Equivalent Consumption Minimization Strategy

• ECMS is a real-time power-split control strategy for HEVs/PHEVs that integrates fuel consumption and battery energy costs through an equivalence factor.
• The methodology uses optimal control formulations, including Pontryagin’s Minimum Principle, to divide driver-demanded power between engine and motor efficiently.
• Variants such as fixed, SOC feedback, and RL-based λ adaptations enable integration into dynamic programming and hierarchical frameworks for robust, real-time energy management.

The Equivalent Consumption Minimization Strategy (ECMS) is a real-time power-split control methodology for hybrid electric vehicles (HEVs) and plug-in hybrid electric vehicles (PHEVs). It is based on augmenting the instantaneous fuel consumption with an “equivalent” electric energy cost, allowing the system to trade off fuel and battery usage efficiently at each control interval, and is underpinned by principles from optimal control and the Pontryagin Minimum Principle (PMP). ECMS delivers near-optimal energy management with computational tractability and provides a formalism to systematize energy equivalence through a key parameter, the equivalence factor λ (Deshpande et al., 2020, Zhang et al., 2024, Razi et al., 2020).

1. Mathematical Formulation of ECMS

The foundational ECMS objective is the minimization of a pseudo-consumption Hamiltonian that combines instantaneous fuel usage and the electric power drawn from the battery, mapped into an equivalent fuel cost:

$$J = \int_{t_0}^{t_f} \left[ P_f(t) + \lambda(t) P_e(t) \right] dt$$

where

• $P_f(t)$ is the fuel power consumption, typically $P_f = \dot{m}_{\text{fuel}}(t) \cdot \text{LHV}$,
• $P_e(t)$ is the battery power (positive for discharge),
• $\lambda(t)$ is the equivalence factor (dimensionless), converting battery power to fuel-equivalent units (Zhang et al., 2024).

In discrete-time implementations:

$$u^*_k = \arg\min_{u_k} \, \dot{m}_{f,k}(x_k, u_k) + \lambda \, f_{\text{pen},k}(x_k) \frac{P_{\text{batt},k}^{\text{des}}(x_k, u_k)}{Q_{\text{lhv}}}$$

where $x_k$ is the state (including SoC and velocity); $u_k$ denotes the control; $f_{\text{pen},k}$ penalizes deviations from SoC constraints (Deshpande et al., 2020).

The solution at each control interval is to split the driver-demanded power between engine and motor by minimizing the instantaneous Hamiltonian:

$$H(t) = P_f[P_{\text{ICE}}(t)] + \lambda(t) P_{\text{bat}}(t)$$

under system constraints, where $P_f(t)$0 (Zhang et al., 2024). In convexified settings, closed-form or second-order approximations to this optimization can be derived, with additional penalty terms encoding interior-point treatments of battery SoC boundaries (Razi et al., 2020).

2. Equivalence Factor (λ): Determination and Adaptation

The efficacy and optimality of ECMS depend critically on the choice and evolution of the equivalence factor $P_f(t)$1, which encodes the marginal value of electrical energy relative to fuel. Common approaches include:

• Constant λ (offline tuning): $P_f(t)$2 is fixed via long-horizon open-loop optimal control (e.g., Dynamic Programming), often with a shooting method to ensure global charge-sustaining behavior. Example: $P_f(t)$3 with a small $P_f(t)$4 (Deshpande et al., 2020).
• SOC feedback adaptation: In adaptive ECMS (A-ECMS), $P_f(t)$5 is explicitly adapted using proportional or integral feedback on SoC error:

$P_f(t)$6

(Zhang et al., 2024).

• Interior-point penalty functions: The battery penalty factor $P_f(t)$7 is modulated using tangent or logarithm penalties to represent soft SoC constraints:
  • Tangent-type:

  $P_f(t)$8

  (Razi et al., 2020).

• Horizon-extended or RL-based λ: Recent advances (RL-ECMS) generate $P_f(t)$9 via a state-dependent policy learned through reinforcement learning, using both past load and predicted future speed as inputs. This approach enables adaptation to longer-horizon disturbances, road events, and modeling error (Zhang et al., 2024).

3. Embedding ECMS into Hierarchical and Hybrid Optimization Architectures

While the classic ECMS solves a strictly myopic, stage-wise minimization, it can be embedded inside longer horizon optimization frameworks:

• Dynamic Programming–ECMS hybrid (DP-ECMS): For optimal eco-driving control, total powertrain torque is selected via DP, while the engine/motor power split is determined by ECMS. The DP-ECMS Bellman recursion takes the form:

$P_f = \dot{m}_{\text{fuel}}(t) \cdot \text{LHV}$0

with $P_f = \dot{m}_{\text{fuel}}(t) \cdot \text{LHV}$1 solved by ECMS at each DP iteration (Deshpande et al., 2020).

• Receding-horizon/rolling DP-ECMS: For real-time implementation, a truncated DP is solved online over a moving window (N_H steps), with $P_f = \dot{m}_{\text{fuel}}(t) \cdot \text{LHV}$2 selected at each interval by grid search and parallel rolling DP execution. The final cost is bootstrapped with the offline-computed value function (Deshpande et al., 2020).
• RL-ECMS integration: A reinforcement learning agent, trained via TD3, maps the extended state (including past/future load) into $P_f = \dot{m}_{\text{fuel}}(t) \cdot \text{LHV}$3, which is then used in ECMS to compute the instantaneous optimal power split (Zhang et al., 2024).

4. Practical Implementation, State Constraints, and Robustness

ECMS architectures are tailored to real-time deployment in HEV/PHEV supervisory controllers:

• State and Input Constraints: Battery power $P_f = \dot{m}_{\text{fuel}}(t) \cdot \text{LHV}$4 is constrained by actuator and battery limits; SoC is subject to admissible bounds. Actuator constraints are enforced through projection or penalty methods; SoC boundaries are usually handled via smooth interior-point penalty terms rather than hard thresholding (Razi et al., 2020).
• Convexity and Solvability: The composite stage cost (fuel-rate plus linear penalty in battery power) is convex under broad conditions, enabling efficient optimization. Second-order Taylor approximations may be leveraged to allow explicit updates for the control variable in each step (Razi et al., 2020).
• Robustness to Measurement Noise: A soft margin (offset in constraint boundaries) is designed to accommodate SoC measurement errors, ensuring that true SoC remains strictly within operational bounds (Razi et al., 2020).
• Disturbance Rejection and Adaptivity: RL-based ECMS controllers achieve higher disturbance resistance, yielding lower fuel-penalty increases under injected state noise and maintaining engine operation in high-efficiency regions (Zhang et al., 2024).

5. Performance Benchmarks and Trade-offs

Quantitative assessments demonstrate that:

• Sub-Optimality: DP-ECMS, when compared to full-state full-control DP, incurs at most a 2% cost penalty while reducing both state and action space cardinalities by an order of magnitude. Receding-horizon DP-ECMS can maintain suboptimality below 1% even with limited lookahead and grid resolution (Deshpande et al., 2020).
• Computation: Full-route DP-ECMS is feasible offline with coarse state spaces. Receding-horizon DP-ECMS achieves real-time capability (online computation on embedded processors) via horizon truncation and parallel λ grid search, with speedups of 10×–100× in DP table size (Deshpande et al., 2020).
• Fuel Economy: RL-ECMS outperforms both A-ECMS and rule-based controllers in real and simulated scenarios. Test-cycle results (WLTC) evidence ~2% lower fuel consumption versus A-ECMS and ~4.6% versus rule-based control, with engine start-stop frequency and robustness also improved under RL-ECMS (Zhang et al., 2024).
• Actuation: ECMS approaches may slightly penalize torque tracking if λ is miscalibrated; parameterization of the penalty terms is critical to balancing fuel efficiency and drivability (Razi et al., 2020).

Algorithm	Fuel Consumption (L/100 km)	Torque Tracking (%)	Robustness Comment
DP (offline)	4.58	–	Global optimum
RL-ECMS	4.93	–	High disturbance rejection
A-ECMS	5.03	–	Less robust
Rule-Based	5.17	–	Baseline
ECMS (tan/log)	5.54–18.9 g/s (ex)	99.3–99.6	Robust with soft bounds
LQT	5.55–18.9 g/s (ex)	99.7–100.0	Reference tracking

6. Limitations, Variants, and Extensions

• Shortcomings of Instantaneous λ Update: Adaptations based solely on instantaneous SoC error or ultra-short horizon driving cycles (1–5 s) cannot guarantee global fuel economy optimality, are sensitive to traffic and disturbance, and result in higher SoC and engine activity fluctuations (Zhang et al., 2024).
• Criticality of Equivalence Factor Selection: ECMS optimality is highly sensitive to correct tuning of the reference $P_f = \dot{m}_{\text{fuel}}(t) \cdot \text{LHV}$5 or $P_f = \dot{m}_{\text{fuel}}(t) \cdot \text{LHV}$6; inadequate settings degrade both fuel savings and torque delivery. Techniques such as soft penalty formulations, RL-based mapping, and horizon-extended controllers address some of these issues (Razi et al., 2020, Zhang et al., 2024).
• Integration with Hierarchical and Learning Controllers: Embedding ECMS within a broader optimization or learning framework (DP, ADP, or RL-augmented) enables adaptation to real-world traffic, road networks, and vehicle uncertainty, while preserving real-time feasibility (Deshpande et al., 2020, Zhang et al., 2024).

7. Applications and Research Directions

ECMS underpins the power-split logic in numerous HEV/PHEV architectures for both research and production powertrain supervisory control. Recent trends focus on:

• Further embedding within hierarchical eco-driving and velocity planning loops for connected, automated, and networked vehicles (Deshpande et al., 2020).
• Hybridized controllers (ECMS+RL) that blend interpretability and explainability with disturbance-robust, globally-efficient policies suitable for hardware-in-the-loop and OEM deployment (Zhang et al., 2024).
• Enhanced constraint-handling, real-world robustness, and systematic tuning of equivalence factors through convex optimization, RL, and system identification (Razi et al., 2020).
• Ongoing benchmarking against linear quadratic tracking, dynamic programming, and model-predictive control methods to define the empirical boundaries of ECMS optimality and computational efficiency.

The versatility and modularity of ECMS remain distinctive; its integration with modern data-driven and predictive control frameworks continues to be a central topic in energy management research for advanced hybrid vehicle platforms.