{ihg}} , and the HVAC thermal power Q_{\text{HVAC}} , which represents the thermal energy exchanged between the HVAC system and the zone. After discretization with sampling period \tau , the system can be written in linear state-space form: \displaystyle x_{k+1}=A_{d}x_{k}+B_{d}u_{k},\qquad y_{k}=Cx_{k}, (8) where the discrete-time matrices are obtained via zero-order hold discretization of the continuous-time dynamics ( 7 ) with sampling period \tau , specifically A_{d}=e^{A_{c}\tau} and B_{d}=A_{c}^{-1}(A_{d}-I)B_{c} . The state vector is x_{k}=[T_{1},\,T_{2},\,T_{z}]^{\top} and the input vector u_{k}=[T_{0},\,q_{\text{sol}}A,\,Q_{\text{ihg}},\,Q_{\text{HVAC}}]^{\top} . The HVAC electrical power consumption is related to the thermal input via P_{\text{el},k}=Q_{\text{HVAC},k}/\text{COP} . For each building, the indoor temperature is constrained to remain within a predefined comfort band, \displaystyle T_{z}^{\min}\leq T_{z,k}\leq T_{z}^{\max},\qquad\forall\,k, (9) and the HVAC heat flow is limited by its rated capacity, \displaystyle-Q_{\max}\leq Q_{\text{HVAC},k}\leq Q_{\max}. (10) To avoid rapid fluctuations in the HVAC operation, we impose ramp-rate constraints, \displaystyle Q_{\text{HVAC},k}-Q_{\text{HVAC},k-1} \displaystyle\leq R_{\uparrow}, (11) \displaystyle Q_{\text{HVAC},k-1}-Q_{\text{HVAC},k} \displaystyle\leq R_{\downarrow},\qquad\forall\,k\geq 1, (12) Together with the discrete-time dynamics ( 8 ), these constraints define the feasible set of building trajectories over the MPC horizon. A sampling period of \tau=1 h is adopted throughout, consistent

Read the complete text