Economic MPC Optimization — Cost-Optimal Real-Time Control for Energy, Buildings, and Process Industries

Why This Matters (Executive Summary)

A tracking controller follows a setpoint. That setpoint encodes the economic objectives — but only at the operating conditions for which it was computed. When disturbances are large, when the plant operates close to constraints, when the process never settles, when prices change faster than the Real-Time Optimization (RTO) cycle, or when multiple operating regimes produce costly transitions — the setpoint loses validity and the feedback controller operates the plant uneconomically without any mechanism to correct it.

Economic Model Predictive Control (EMPC) removes this gap by embedding the economic objective directly into the feedback controller. Disturbance rejection and economic optimization become a single closed-loop operation. Published results: 17%+ primary energy reduction in buildings, minimum-fuel greenhouse climate control, price-arbitrage dispatch for a 1 MW battery storage system, improved power capture with reduced structural loads on wind turbines.

When RTO + tracking MPC is insufficient and EMPC applies: unplanned perturbations or model mismatch invalidate the RTO setpoint or DRTO trajectory — RTO optimises steady-state setpoints, Dynamic RTO (DRTO) extends this to planned transient trajectories, but neither has feedback at the economic level to handle unplanned disturbances in real time; frequent switching between operating regimes causes losses at the transitions; the optimal operation is periodic rather than a fixed steady state; RTO and MPC use different models making computed setpoints unreachable; economic signals change faster than the RTO update rate; disturbances are persistent enough that the process never reaches steady state. If none of these hold, a well-tuned RTO + tracking MPC is the simpler and sufficient choice.

The Design Pattern Explained

A classic tracking controller — Proportional-Integral-Derivative (PID), Linear Quadratic Gaussian (LQG), H∞, or standard tracking Model Predictive Control (MPC) — follows a setpoint. That setpoint is the expression of economic plant operation objectives translated into a target value the feedback controller can pursue.

In simple cases an operator sets the target directly. When the relationship between setpoint and economic objectives becomes more complex — due to multi-variable interactions, system dynamics, or time-varying operating conditions — the setpoints, or full reference trajectories, may be calculated offline from an optimal control problem (OCP) or online by an optimization layer running above the feedback controller. Two variants exist: Real-Time Optimization (RTO) computes economically optimal steady-state setpoints assuming the process is at or near steady state; Dynamic Real-Time Optimization (DRTO) extends this to planned transient trajectories — grade changes, batch phase transitions, planned startups — where the dynamic path between operating points matters economically. Both keep setpoint or trajectory calculation separate from the tracking feedback controller: the feedback controller’s role is to reject perturbations and compensate for modelling errors, driving the plant towards whichever setpoint or trajectory it receives.

The limitation appears when large perturbations, operation close to constraints, or significant modelling errors cause the feedback action itself to negatively affect the economic objectives of plant operation — for example by continuously shifting the operating point away from the cost optimum during disturbance recovery or constraint saturation. In this situation the tracking controller is doing its job correctly (minimizing setpoint deviation) while the plant operates uneconomically, because the RTO setpoint or DRTO trajectory is computed for nominal conditions — RTO and DRTO can handle planned transitions, but once an unplanned perturbation pushes the system off the nominal path, neither has feedback at the economic level to re-optimise in real time.

Economic MPC (EMPC) resolves this by applying the feedback action directly on the economic objective. Instead of tracking a setpoint, the controller optimizes the economic cost function in closed loop, handling perturbations and modelling errors while simultaneously maintaining cost-optimal operation across the full dynamic trajectory — including during transitions and disturbance recovery where the layered architecture loses its validity.

A second, distinct use case arises when the plant operates in multiple regimes, each served by a separate tracking controller optimized for that regime. The switching zones between controllers are areas of poor economic performance: neither controller is designed for the transition, control authority may be limited near the switching boundary, and the system may oscillate or dwell in the transition region. If regime switches are frequent — as in a wind turbine alternating between partial-load operation (maximum power point tracking, generator torque control) and full-load operation (rated power, pitch control) depending on wind speed — the accumulated economic losses in the switching zones are significant. A single Economic Nonlinear MPC (NMPC) replacing all regime-specific tracking controllers handles these transitions naturally: the economic objective and the full nonlinear dynamics are encoded once, the optimizer finds the globally optimal trajectory across all operating regimes, and the switching itself becomes part of the optimization rather than a gap between controllers.

A third structural weakness of layered RTO + tracking MPC is model inconsistency. RTO typically uses a simplified steady-state model; the tracking MPC uses a separate dynamic model — different structures, parameters, and timescales. When RTO computes an optimal setpoint using its model, the dynamic model used by MPC may predict that setpoint is unreachable or may settle at a slightly different operating point. The result: the economic optimum computed by RTO is never actually achieved in practice, because the two layers are optimizing with respect to different models. EMPC uses a single dynamic model, eliminating this inconsistency by construction.

A fifth case arises when economic conditions — electricity prices, feedstock costs, demand signals, CO₂ tariffs — change on a timescale shorter than the RTO update cycle. RTO is typically executed every few minutes to hours; real-time electricity markets and demand response signals can change every 15 minutes or faster. A setpoint computed at the start of a pricing interval is already suboptimal by the end of it. EMPC incorporates price and demand forecasts directly into the control horizon and reoptimizes at every control step, continuously adapting the operating trajectory to the latest economic signals.

Finally, RTO’s steady-state assumption requires that the process actually reaches and stays near steady state between updates. When disturbances are persistent or frequent — continuous feed composition variation, weather-driven thermal loads, fluctuating grid conditions, variable inlet quality — the process never settles, and every RTO update is computed for a state the system has already left. EMPC treats the disturbance stream as a continuous input to the dynamic optimization: there is no steady-state requirement, and the controller remains economically valid regardless of how frequently operating conditions change.

The architecture typically includes: (1) a dynamic process model capturing energy/momentum/mass balances, (2) time-varying economic parameters as inputs (prices, forecasts, demand), (3) hard constraints on safety, quality, and actuator limits, and (4) longer prediction horizons (hours to days) to capture the relevant economic dynamics.

Economic objective functions — a comprehensive inventory: The full space of economic objectives for process control includes:

• Energy and fuel cost: electricity price × kWh consumed; natural gas price × Nm³ used; district heating tariff; steam cost per tonne.
• Throughput and production value: revenue per unit produced; batch cycle-time value; capacity utilization premium.
• Raw material and feedstock cost: reagent price × consumption; yield loss cost (off-spec product value).
• Component wear and maintenance cost: actuator cycling penalty; fatigue-life consumption rate; bearing/seal degradation cost; pump/compressor wear index.
• Emissions and environmental cost: carbon price × CO₂ intensity; NOₓ/SOₓ abatement cost; effluent treatment cost.
• Demand charge and peak power: peak power billing (€/kW/month); ramp rate penalty from grid operator.
• Energy storage arbitrage and grid services: buy-low/sell-high price differential; frequency regulation ancillary service revenue; capacity market revenue.
• Battery and equipment degradation: cycle aging cost (€/cycle); calendar aging cost at given state-of-charge (SOC) and temperature; catalyst deactivation rate.
• Quality and comfort premium: price differential for higher-specification product; thermal comfort compliance cost; crop quality yield multiplier.
• Water and utility consumption: cooling water cost; compressed air cost per Nm³; deionized water cost.
• Labor and operational overhead: operator intervention cost; shift scheduling efficiency.
• Risk and reliability: expected downtime cost; safety margin utilization cost; insurance premium reduction.

In practice, the chosen objective is a weighted combination of the terms most relevant to the process economics — the weighting translates operational priorities into a mathematical cost function, and the optimal compromise between the competing objectives is calculated.

Applications & Reference Implementations

Application 1: Wind Turbine Control — Economic NMPC on the NREL 5-MW Benchmark

An Economic Nonlinear MPC (NMPC) was developed by an IAV GmbH-led consortium of industrial and academic partners and applied to the National Renewable Energy Laboratory (NREL) 5-MW reference wind turbine — work that subsequently evolved into IAV Larus, a commercial Lidar-based pitch control product for wind turbines. IAV referred to this control concept internally as Realtime Supervisory Control (RSC) / Echtzeit Betriebsführung (EBF). A conventional wind turbine control system operates in two distinct regimes — partial-load (below rated wind speed: maximize power capture via generator torque control) and full-load (above rated wind speed: limit power via pitch control) — each served by a separate tracking controller. The transition zone between these regimes is a region of poor economic performance: both the power capture efficiency and structural load management are suboptimal as the system switches between controllers. The NMPC replaces both regime-specific controllers with a single economic cost function that directly trades off three competing objectives: power capture value, structural fatigue load (tower bending oscillations), and actuator wear (pitch utilization rate) — without a separate setpoint generator or regime-switching logic. Closed-loop simulation benchmarks showed better generator-speed tracking, softer pitch utilization, improved power capture, and reduced tower oscillations — with higher power fluctuations noted as a trade-off. Dr. Noga was an IAV employee and contributed to this project. ¹

Application 2: LHC Cryogenics — Economic NMPC for Superfluid Helium at 2 K

An output-feedback economic NMPC was applied to the superfluid helium cryogenic circuit of the Large Hadron Collider (LHC) at CERN (European Organization for Nuclear Research). The system operates at a setpoint of 1.9 K. Exceeding 2.1 K triggers the interlock and de-energises the magnets; above 2.16 K superfluid helium transitions to normal fluid, causing a quench. The margins are hard constraints, not soft targets. The cryogenic circuit is a strongly coupled multi-input multi-output system with nonlinear thermo-hydraulic dynamics: valve flow characteristics, heat exchange, and helium phase behaviour are all nonlinear, and the control inputs interact through shared thermal mass. The system is also spatially distributed — what is controlled is the maximum temperature across a 214 m long sector, not a single point. The spatial location of that maximum shifts dynamically depending on heat loads and flow conditions, making a fixed-setpoint tracking approach inadequate: the economic NMPC formulation is necessary precisely because the quantity to be constrained (the spatial maximum) is itself a dynamic, state-dependent function of the system. The legacy Proportional-Integral (PI) controllers, tuned independently for each loop, could not handle the nonlinearities, inter-loop couplings, and dynamically shifting spatial maximum simultaneously. To avoid constraint violations they had to be detuned to the point where control action became very slow — accepting poor disturbance rejection as the price of safety.

The NMPC used a first-principles thermo-hydraulic model paired with a Luenberger Observer and Moving Horizon Estimation (MHE) for state reconstruction, explicitly accounting for the nonlinear couplings in every optimization cycle. The result was superhuman: the NMPC stabilized bath temperature approximately 10× faster than the legacy PI system after the same perturbation, while never violating the 2.1 K hard constraint throughout recovery. Two configurations were tested — two control valves alone, and valves combined with 12 electric heaters — both outperforming PI control by a wide margin. Computation times of approximately 1 s for estimation and 7–14 s per optimization cycle are well within the slow thermal time constants of the cryogenic circuit. The NMPC remained a proof-of-concept and did not enter production: subsequent changes to the mechanical layout of the cryogenic circuit and additional systems limiting the magnitude of perturbations made the slow PI response acceptable for operational purposes, removing the immediate business case for deployment. Dr. Noga contributed to this work during his PhD as a CERN employee. ²

Application 3: Swiss Office Building — MPC Achieves 17% Energy Reduction (OptiControl-II)

The ETH Zurich OptiControl-II project — conducted in collaboration with Siemens Building Technologies as research partner — implemented economic MPC for a fully occupied Swiss office building, controlling thermally activated building systems (TABS), air handling units (AHU), and blinds over a seven-month field deployment. The economic objective explicitly minimizes non-renewable primary energy (NRPE) cost weighted by time-varying electricity tariffs. Results: approximately 17% NRPE reduction, translating to roughly 21.6 MWh/year per floor, with improved comfort compliance maintained throughout. Siemens was a research partner in the project and has since integrated these capabilities into commercial building products (Building X Energy Manager, Climatix C600) marketed as “smart optimization”. The pattern itself — economic MPC with a thermal mass as a storage buffer and time-varying energy prices as the economic driver — transfers directly to industrial utility systems: compressed air, chilled water, steam, and process heat networks where the same price-arbitrage and load-shifting logic applies at larger scale with higher economic stakes. ³

Application 4: Commercial Greenhouse — Infinite-Horizon Economic MPC Minimizes Heating Cost

Researchers at Wageningen University & Research (WUR) implemented an Infinite-Horizon Economic MPC for a commercial glass greenhouse, with the explicit economic objective of minimizing natural gas heating costs while maintaining indoor climate conditions for crop growth. The EMPC exploits a realistic multi-day weather and gas-price forecast to pre-heat the greenhouse thermal mass during low-price periods and reduce heating during peak-price windows. Compared to a conventional setpoint controller, the EMPC demonstrated measurable reductions in heating energy use while satisfying all crop-climate constraints (temperature bounds, humidity). The greenhouse case is a clean instance of economic MPC because the control horizon (hours to days), the economic driver (gas price × heating energy), and the soft constraints (climate bounds for crop biology) are all explicitly encoded in the cost function. The specific WUR research was not commercialized as a named product; Dutch greenhouse control vendors such as Priva and Argus offer climate computer systems with optimization features, and MPC is gradually entering their product lines industry-wide. ⁴

Application 5: Battery Energy Storage — Economic MPC for Price Arbitrage and Grid Services

ETH Zurich implemented MPC-based dispatch for a real 1 MW / 0.56 MWh battery energy storage system (BESS) operating in the Swiss grid, targeting both electricity price arbitrage and primary frequency regulation. The economic MPC optimized charge/discharge schedules against time-varying spot prices and frequency deviation signals while enforcing hard constraints on state-of-charge (SOC) limits, charge/discharge rates, and power electronics thermal limits. The study quantified achievable revenues across multiple grid service modes and identified the operating strategy that respects battery degradation constraints — directly addressing the BESS owner’s economic objective. This application was fully commercialized: the Zurich BESS was owned and operated by EKZ (the utility of the Canton of Zurich) as live grid infrastructure, with the MPC dispatch running in production. The pattern has since become industry standard — 66% of all US utility-scale battery capacity included arbitrage among its applications in 2024, and commercial BESS optimization software (PLEXOS, PCI Energy Solutions, Energy Exemplar) now implements exactly this logic at scale. ⁵

What This Means for Your Operations

Economic MPC delivers the largest returns when one or more of the following apply: (1) large disturbances, near-constraint operation, or model mismatch cause economically costly transients that RTO cannot correct; (2) the plant switches frequently between operating regimes, each with its own controller, with poor economics in the transition zones; (3) the optimal operation is cyclic rather than a fixed steady state; (4) the RTO and MPC use different models, making computed setpoints unreachable; (5) economic signals such as energy prices or demand change faster than the RTO update cycle; (6) disturbances are persistent or frequent enough that the process never reaches the steady state RTO assumes. If none of these conditions hold, a well-tuned RTO + tracking MPC is likely sufficient and simpler to commission.

Common prerequisites: a working regulatory control layer (stable base-layer PID/MPC), an RTO or equivalent setpoint strategy already in place, measurable economic drivers (energy meters, production counters), a dynamic process model, and access to forecast data (weather, prices, demand) if the economics are time-varying.

How We Deliver This (Engagement Model)

• Phase 0: Non-Disclosure Agreement (NDA) + data request — share process data, energy bills, control architecture, and current operating strategy.
• Phase 1: Fixed-scope discovery (2-4 weeks) — economic baseline quantification, dynamic model identification, constraint mapping, and feasibility/Return on Investment (ROI) estimate for economic MPC.
• Phase 2: Implementation + validation + commissioning — economic MPC development, simulation validation against historical data, online commissioning with operator oversight.
• Phase 3: Monitoring + training + scaling — performance dashboards tracking economic key performance indicators (KPIs), operator training on the economic objective, and extension to additional units or sites.

Typical KPIs to Track

• Energy efficiency: kWh/unit produced, non-renewable primary energy consumption, specific energy cost
• Revenue / savings: arbitrage revenue (€/MWh), heating cost reduction (€/m²/year), demand charge reduction
• Quality/comfort: Temperature variability, comfort compliance hours, crop yield impact
• Operator burden: Manual interventions per shift, time spent adjusting setpoints
• Constraint compliance: SOC limit violations, actuator saturation frequency, safety constraint margin

Risks & Prerequisites

• Model complexity: Economic MPC often requires longer prediction horizons and more detailed models than tracking MPC, increasing computational cost and commissioning effort.
• Economic data integration: Time-varying prices, forecasts, and demand signals must be reliably fed into the controller — data pipeline quality is critical.
• Operator trust: Operators accustomed to fixed setpoints may resist a controller that moves the operating point; training and transparent dashboards are essential.
• Baseline quantification: Without a clear energy/cost baseline, it is impossible to demonstrate ROI — invest in measurement before optimization.
• Regulatory control must work first: Economic MPC sits on top of a stable base layer; if PID loops are poorly tuned or sensors are unreliable, fix those first.

FAQ

Q: How is economic MPC different from real-time optimization (RTO) and dynamic RTO (DRTO)? A: RTO computes economically optimal steady-state setpoints and passes them to a tracking layer — it assumes the process is at or near steady state. DRTO extends this to planned transient trajectories: it can optimise the dynamic path through a grade change, batch phase transition, or planned startup. Both work well under nominal conditions. The shared limitation is the absence of feedback at the economic level: once an unplanned perturbation pushes the process off the nominal setpoint or trajectory, neither RTO nor DRTO can re-optimise in real time — the tracking controller continues to pursue a stale target while the plant operates uneconomically. EMPC resolves this by embedding the economic objective directly into the feedback controller, so disturbance rejection and economic optimisation are a single closed-loop operation that remains valid regardless of unplanned perturbations.

Q: What energy savings can we realistically expect? A: Published results range from approximately 17% NRPE reduction in buildings to measurable heating cost savings in greenhouses and arbitrage revenues in energy storage. Actual savings depend on the current baseline, process dynamics, and the magnitude of time-varying economic drivers. A Phase 1 discovery engagement quantifies the realistic potential for your specific system.

Q: Does economic MPC require a complete process model? A: The model must capture the dominant energy/material dynamics and the key economic drivers. It does not need to model every physical detail. In practice, a combination of first-principles structure and data-driven parameter identification yields models that are sufficient for economic optimization within 4-8 weeks of commissioning effort.

Q: Can economic MPC work with our existing control infrastructure (DCS/PLC)? A: Yes. Economic MPC typically runs as a supervisory layer that sends setpoints or trajectories to existing regulatory controllers. Integration requires an OPC (OLE for Process Control) or similar data interface to the Distributed Control System (DCS) or Programmable Logic Controller (PLC), not a control system replacement.

Q: Can Economic MPC find a periodic operating cycle that is more profitable than any fixed setpoint? A: Yes — and this is a capability that RTO + tracking MPC structurally cannot match. RTO is designed to find an optimal steady-state operating point and hand it to a tracking controller. If the economically optimal strategy is not a steady state but a repeating cycle — periodic feeding, temperature cycling, pressure swing — RTO will never find it, because it only searches the space of steady states. EMPC’s objective function has no requirement to converge to a steady state: the optimizer finds and maintains a periodic orbit if that is cheaper. This has been demonstrated in pressure swing adsorption, bioreactor fed-batch operation, and cyclic catalytic processes.

Q: Our RTO and MPC use different models. Does this cause problems? A: Yes — this is a well-documented structural weakness of layered architectures. RTO typically uses a simplified steady-state model for speed; MPC uses a separate dynamic model for trajectory tracking. When the two models disagree, the setpoint RTO computes as optimal is not achievable according to the dynamic model: the MPC will settle at a slightly different operating point, and the economic optimum is never actually reached. EMPC uses a single dynamic model for both economic optimization and control, eliminating this inconsistency by construction and ensuring that what the optimizer plans is what the controller executes.

Q: Our energy prices and demand signals change every 15 minutes. Can the RTO keep up? A: Typically not at that rate. RTO is usually executed every few minutes to hours and assumes economic conditions are approximately constant over its update interval. When prices, demand charges, or grid signals change on a 15-minute or shorter cycle — as in real-time electricity markets and demand response programmes — the setpoint from the previous RTO run is already suboptimal before the next update arrives. EMPC incorporates price and demand forecasts directly into the control horizon and reoptimizes at every control step (seconds to minutes), continuously adapting the operating trajectory to the latest economic signals without waiting for the next RTO cycle.

Q: Our process never really reaches steady state — disturbances are constant. Is RTO still valid? A: No — RTO’s steady-state assumption requires the process to actually settle near the computed setpoint between updates. If disturbances are persistent or frequent — continuous feed composition variation, fluctuating ambient conditions, variable grid load — the process is always in a transient state and the RTO setpoint is always computed for conditions the plant has already left. EMPC treats the disturbance stream as a continuous input to the dynamic optimization: there is no steady-state requirement, and the controller remains economically valid regardless of disturbance frequency. This makes EMPC the natural choice for processes driven by weather, markets, or variable feedstock quality.

Q: Can Economic MPC help when our plant switches between different operating regimes? A: Yes — this is one of the strongest use cases for Economic NMPC. Plants that operate in multiple regimes (e.g., partial-load vs. full-load in wind turbines, different reaction phases in batch processes, or heating vs. cooling in Heating, Ventilation, and Air Conditioning (HVAC)) typically use separate tracking controllers optimized for each regime. The transition zone between controllers is a region of poor economic performance: neither controller handles the switch well, and if transitions are frequent, the accumulated losses are significant. A single Economic NMPC encodes the full nonlinear dynamics and the economic objective once — the optimizer finds the globally optimal trajectory across all regimes, and regime transitions become part of the optimization rather than a gap between controllers.

Q: What is the difference between linear economic MPC and nonlinear economic MPC (NMPC)? A: Whether the underlying optimization is linear or nonlinear depends on both the objective function and the constraints together — not the dynamics model alone. Linear economic MPC applies when the optimization objective is linear or at most quadratic AND all constraints are linear (the process model enters as linear equality constraints), yielding a convex Quadratic Program (QP) or Linear Program (LP): globally optimal solution guaranteed, fast and reliable even on modest hardware. A linear process model is necessary but not sufficient — a linear model paired with nonlinear constraints or an objective more complex than quadratic still results in a Nonlinear Program (NLP). Nonlinear economic MPC (NMPC) is required whenever the dynamics model is nonlinear, or when constraints or the cost function are nonlinear; the optimization is a non-convex NLP: significantly more demanding, local optima are possible, solver initialization and warm-starting are critical engineering decisions. In practice, however, nonlinear optimization works surprisingly well when the problem is carefully formulated and globalization and fallback strategies are implemented — this is a matter of engineering craft rather than a fundamental barrier. Dr. Noga specializes in Economic NMPC. In the applications on this page: buildings, greenhouse, and battery storage use linear dynamics with QP/LP structure; wind turbine and LHC cryogenics require full NMPC due to nonlinear aerodynamic/structural and thermo-hydraulic models respectively.

Q: How quickly can economic MPC deliver measurable results after commissioning? A: In practice, economic MPC can be commissioned and delivering measurable economic results within a few weeks of engagement start, provided the base regulatory layer is stable and process data is available. A typical progression: 1–2 weeks for model identification and economic baseline quantification, 1–2 weeks for offline simulation and validation, 1–2 weeks for online commissioning and operator acceptance. The exact timeline depends on data availability, DCS integration complexity, and the number of economic objective terms.

Book a 30-Minute Discovery Call

Ready to explore whether this pattern fits your system?

Dr. Rafal Noga — Independent APC/MPC Consultant

📧 Email me · 🌐 noga.es

Fixed-scope discovery · NDA-first · DACH on-site available

Public References

Footnotes

1. Schild, “Control-oriented modeling and controller design for wind turbines” — presentation by IAV engineer at the University of Freiburg, delivered as part of the IAV GmbH-led consortium (2018). This work evolved into IAV Larus — Lidar-based pitch control for wind turbines. IAV internal designation: Realtime Supervisory Control (RSC) / Echtzeit Betriebsführung (EBF). https://www.iav.com/de/produkte-und-services/iav-larus/ ↩

2. “NMPC for the Superfluid Helium Cryogenic Circuit of the LHC” (IFAC PapersOnLine, 2015). Dr. Noga contributed to this work during his PhD as a CERN employee. ↩

3. Sturzenegger et al., “Model Predictive Climate Control of a Swiss Office Building: Implementation, Results, and Cost-Benefit Analysis” (IEEE TCST / ETH Research Collection, 2016). https://www.research-collection.ethz.ch/bitstreams/1a73128f-0bc3-4f40-a001-39d53e0cf491/download ↩

4. Van Beveren et al., “Minimal heating energy use combining an Infinite Horizon Economic Model Predictive Controller with a Realistic Disturbance Forecast for a Greenhouse” (Biosystems Engineering, Wageningen University & Research, 2015). https://doi.org/10.1016/j.biosystemseng.2015.06.004 ↩

5. Koller et al., “Review of Grid Applications with the Zurich 1 MW Battery Energy Storage System” (Electric Power Systems Research, ETH Zurich, 2015). https://doi.org/10.1016/j.epsr.2014.06.023 ↩