Data-driven predictive control of battery energy storage with physics-informed sequence modeling

Integrating battery energy storage systems (BESS) into industrial sites with photovoltaic generation and time-varying tariffs requires controllers that combine economic performance, low real-time computational cost, and robustness to forecast errors. Model predictive control (MPC) achieves near-optimal cost but solves a mixed-integer optimization at every step using a linearized battery model, since nonlinear formulations are intractable for real-time deployment. Learning-based controllers remove this online optimization, but they are typically trained without any knowledge of the equipment they operate on; they are tuned site by site, and they provide no structural guarantee that operational constraints will be met. Closing this gap is the motivation of this work. This paper proposes PI-DT, a data-driven predictive controller based on a transformer sequence model conditioned on the return-to-go, trained offline on linear MILP expert actions executed on a nonlinear Thévenin two-branch RC battery simulator, and informed by per-battery physics, so that a single network operates batteries of different capacity, rated power, and efficiency without per-site retraining. An asymmetric minimum terminal state of charge (SOC) constraint, closer to industrial practice than a fixed target, is enforced consistently across expert, MPC, and the proposed controller, and enforced at inference through a horizon-aware projection on the action. The evaluation uses the public EMSx dataset (60 industrial sites, 37,442 site-days, Spanish PVPC tariff) and compares against the Rolling MPC, behavior cloning (BC), and Implicit Q-Learning (IQL), all assessed under the same nonlinear battery dynamics. On 5577 test site-days, the proposed controller reaches an aggregate cost gap of 1.39% versus the MILP, below 1.54% (BC), 1.61% (MPC), and 1.96% (IQL). The learning baselines fall below the minimum terminal SOC by up to 8.4 and 29 percentage points, respectively; the projection makes the reserve feasible by construction under the linear SOC model on which it is computed, and on the nonlinear plant the worst-case residual deficit of the proposed controller is 0.14 pp, within the 1 pp operational tolerance. A 3-fold cross-validation on held-out sites shows it transfers to unseen batteries with negligible degradation. The proposed controller therefore attains MPC-level economics without an online solver and with a structurally enforced end-of-day reserve, which makes it a candidate for embedded deployment in industrial energy management systems, subject to hardware-in-the-loop validation.

Authors

Institutions

Publication Details

Journal
Journal of Energy Storage
Published
2026-11-16
DOI
https://doi.org/10.1016/j.est.2026.125523
Primary Topic
Advanced Control Systems Optimization
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Data-driven predictive control of battery energy storage with physics-informed sequence modeling

Vicente Torres‐Sanz, Piedad Garrido, Ricardo Jorge Bessa, Julio A. Sanguesa et al.
1 links
Journal of Energy Storage
Advanced Control Systems Optimization
article

Data-driven predictive control of battery energy storage with physics-informed sequence modeling

Vicente Torres‐Sanz, Piedad Garrido, Ricardo Jorge Bessa, Julio A. Sanguesa, Francisco J. Martínez, Pablo Donate
article en

Abstract

Integrating battery energy storage systems (BESS) into industrial sites with photovoltaic generation and time-varying tariffs requires controllers that combine economic performance, low real-time computational cost, and robustness to forecast errors. Model predictive control (MPC) achieves near-optimal cost but solves a mixed-integer optimization at every step using a linearized battery model, since nonlinear formulations are intractable for real-time deployment. Learning-based controllers remove this online optimization, but they are typically trained without any knowledge of the equipment they operate on; they are tuned site by site, and they provide no structural guarantee that operational constraints will be met. Closing this gap is the motivation of this work. This paper proposes PI-DT, a data-driven predictive controller based on a transformer sequence model conditioned on the return-to-go, trained offline on linear MILP expert actions executed on a nonlinear Thévenin two-branch RC battery simulator, and informed by per-battery physics, so that a single network operates batteries of different capacity, rated power, and efficiency without per-site retraining. An asymmetric minimum terminal state of charge (SOC) constraint, closer to industrial practice than a fixed target, is enforced consistently across expert, MPC, and the proposed controller, and enforced at inference through a horizon-aware projection on the action. The evaluation uses the public EMSx dataset (60 industrial sites, 37,442 site-days, Spanish PVPC tariff) and compares against the Rolling MPC, behavior cloning (BC), and Implicit Q-Learning (IQL), all assessed under the same nonlinear battery dynamics. On 5577 test site-days, the proposed controller reaches an aggregate cost gap of 1.39% versus the MILP, below 1.54% (BC), 1.61% (MPC), and 1.96% (IQL). The learning baselines fall below the minimum terminal SOC by up to 8.4 and 29 percentage points, respectively; the projection makes the reserve feasible by construction under the linear SOC model on which it is computed, and on the nonlinear plant the worst-case residual deficit of the proposed controller is 0.14 pp, within the 1 pp operational tolerance. A 3-fold cross-validation on held-out sites shows it transfers to unseen batteries with negligible degradation. The proposed controller therefore attains MPC-level economics without an online solver and with a structurally enforced end-of-day reserve, which makes it a candidate for embedded deployment in industrial energy management systems, subject to hardware-in-the-loop validation.

Journal of Energy StorageVol. 183
Universidade do Porto (PT), Universidad de Zaragoza (ES), INESC TEC (PT)
Ministerio de Ciencia, Innovación y Universidades, Gobierno de Aragón, HORIZON EUROPE Framework Programme
Affordable and clean energy, Industry, innovation and infrastructure
Advanced Control Systems Optimization
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon