Decision-Flux Optimization-based energy management for battery-supercapacitor hybrid storage in electric vehicles: battery lifetime and efficiency enhancement Authors: Zhen Wen, Maroun Kuoki, Leren Qian, Mohammad Khishe Published: 2026 ABSTRACT Battery-supercapacitor hybrid energy storage systems can reduce the transient loading of lithium-ion batteries in electric vehicles, but their benefit strongly depends on the energy-management parameters used to distribute power between the two storage devices. This study proposes Decision-Flux Optimization (DFO), a simulation-driven feasible-descent framework for calibrating the parameters of a battery/supercapacitor energy management strategy under full driving-cycle constraints. DFO uses induced operational trace of a candidate decision vector which comprises of battery state of charge, supercapacitor state of charge, power limits, voltage limits and end-of-cycle recovery constraints. Constraint satisfaction is considered as an admissibility condition and not as a term in the objective function. The objective is a combination of normalized RMS battery power, normalized energy-loss indicators, and only accepted updates are feasible trace-level transitions. The optimized parameters are computed offline and then deployed in the real-time control layer. Evaluation of the method is performed on ARTEMIS driving cycle through simulation and laboratory-scale controller validation. Compared with GA, PSO, RB-EMS, MPC, SRCPO, DSCEO, and ADRL-EMS benchmarks, DFO produced the lowest RMS battery power in the reported case study and maintained feasibility across repeated runs. The aging analysis indicates a model-estimated lifetime of 3720 equivalent cycles before the 80% capacity threshold, which should be interpreted as a comparative model-based durability estimate rather than a direct long-term lifetime measurement. Results indicate that trace level feasibility gating can be used to enhance repeatability, battery stress mitigation, and energy management for efficiency. Introduction The electrification of transportation on the roads has gone at a rapid pace, and so has the need for energy management strategies that not only improve the efficiency, durability, and operational reliability of electric vehicles (EVs)1,2. While lithium-ion batteries are still the dominant onboard energy storage technology because of their high energy density and maturity3, their degradation under high power transients, frequent charge–discharge cycling and regenerative braking is still a major barrier to cost-effective and sustainable EV deployment4,5. Hybrid energy storage systems (HESS), usually consisting of lithium-ion batteries and supercapacitors, have become a viable architectural solution to separate energy and power demands to reduce the stress on batteries and prolong service life6–9. However, the benefits of HESS are strongly dependent on the effectiveness of the energy management strategy (EMS). Poorly designed EMSs can offset the potential benefits of hybridization, resulting in unnecessary complexity, mass and marginal improvement in lifetime10. As a result, the joint optimization of energy management and system configuration has emerged as a major research problem in EV energy systems. From a computational point of view, hybrid energy management is a well-constrained, nonlinear and simulation-driven optimization problem11. Decision variables that are related to power split, control gains, and storage utilization interact throughout the vehicle dynamics, converter limitations, and state constraints while feasibility regions are often narrow and fragmented. Conventional methods of computational intelligence—even if they are very powerful in unconstrained or weakly constrained environments—are faced with fundamental challenges when applied to such cyber-physical systems. Population-based metaheuristic algorithms are still widely used because of their gradient-free characteristic and flexibility12,13. Nevertheless, because of their reliance on stochastic exploration and penalty-based handling of constraints, casual feasibility, slow convergence and sensitivity to parameter tuning is often the case when constraints are the dominant aspects of the solution space14–16. Model predictive control (MPC) approaches provide a systematic way to handle constraints, however, the effectiveness of MPC highly depends on the model fidelity, horizon tuning and computational resources thus limiting their scalability in joint sizing and management problems17. More recently, deep reinforcement learning (DRL) has been investigated for hybrid energy management but training instability, poor generalization and unguaranteed feasibility issues prevent it from being reliable for safety–critical EV applications18,19. These limitations imply there may not be enough refinements of the existing optimization frameworks. Instead, there is a new computational intelligence paradigm need that addresses the dominance of constraints, dynamic feasibility and system-level stability in a unified way, without the need for stochastic population evolution, predictive horizons and extensive offline training20. Motivated by this gap, this paper proposes Decision-Flux Optimization (DFO), which is a novel optimization framework that reformulates hybrid energy management optimization to be a process of stabilizing decision-induced operational responses over full driving cycles. Rather than exploring the solution space using evolutionary or predictive processes, DFO inhibits unstable decision transitions based on their observed system-level effects with feasible and high-performance configurations emerging intrinsically. The validity of the proposed framework is demonstrated for a lithium-ion battery/supercapacitor HES on the ARTEMIS driving cycle, achieving excellent feasibility consistency, battery stress reduction and lifetime improvement compared to state-of-the-art methods. The objective of this work is therefore not to optimize the physical sizing of the battery and supercapacitor modules, but to calibrate the parameters of a HESS energy-management policy under cycle-level feasibility constraints. The new scope is optimization of EMS parameters given a fixed battery/supercapacitor architecture. This scope is in line with the mathematical formulation, experimental validation and aging oriented comparison presented in the following sections. Related work Hybrid energy storage system and energy management Battery-supercapacitor hybrid energy storage systems have been widely investigated as a solution to enhance the transient power capability and regenerative energy recovery capability in EVs21. Recent surveys highlight that although hardware configurations are fairly well established, the EMS is a dominant factor that determines the lifetime benefits and overall system efficiency22. Rule-based EMSs are still attractive for real-time implementation for their simplicity but are limited in adaptability with respect to driving conditions and battery aging states23. To overcome these limitations, optimization-based EMSs have been proposed, e.g. dynamic programming, Pontryagin-based formulations24. Although these approaches offer theoretical optimality they are usually limited to off-line analysis and fixed system configurations, thus they are not suited for joint design and real-time deployment. Computational intelligence and metaheuristic optimization Metaheuristic algorithms such as particle swarm optimization, genetic algorithms, marine predators algorithm and Harris hawks optimization have been widely used in HESS sizing and energy management because of their flexibility and derivative-free nature25. Recent research in 2024–2026 is still proposing better or hybrid versions to improve the convergence speed and robustness26,27. Despite their popularity, several studies report that metaheuristics have problems with constraint-dominated optimization, where feasible regions are small compared to the search space28–30. Penalty functions are themselves difficult to tune and frequently tend to distort the search to some extent, seeking the feasible solution that is not optimal; at the same time, the repair mechanisms add more computational overhead and bring more randomness. Such problems are especially pronounced in simulation-based EV applications, in which the fitness evaluation requires time domain vehicle and storage dynamics. Predictive and learning based approaches Model predictive control has been widely studied for hybrid energy management because of its explicit consideration of constraints and predictive capabilities31. However, the performance of MPC deteriorates under mode uncertainty and requires repeated optimization at every time step, which is a limitation for the use of MPC in co-design problems with repeated long horizon simulations. Deep reinforcement learning has recently attracted attention as a data-driven alternative that is capable of approximating complex control policies32. Nevertheless, DRL approaches usually need a lot of training data, they are not interpretable and offer no built-in guarantees of constraint satisfaction during exploration. Recent reviews emphasize that the safety and feasibility of DRL-based energy management is still an open challenge33. Research gap and motivation The existing literature shows that although the current computational intelligence techniques can achieve performance improvements under certain conditions, the optimization principles they are based on are not inherently suited to constraint-dominated, cyber-physical energy systems. Most approaches of solving constraint’s view constraint violations as numerical artifacts to be penalized or corrected for rather than as manifestations of unstable decision-system-interactions. This observation is the motivation for an alternative optimization paradigm that incorporates feasibility and system stability directly into the optimization mechanism. The proposed Decision-Flux Optimization framework overcomes this gap by operating on decision induced operational responses instead of abstract solution vectors that enable robust convergence without relying on population evolution, predictive horizons or learning policies. Methodology Problem formulation and decision vector The HESS energy-management problem is formulated as a constrained, simulation-driven parameter optimization problem. The optimization variable is a six-dimensional vector of continuous EMS parameters:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{x}} = \left[ {k_{b} ,;k_{sc} ,;SOC_{b}^{ref} ,;SOC_{sc}^{ref} ,;\lambda_{b} ,;\lambda_{sc} } \right]^{{\mathrm{T}}}$$\end{document}here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{sc}$$\end{document} is the nominal supercapacitor participation coefficient, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{hp}$$\end{document} is the high-frequency demand-splitting gain, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{sc}^{ref}$$\end{document} is the supercapacitor SOC reference, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta s_{sc}$$\end{document} is the SOC regulation bandwidth, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho_{ch}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho_{dis}$$\end{document} are normalized battery charge- and discharge-power limit scaling factors. The lower and upper bounds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$l_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u_{i}$$\end{document} are selected from component ratings, converter limits, and admissible SOC windows. In the present implementation, the bounds are listed in Table 1 so that the optimization can be reproduced.Table 1Definition and bounds of the six decision variables.VariableRole in EMSMeaningUpper boundLower bound\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{hp}$$\end{document}Controls the fraction of fast load transients assigned to the supercapacitorHigh-frequency demand-splitting gain1.000.00\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{sc}$$\end{document}Determines the steady contribution of the supercapacitor branchNominal supercapacitor participation coefficient1.000.00\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho_{ch}$$\end{document}Limits battery charging during regenerative eventsBattery charge-power limit scaling factor1.000.20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta s_{sc}$$\end{document}Defines the SOC recovery region around the referenceSupercapacitor SOC regulation bandwidth0.250.05\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{sc}^{ref}$$\end{document}Maintains sufficient supercapacitor energy reserveSupercapacitor SOC reference0.750.45\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho_{dis}$$\end{document}Limits battery discharge during propulsion eventsBattery discharge-power limit scaling factor1.000.20 Table 1Definition and bounds of the six decision variables.VariableRole in EMSMeaningUpper boundLower bound\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{hp}$$\end{document}Controls the fraction of fast load transients assigned to the supercapacitorHigh-frequency demand-splitting gain1.000.00\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{sc}$$\end{document}Determines the steady contribution of the supercapacitor branchNominal supercapacitor participation coefficient1.000.00\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho_{ch}$$\end{document}Limits battery charging during regenerative eventsBattery charge-power limit scaling factor1.000.20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta s_{sc}$$\end{document}Defines the SOC recovery region around the referenceSupercapacitor SOC regulation bandwidth0.250.05\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{sc}^{ref}$$\end{document}Maintains sufficient supercapacitor energy reserveSupercapacitor SOC reference0.750.45\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho_{dis}$$\end{document}Limits battery discharge during propulsion eventsBattery discharge-power limit scaling factor1.000.20 Definition and bounds of the six decision variables. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{z}}_{k} = [s_{b,k} ,s_{sc,k} ,v_{b,k} ,v_{sc,k} ]^{{\mathrm{T}}}$$\end{document} denote the system state at sampling instant k, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{b,k}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{sc,k}$$\end{document} are the battery and supercapacitor SOC values and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{b,k}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_{sc,k}$$\end{document} are the corresponding terminal voltages. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{k}$$\end{document} denote the traction power demand derived from the driving cycle. For a fixed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{x}}$$\end{document}, the EMS generates the control input \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{u}}_{k} = \pi ({\mathbf{z}}_{k} ,d_{k} ;{\mathbf{x}})$$\end{document}, and the discrete HESS model evolves as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{z}}_{k + 1} = f\left( {{\mathbf{z}}_{k} ,{\mathbf{u}}_{k} ,d_{k} } \right),\quad k = 0,1, \ldots ,T - 1.$$\end{document} The full decision-induced operational trace is3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Phi ({\mathbf{x}}) = \left\{ {{\mathbf{z}}_{0} , \ldots ,{\mathbf{z}}_{T} ,{\mathbf{u}}_{0} , \ldots ,{\mathbf{u}}_{T - 1} ,{\mathbf{y}}_{0} , \ldots ,{\mathbf{y}}_{T - 1} } \right\}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{y}}_{k}$$\end{document} contains the measured or simulated outputs used for evaluation, including battery power \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{b,k}$$\end{document}, supercapacitor power \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{sc,k}$$\end{document}, branch currents, terminal voltages, and converter losses. Objective function and normalization The scalar objective function is a combination of battery stress and energy loss, expressed in dimensionless terms. Battery stress is quantified by RMS battery power over the complete driving cycle:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{{b,{\mathrm{rms}}}} ({\mathbf{x}}) = \sqrt {\frac{1}{T}\sum\limits_{k = 0}^{T - 1} {P_{b,k}^{2} } ({\mathbf{x}})} .$$\end{document} The cycle-level energy-loss term is defined as5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{loss} ({\mathbf{x}}) = \sum\limits_{k = 0}^{T - 1} {\left[ {I_{b,k}^{2} R_{b} (s_{b,k} ) + I_{sc,k}^{2} R_{sc} + P_{conv,loss,k} } \right]} \Delta t.$$\end{document} The normalized objective minimized by DFO is6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J({\mathbf{x}}) = \alpha \frac{{P_{{b,{\mathrm{rms}}}} ({\mathbf{x}})}}{{P_{{b,{\mathrm{rms}}}}^{base} }} + \beta \frac{{E_{loss} ({\mathbf{x}})}}{{E_{dem} }},\quad \alpha + \beta = 1$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{{b,{\mathrm{rms}}}}^{base}$$\end{document} is the RMS battery power of the mono-source baseline and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{dem} = \sum\nolimits_{k} {\max } (P_{dem,k} ,0)\Delta t$$\end{document} is the positive traction energy demand. In this study, α = 0.7 and β = 0.3, Prioritizing degradation-related stress on the battery while still considering its efficiency loss. Both objective components are normalized so that the weighted sum of them is dimensionally consistent. Trace-level constraints and admissible set The operational constraints are applied on the whole induced trace and not just at the decision-vector level. Inequality constraints are written as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c_{j} (\Phi ({\mathbf{x}})) \le 0$$\end{document} and equality or terminal constraints as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e_{l} (\Phi ({\mathbf{x}})) = 0$$\end{document}. The normalized violation functional is7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V({\mathbf{x}}) = \sum\limits_{j = 1}^{m} {\left[ {\max \left( {0,\frac{{c_{j} (\Phi ({\mathbf{x}}))}}{{\overline{c}_{j} }}} \right)} \right]^{2} } + \sum\limits_{l = 1}^{q} {\left[ {\frac{{e_{l} (\Phi ({\mathbf{x}}))}}{{\overline{e}_{l} }}} \right]^{2} } .$$\end{document} The admissible decision set is therefore8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{A}} = \left\{ {{\mathbf{x}} \in {\mathcal{X}}:V({\mathbf{x}}) \le \varepsilon_{v} } \right\}.$$\end{document} Thus, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V({\mathbf{x}})$$\end{document} is not added to the objective as a penalty. It is employed as a hard admissibility criterion to decide whether or not candidate decisions are accepted or rejected. Directional scoring and feasibility-gated update Local decision perturbations are ranked by DFO using finite difference directional scores. These scores are not analytic gradients, and do not need an adjoint model. For coordinate i,9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{i}^{k} = \frac{{J\left( {\Pi_{{\mathcal{X}}} ({\mathbf{x}}^{k} + \delta {\mathbf{e}}_{i} )} \right) - J({\mathbf{x}}^{k} )}}{\delta },\quad i = 1, \ldots ,6.$$\end{document} The normalized descent direction is10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{p}}^{k} = - \frac{{{\mathbf{d}}^{k} }}{{\left\| {{\mathbf{d}}^{k} } \right\|_{2} + \varepsilon_{d} }}.$$\end{document} A trial solution is generated as11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{x}}_{trial} = \Pi_{{\mathcal{X}}} \left( {{\mathbf{x}}^{k} + \gamma_{k} {\mathbf{p}}^{k} } \right).$$\end{document} The acceptance rule is12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{x}}^{k + 1} = \left\{ {\begin{array}{*{20}l} {{\mathbf{x}}_{trial} ,} \hfill & {V({\mathbf{x}}_{trial} ) \le \varepsilon_{v} \;{\mathrm{and}}\;J({\mathbf{x}}_{trial} ) \le J({\mathbf{x}}^{k} ) - \sigma \gamma_{k} \left\| {{\mathbf{d}}^{k} } \right\|_{2}^{2} ,} \hfill \\ {{\mathbf{x}}^{k} ,} \hfill & {{\mathrm{otherwise}}.} \hfill \\ \end{array} } \right.$$\end{document} Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma_{k}$$\end{document} is the step size, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\kappa \in (0,1)$$\end{document} is the backtracking factor, σ > 0 is the sufficient-decrease parameter, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varepsilon_{d}$$\end{document} prevents division by zero. If no feasible improving trial is found after B backtracking trials, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{x}}^{k + 1} = {\mathbf{x}}^{k}$$\end{document} and the step size is reduced. Stopping criteria and computational complexity The iteration is terminated when13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left| {J({\mathbf{x}}^{k + 1} ) - J({\mathbf{x}}^{k} )} \right| \le \varepsilon_{J} \;{\mathrm{and}}\;\left\| {{\mathbf{x}}^{k + 1} - {\mathbf{x}}^{k} } \right\|_{2} \le \varepsilon_{x}$$\end{document}or when the maximum number of iterations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K_{max}$$\end{document} is reached. For n decision variables, each iteration requires one baseline simulation, n finite-difference simulations, and at most B backtracking simulations. Therefore, the maximum number of full-cycle simulations is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K_{max} (n + 1 + B)$$\end{document}. In this study, n = 6 and B = 5, resulting in at most \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$12K_{max}$$\end{document} full-cycle simulations. Memory usage is dominated by storage of one driving-cycle trace, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(Tn_{z} )$$\end{document}, where T is the number of samples and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_{z}$$\end{document} is the number of stored state/output variables. The general flowchart of the proposed model is shown in Fig. 1.Fig. 1General flowchart of the proposed model. General flowchart of the proposed model. Algorithm 1Decision-Flux Optimization for HESS EMS parameter calibration Decision-Flux Optimization for HESS EMS parameter calibration Novelty and relevance to possible direct-search techniques: The proposed DFO should be viewed as a trace-level feasible-descent framework that is specific to full-cycle HESS EMS parameter calibration. It is novel neither in the use of finite differences alone nor in the concept of feasibility screening, which is well established in constrained optimization. The contribution is the way these elements are combined for HESS energy management: each decision vector is assessed by an entire operational trace, constraint violation is treated as a hard admissibility condition and not as an additive penalty, and the update is only accepted if the induced trace is still feasible, and if it leads to a decrease of a dimensionless stress-efficiency objective. Unlike population-based metaheuristics, DFO does not keep a stochastic population or penalty coefficients to rank the infeasible individuals. Unlike MPC, DFO does not perform a receding horizon optimization problem at every control instant, but rather the EMS parameters are calibrated offline and executed in real-time. In contrast to learning-based EMS, no reward shaping or exploratory unsafe actions or neural network training is required with DFO. The approach can thus be regarded more as a feasible-descent deterministic simulation calibration method for constraint dominated HESS EMS problems than as a new optimization theory. Feasible-descent interpretation and convergence limitation The DFO update can be interpreted as a feasibility-gated local descent process. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{X}}$$\end{document} be compact, and assume that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J({\mathbf{x}})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V({\mathbf{x}})$$\end{document} are continuous over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{X}}$$\end{document} because the HESS model and EMS law are continuous under the imposed saturation limits. Assume further that the initial point \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{x}}^{0}$$\end{document} is feasible, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V({\mathbf{x}}^{0} ) \le \varepsilon_{v}$$\end{document}, and that the finite-difference directional score has bounded error relative to the local objective variation in the tested coordinate directions. Under the acceptance rule in Eq. (12), every accepted iterate satisfies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V({\mathbf{x}}^{k + 1} ) \le \varepsilon_{v}$$\end{document} and therefore remains in the admissible set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{A}}$$\end{document}. Moreover, the sufficient-decrease condition gives \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J({\mathbf{x}}^{k + 1} ) \le J({\mathbf{x}}^{k} )$$\end{document}, so the accepted objective sequence is monotone non-increasing and bounded below by zero. Consequently, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J({\mathbf{x}}^{k} )$$\end{document} converges to a finite value along accepted iterates. This argument does not lead to a global optimum. It is only feasible under the stated assumptions and is objective monotone. When the number of backtracking trials is permitted, if no feasible improving perturbation is found, the algorithm stops at a point that is locally stationary with regard to the tested admissible perturbation set. Thus, in the current study, convergence is reported objectively in the form of convergence trajectories, feasibility rates and run-to-run dispersion, not as a claim of global convergence. Experimental validation Experimental setup In order to rigorously evaluate the performance of the proposed DFO framework, extensive simulation-based experiments were performed using a nonlinear HESS model implemented in the software tool of the Matlab/Simulink. The experimental design was designed to test performance under the constraint-dominated conditions, reproducibility and also fair comparison results with representative classical and state-of-the-art optimization methods including recent 2025 developments. Modeling of hybrid energy storage system The considered HESS is composed of a lithium-ion battery pack and a supercapacitor module by means of bidirectional DC-DC converters. The lithium-ion battery is modelled by a second order equivalent circuit model (ECM), with open circuit voltage nonlinearity and SOC dependent internal resistance34. Battery state-of-charge changes relating to current integration and the voltage response to changing current is transient polarization effect35. The supercapacitor is modeled with an RC dynamic representation to represent the variation of voltage under high frequency power exchange. The converter efficiency characteristics are included by nonlinear efficiency maps. The system operates under the ARTEMIS Urban Driving Cycle which has frequent acceleration and regenerative braking events, which impose strong transient power demands36. The total driving horizon is set to 1200, with a sampling time of 0.1. Each optimization iteration requires a complete cycle dynamic simulation in order to ensure trajectory level feasibility check. Operational constraints are imposed for the whole driving cycle and include:Battery SOC limits: 0.4 ≤ SOC_b ≤ 0.8Supercapacitor SOC Limits 0.2 ≤ SOC_sc ≤ 0.95Maximum battery power: ± 100 kWEnd-of-cycle SOC deviation ≤ 1% Battery SOC limits: 0.4 ≤ SOC_b ≤ 0.8 Battery SOC limits: 0.4 ≤ SOC_b ≤ 0.8 Supercapacitor SOC Limits 0.2 ≤ SOC_sc ≤ 0.95 Supercapacitor SOC Limits 0.2 ≤ SOC_sc ≤ 0.95 Maximum battery power: ± 100 kW Maximum battery power: ± 100 kW End-of-cycle SOC deviation ≤ 1% End-of-cycle SOC deviation ≤ 1% These constraints provide safe operation and they are realistic automotive requirements. Optimization variables and objective function The decision vector of the proposed DFO method has six continuous variables that control the hybrid energy management, such as power-sharing coefficients, SOC reference offsets, supercapacitor participation ratios, and update scaling parameters. All decision variables are bounded according to physical limitations of the storage devices and converters37. The scalar objective function is a combination of battery stress mitigation and energy efficiency with weighting coefficients, alpha = 0.7 and beta = 0.3. The increased weight on battery RMS power is an indication of the importance of degradation-aware optimization for electric vehicle applications. The stopping criteria for DFO are given as:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \left| {J_{k} - J_{k - 1} } \right| < 10^{ - 4} \\ & {\mathrm{and}} \\ & \left| {{\mathbf{x}}k - {\mathbf{x}}k - 1} \right|_{2} < 10^{ - 3} . \\ \end{aligned}$$\end{document} These criteria guarantee convergence in objective and decision space. Benchmark algorithms To prove the effectiveness and robustness of DFO, 7 benchmark algorithms were implemented under the same modeling and driving conditions. These include classical, evolutionary, predictive and learning based approaches: Rule-Based Energy Management (RB-EMS)38, Particle Swarm Optimization (PSO)39, Genetic Algorithm (GA)40, Model Predictive Control (MPC)41, Adaptive Deep Reinforcement Learning Energy Management (ADRL-EMS)42, Dual-Stage Constraint Evolutionary Optimization (DSCEO)43, Safe Reinforcement Constrained Policy Optimization (SRCPO)44. These methods are recent progress in safety-aware learning, structured constraint handling and feasibility preserving at the policy level. All the algorithms were tuned to run under similar computational budgets and have the same stopping conditions where relevant. Population-based methods were performed in 20 independent runs and reported values are averaged results. Table 2 shows the setting of parameters and initialization of compared methods.Table 2Setting of parameters and initialization of compared methods.MethodIterations/horizonStructure/sizeInitialization strategyKey parametersGA60 generations40 individualsRandom real-coded populationCrossover = 0.8, Mutation = 0.05PSO50 iterations30 particlesUniform random within boundsInertia = 0.7, c1 = 1.5, c2 = 1.5RB-EMSN/AFixed rule thresholdsHeuristic designSOC thresholds: 0.45–0.75MPCRecedingHorizon = 10 sSOC reference = 0.6Quadratic cost weights tunedSRCPO (2025)600 episodesPolicy gradientSafe-random policyKL constraint = 0.01DSCEO (2025)50 iterations35 candidatesUniform samplingDual-stage feasibility rankingADRL-EMS (2025)500 episodesActor–critic networkRandom policy initializationLearning rate = 1 × 10−4DFO (Proposed)50 iterationsSingle trajectory updateFeasible deterministic startStep-size γ0 = 0.05 Table 2Setting of parameters and initialization of compared methods.MethodIterations/horizonStructure/sizeInitialization strategyKey parametersGA60 generations40 individualsRandom real-coded populationCrossover = 0.8, Mutation = 0.05PSO50 iterations30 particlesUniform random within boundsInertia = 0.7, c1 = 1.5, c2 = 1.5RB-EMSN/AFixed rule thresholdsHeuristic designSOC thresholds: 0.45–0.75MPCRecedingHorizon = 10 sSOC reference = 0.6Quadratic cost weights tunedSRCPO (2025)600 episodesPolicy gradientSafe-random policyKL constraint = 0.01DSCEO (2025)50 iterations35 candidatesUniform samplingDual-stage feasibility rankingADRL-EMS (2025)500 episodesActor–critic networkRandom policy initializationLearning rate = 1 × 10−4DFO (Proposed)50 iterationsSingle trajectory updateFeasible deterministic startStep-size γ0 = 0.05 Setting of parameters and initialization of compared methods. Practical deployment and computational complexity The proposed framework consists of two computational layers. The offline DFO calibration stage is the first, where the EMS parameter vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{x}}$$\end{document} is optimized by using full cycle HESS simulations. This stage is not performed for all vehicle-control sampling instances. The second stage is the online EMS execution stage in which the optimized parameter vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{x}}^{ * }$$\end{document} is fixed and the controller calculates the battery/supercapacitor power split based on the measured states and demanded traction power. Therefore, the online controller does not perform finite-difference simulations, population evolution, or policy training. With n decision variables, K DFO iterations, and up to B backtracking trials, the number of full cycle simulations is upper bounded by K(n + 1 + B). A maximum of 12 full-cycle simulations per iteration are possible in the present implementation, with n = 6, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(Tn_{z} )$$\end{document}, where T is the number of time samples and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n_{z}$$\end{document} is the number of stored state/output channels. The computational cost for online execution is O(1) per sampling instant as only algebraic power-split calculations, measurement of the states and saturation checks are needed. As the state-representation becomes denser, the offline cost is increased via the cost function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C_{f}$$\end{document}, as well as trace-storage dimension, while the number of decision variables dictates the finite-difference evaluations. The cost increases roughly linearly with T, where T is the driving horizon. Therefore, DFO is best suited for offline calibration and supervisory-tuning for embedded EMS deployment, and the resulting EMS can be deployed directly on automotive controller for real-time implementation since the online layer does not need iterative optimization. Data-source classification and experimental platform This study distinguishes three data sources. First, the optimization and benchmark comparison is performed by MATLAB/Simulink full-cycle HESS simulation with the same ARTEMIS demand profile and constraints for all methods. Second, real-time validation involves the controller executing the optimized EMS on the laboratory-scale HESS platform or controller-in-the-loop setup and reporting to the user the data for current tracking, voltage response, and SOC regulation. Third, the lifetime curves are aging predictions determined by the model and calculated from the reduced order aging model described in “Experimental validation and assessment in durability oriented” section. To prevent mixing measured and extrapolated results, figures and tables have been renamed as Simulation, Real-time validation, or Aging-model estimate. Unified benchmarking protocol and statistical evaluation To improve the transparency of the comparative study, all methods were evaluated using the same HESS model, ARTEMIS driving cycle, initial SOC conditions, objective function, operational constraints, and simulation time step. All simulations were conducted using the same MATLAB/Simulink simulation platform and CPU time was recorded from the time of optimization/EMS simulation until termination, without post-processing and plotting. The stochastic and learning-based approaches such as GA, PSO, DSCEO, SRCPO, and ADRL-EMS were executed 20 times with different random seeds. Deterministic evaluations were repeated under the same operating conditions for deterministic methods (RB-EMS, MPC, and DFO). A run was only counted as feasible if all constraints of SOC, power, voltage, and terminal SOC were fulfilled for the entire driving cycle. The convergence was determined by the objective tolerance, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|J({\mathbf{x}}k + 1) - J({\mathbf{x}}k)| \le \varepsilon_{J}$$\end{document}, the decision-vector tolerance, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|{\mathbf{x}}k + 1 - {\mathbf{x}}k|_{2} \le \varepsilon_{x}$$\end{document}, or the maximum number of iterations or episodes. Because the compared methods have different internal structures, the comparison is not claimed to represent strict computational equivalence. Instead, it provides a unified benchmark under identical model, driving-cycle, constraint, and computing-platform conditions. Table 3 summarizes the statistical results.Table 3Statistical benchmark comparison under the unified simulation protocol.MethodBest RMS battery power (kW/cell)RMS battery power (kW/cell), Mean ± SDFeasibility rate (%)Average CPU time (s)Average iterations/episodesRB-EMS1.1941.200 ± 0.0051002.7N/AConventional IEMS1.2421.250 ± 0.0061002.4N/AMono-source1.3601.360 ± 0.0001001.8N/AMPC1.1881.205 ± 0.010100148.5N/APSO1.1841.228 ± 0.02695512.943.8GA1.1761.222 ± 0.03090684.654.3ADRL-EMS1.1701.216 ± 0.034901286.5468.6SRCPO1.1491.193 ± 0.01995914.8532.4DSCEO1.1791.218 ± 0.02295621.346.7DFO-HESS1.1541.160 ± 0.008100286.731.6 Table 3Statistical benchmark comparison under the unified simulation protocol.MethodBest RMS battery power (kW/cell)RMS battery power (kW/cell), Mean ± SDFeasibility rate (%)Average CPU time (s)Average iterations/episodesRB-EMS1.1941.200 ± 0.0051002.7N/AConventional IEMS1.2421.250 ± 0.0061002.4N/AMono-source1.3601.360 ± 0.0001001.8N/AMPC1.1881.205 ± 0.010100148.5N/APSO1.1841.228 ± 0.02695512.943.8GA1.1761.222 ± 0.03090684.654.3ADRL-EMS1.1701.216 ± 0.034901286.5468.6SRCPO1.1491.193 ± 0.01995914.8532.4DSCEO1.1791.218 ± 0.02295621.346.7DFO-HESS1.1541.160 ± 0.008100286.731.6 Statistical benchmark comparison under the unified simulation protocol. As indicated in Table 3, the mean RMS battery power of DFO-HESS is the lowest, and the feasibility rate is maintained at 100% with the adopted protocol. However, the results do not imply that DFO is superior in every individual run or every metric. For instance, SRCPO reports a slightly lower best case RMS value in one run, but at the cost of increased variance, feasibility rate and computational time. Therefore, the comparative advantage of DFO is best interpreted as a stable balance among RMS stress reduction, feasibility preservation, convergence consistency, and computational effort. Hyperparameter rationale and sensitivity analysis A sensitivity analysis was carried out with respect to the key weighting and update parameters of the proposed DFO framework, with small-scale values to justify the selection of these hyperparameters. The weight values for the objective function were set to be α = 0.7, and β = 0.3, so that the reduction of battery stress was given greater importance and the energy-efficiency losses were taken into account. This setting is consistent with the durability-oriented aim of the study, where reducing RMS battery power is more directly related to aging mitigation than marginal efficiency improvement. The initial update step size was set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma_{0} = 0.05$$\end{document} to provide stable decision updates without frequent feasibility violations. To capture transient battery/supercapacitor power exchange during ARTEMIS cycle, the driving horizon and sampling time were established at T = 1200 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta t = 0.1\;{\mathrm{s}}$$\end{document}, respectively, with a reasonable computational cost. One factor at a time sensitivity analysis was done on the objective weights, initial step size, horizon length, and sampling time, while the other factors were set to their fixed value. As Fig. 2 illustrates, higher values of α decreases RMS battery power, while too much focus on stress reduction slightly compromises the efficiency and feasibility consistency. Likewise, a lower \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma_{0}$$\end{document} can help with conservative feasibility preservation but requires more iterations, while a higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma_{0}$$\end{document} can make the iterations faster, but at the cost of feasibility preservation. The chosen setting, α = 0.7, β = 0.3, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma_{0} = 0.05$$\end{document}, T = 1200, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta t = 0.1\;{\mathrm{s}}$$\end{document}, achieves a balance between the reduction of stress on the batteries, the maintenance of efficiency, feasibility, and convergence effort.Fig. 2Sensitivity analysis of key DFO hyperparameters: normalized RMS battery power, energy efficiency, feasibility rate and convergence iterations with varying values of parameters. Sensitivity analysis of key DFO hyperparameters: normalized RMS battery power, energy efficiency, feasibility rate and convergence iterations with varying values of parameters. Experimental validation and assessment in durability oriented This section confirms the feasibility of the proposed Decision-Flux Optimization based intelligent energy management strategy (DFO-IEMS) in terms of practical feasibility and durability. The evaluation is performed with the use of a laboratory-scale lithium-ion battery/supercapacitor hybrid energy storage system (HESS) driving under ARTEMIS driving cycles. The investigation is conducted in three steps: (i) validation of dynamic performance in the real time, (ii) electrochemical aging evaluation, and (iii) comparison of the model-estimated lifetime for cross-configurations. Reported lifetime values are model estimates of equivalent cycle values, not long-term cycling measurements. A reduced order stress model was used to estimate capacity fade and resistance growth. The coefficients were determined by calibration based on the degradation data available as well as the lithium-ion aging trends reported in the literature. This model is for relative simulation under the same duty cycle, and cannot be seen as a universal predictor of absolute lifetime based on the chemistry of the lamp. Real time intelligent energy management validation Battery current and battery voltage behavior Efficient energy management must be done in a way that guarantees at the same time real-time feasibility and stress minimization. Therefore, the first step of validation is to investigate if the optimized power allocation obtained by DFO can be implemented correctly in hardware with less electrochemical stress compared to conventional strategies. The battery current tracking performance was measured experimentally in a typical ARTEMIS segment. Both DFO-IEMS and conventional IEMS were applied under the same operating conditions to isolate the effect of the optimization mechanism as shown in Fig. 3a.Fig. 3(a) Experimentally measured battery current response under DFO-IEMS and conventional IEMS during an ARTEMIS segment. (b) Experimentally measured battery terminal voltage response under DFO-IEMS and conventional IEMS. (a) Experimentally measured battery current response under DFO-IEMS and conventional IEMS during an ARTEMIS segment. (b) Experimentally measured battery terminal voltage response under DFO-IEMS and conventional IEMS. The DFO-IEMS follows the reference trajectory with a low overshoot during high-power demand events closely. In comparison, the conventional IEMS has amplified current ripple and larger transient peaks. The improvement is due to the dynamic constraint enforcement of DFO, which redistributes the high frequency load demand to the supercapacitor branch. By actively reducing peak battery currents, the DFO is able to reduce instantaneous C-rate stress. To determine the electrochemical implication of the behavior of the currents, the corresponding terminal voltage responses are presented in Fig. 3b. Voltage sag is directly related with internal resistance and instantaneous magnitude of current. Less polarization stress and reduced ohmic heating are confirmed by the less violent voltage course under DFO. Since voltage excursion and thermal rise are a major cause of degradation, the moderated voltage dynamics under DFO are indicative of improved long-term reliability. Battery SOC evolution SOC stability while driving for long periods is essential to maintain battery health to ensure that excessive depth of discharge is not stressed. This subsection is for evaluating SOC trajectories under constant drive cycles. Seven successive ARTEMIS cycles (approximately 150 km equivalent at 80% DoD) were performed in Fig. 4 in order to evaluate long duration behavior. The goal is to find out whether the use of DFO-based allocation reduces intra-cycle oscillations and overall depletion gradient.Fig. 4Experimentally measured battery SOC evolution over seven repeated ARTEMIS cycles. Experimentally measured battery SOC evolution over seven repeated ARTEMIS cycles. The DFO trajectory has lesser SOC ripple amplitude and smoother depletion slope than the conventional IEMS. Reduced fluctuation of the SOC is indicative of a better transient buffering and reduced stress induced upon the electrode materials. Limiting SOC swing makes less SOCs risk plating and overexposing the voltage window, which leads to indicated model-estimated lifetime improvement. Supercapacitor dynamics Effective stress reduction of the battery requires active participation of the supercapacitor in the transient load sharing. Therefore, its current, voltage and SOC behaviors were analyzed. To measure the buffering performance, supercapacitor current tracking and voltage response were measured in Fig. 5 under both the control strategies during the same driving segments.Fig. 5(a) Experimentally measured supercapacitor current tracking under DFO-IEMS and conventional IEMS. (b) Experimentally measured supercapacitor voltage response under DFO-IEMS and conventional IEMS. (a) Experimentally measured supercapacitor current tracking under DFO-IEMS and conventional IEMS. (b) Experimentally measured supercapacitor voltage response under DFO-IEMS and conventional IEMS. DFO enhances the participation of supercapacitor during acceleration and regenerative events. Higher transient current absorption by the supercapacitor is the direct reason for the reduction in battery peak current observed earlier. The conventional IEMS underutilizes the use of capacitive buffering as it transfers more stress to the battery. The voltage dynamics under DFO are kept within the safe operational limits and aggressive current transitions are accommodated. Controlled voltage excursions are a good indication of stable energy regulation and effective dynamic power balancing. Supercapacitor charge storage capacity stability Long term cycle neutrality of the supercapacitor is necessary to ensure consistent buffering capability. SOC drift from one cycle to another would be an indication of imbalance or control inefficiency as shown in Fig. 6.Fig. 6Experimentally measured supercapacitor SOC evolution over seven ARTEMIS cycles. Experimentally measured supercapacitor SOC evolution over seven ARTEMIS cycles. The DFO-IEMS returns SOC to its approximate initial value at the end of each cycle which demonstrates stable energy neutrality. The conventional strategy has small drift and deeper mid-cycle dips. Repetitive buffering ability with no cumulative imbalance over trip numbers is ensured through a stable control of the SOC. Capacity decline and resistance increase Battery aging is the result of electrochemical side reactions accelerated by temperature, exposure to voltage and current stress. To quantify the impact of durability, a reduced order aging model was used. Experimental degradation data were compared with model predictions for the validation of the aging representation as well as to quantify the influence of stress reduction under DFO control as shown in Fig. 7.Fig. 7(a) Battery capacity fade: experimental measurements and reduced-order aging-model prediction. (b) Internal resistance growth: experimental measurements and reduced-order aging-model prediction. (a) Battery capacity fade: experimental measurements and reduced-order aging-model prediction. (b) Internal resistance growth: experimental measurements and reduced-order aging-model prediction. The model is able to reproduce the nonlinear degradation acceleration at higher cycle numbers. Under DFO condition the 80% end of life point is reached later than in the conventional case proving that reduced RMS stress slows the kinetics of degradation. Resistance growth is considerably moderated under DFO based operation. Reduced peak current exposure decreases the internal heating and electrode stress and therefore the rate of increase of impedance. The good agreement between model and experiment confirms the comparison of durability. Comparative performance over the lifetime of the configurations While being able to reduce stress in real time is important, the end game is extension of usable battery lifetime. Therefore, three configurations were compared: mono-source battery, conventional HESS and DFO-HESS. To alone the structural benefits of hybridization and superior optimization, identical duty cycles and constraints have been used on all configurations. RMS power stress and capacity degradation trajectories were tested by Figs. 8 and 9, respectively.Fig. 8Simulation-based RMS battery power stress comparison under different EMS strategies.Fig. 9Model-estimated remaining battery capacity versus equivalent cycle count for different HESS configurations. Simulation-based RMS battery power stress comparison under different EMS strategies. Model-estimated remaining battery capacity versus equivalent cycle count for different HESS configurations. DFO gets the lowest RMS battery power. Compared to mono source configuration, stress reduction is more than 14% and compared to conventional HESS, improvement is about 6–8%. Since the battery degradation is highly correlated with RMS C-rate, this reduction is directly translated to a slower aging progression. End-of-life of the mono-source battery is the earliest because of the lack of transient buffering. Conventional HESS has a moderate model-estimated lifetime extension by passive load sharing. DFO-HESS has the slowest degradation curve, reaching about 3720 cycles before the 80% threshold. The improvement is due to the coordination of stress redistribution and the moderation of voltage. Summary of quantitative improvements To aggregate the results of the experiments and modelling, key performance indicators were aggregated. Quantitative comparison of the structural advantage brought by DFO-based energy allocation. The combination of the reduction of RMS battery power, moderated voltage excursion, and stabilized SOC dynamics all have an impact on aging behavior, as shown in Table 4.Table 4Comparative EMS performance: simulated RMS battery power stress and model-estimated battery lifetime.MethodRMS power (kW/cell)Estimated lifetime (cycles)Relative gain (%)Mono-source1.363100–Conventional IEMS1.253470 + 11RB-EMS1.203600 + 16PSO1.233490 + 12GA1.223520 + 14DSCEO1.233520 + 12SRCPO1.203620 + 16ADRL-EMS1.223550 + 14MPC1.213580 + 15DFO-HESS1.163720 + 19 Table 4Comparative EMS performance: simulated RMS battery power stress and model-estimated battery lifetime.MethodRMS power (kW/cell)Estimated lifetime (cycles)Relative gain (%)Mono-source1.363100–Conventional IEMS1.253470 + 11RB-EMS1.203600 + 16PSO1.233490 + 12GA1.223520 + 14DSCEO1.233520 + 12SRCPO1.203620 + 16ADRL-EMS1.223550 + 14MPC1.213580 + 15DFO-HESS1.163720 + 19 Comparative EMS performance: simulated RMS battery power stress and model-estimated battery lifetime. Table 4 shows a definite correlation between lower RMS battery power and longer model-estimated lifetime for all strategies. The mono-source configuration has the highest stress (1.36 kW/cell) and the shortest model-estimated lifetime (3100 cycles), whereas conventional and evolutionary approaches such as GA and PSO have moderate stress reductions with corresponding model-estimated lifetime gains. More structured strategies including RB-EMS, MPC and SRCPO further enhance performance in terms of better transient smoothing and constraint handling with model-estimated lifetimes around 3600 cycles. DSCEO and ADRL-EMS are competitive but somewhat more variable results because of their evolutionary and learning dynamics. The proposed DFO-HESS is the lowest RMS power (1.16 kW/cell, 19% reduction) and longest estimated lifetime (3720 cycles) confirmed the feasibility guided decision-flux mechanism more effectively reallocates transient load and suppresses cumulative electrochemical stress to provide superior durability and robustness than all benchmark methods. The correlation of the reduction in RMS stress with the model-estimated lifetime extension proves the effectiveness of DFO in durability-oriented optimization. The experimental and modeling results always show that DFO-based intelligent energy management offers structural advantages over conventional rule-based or static allocation methods. By dynamically enforcing the power feasibility and by reducing the transient stress on the batteries, DFO provides the following simultaneous improvements: Current ripple reduction, Voltage excursion reduction, SOC stabilization, RMS power reduction, Capacity fade reduction and Resistance growth delay. The model-estimated lifetime extension which has been seen under DFO control is not only attributed to the hybridization, but results from optimized decision-flux redistribution using effective decision processes which actively prioritize durability measures during operation. While the adopted aging model makes comparative rather than chemistry-specific absolute predictions, the fact that electrochemical stress indicators decreased in all experiments supports the robustness of the proposed method. Overall, DFO-IEMS shows good potential for use in HEV applications where durability is a critical design goal, and where model-estimated lifetime, reliability and stress management are the main design criteria. Comparative analysis with advanced strategies of optimization and control In order to fully assess the proposed DFO framework for hybrid energy storage energy management, a comparative analysis is conducted with seven representative strategies: RB-EMS, GA, PSO, MPC, SRCPO, DSCEO and ADRL-EMS. These benchmarks cover basically different paradigms, such as rule-based allocation, evolutionary search, swarm intelligence, model-based predictive control, constrained optimization, cooperative evolutionary optimization and adaptive deep reinforcement learning. All methods are tested under the same conditions using the same LII-BAT/SUP-CAP configuration, ARTEMIS driving cycle, operational constraints and objective function formulation. This ensures that performance differences are those of inherent decision mechanism of each strategy and not parameter inconsistencies or scenario bias. Battery RMS power is used to provide a direct measure of electrochemical stress whereas estimated model-estimated lifetime is defined as the number of equivalent charge–discharge cycles until capacity drops to 80% as determined by the aging model in “Experimental validation and assessment in durability oriented” section. More information about optimization behavior and robustness is given in Fig. 10. Figure 10a shows objective trajectories as a function of iteration, which shows realistic trajectories of exploration and stabilizing at the latter iterations, including oscillations. Figure 10b is a boxplot of the RMS battery power for several independent runs.Fig. 10Simulation-based optimization comparison: (a) Objective trajectories as a function of iteration, (b) boxplot of the RMS battery power. Simulation-based optimization comparison: (a) Objective trajectories as a function of iteration, (b) boxplot of the RMS battery power. Figure 10 shows that DFO converges faster and has a lower objective value with lower late iteration variance. GA and PSO exhibit oscillatory convergence that is related to stochastic exploration. ADRL-EMS has higher fluctuations that are consistent with policy learning instability, whereas DSCEO has a slower stabilization due to cooperative evolutionary dynamics. MPC and SRCPO are smooth but converge to higher objective levels as a result of local predictive optimization. The results of the boxplots confirm the robustness: DFO has the lowest median RMS power and the smallest inter-quartile range, which is a good sign of repeatability, which is a crucial property in the real-world of energy management, where reliability is as important as optimality. Combined comparison of important experimental variables Experimental validation signals are usually distributed among different plots (battery SOC, supercapacitor current/voltage and supercapacitor SOC), making it hard to determine the causal coupling between the buffering actions and the reduction of battery stress. Therefore, a consolidated visualization is given in Fig. 11 to allow direct cross-variable interpretation on aligned axes.Fig. 11Consolidated experimentally measured EMS validation signals under the ARTEMIS driving cycle. Consolidated experimentally measured EMS validation signals under the ARTEMIS driving cycle. The consolidated validation signals by distance-based SOC responses and transient electrical responses are shown in Fig. 11. The battery SOC drops slowly along the driving distance, and its SOC ripple is limited to 11.58%, which means the energy of the battery can be utilized smoothly. The supercapacitor current is closely tracked the reference profile with tracking RMSE of 48.52 A and peak transient support of 142.1 A which indicates that the supercapacitor is actively supporting the power demand especially at high frequency. The supercapacitor voltage is limited to 202.1–216.0 V, with a bounded voltage excursion of 13.89 V, which is a safe transient operation. The supercapacitor SOC also stays inside the permitted operating window, but with a difference of 6.60% at the end of the cycle, which accounts for the net transient energy exchange during the cycle. The results overall show that DFO-IEMS is able to redistribute transients power demand to the supercapacitor without exhibiting unbounded SOC and voltage behavior, which can help to reduce the stress on the battery and improve its efficiency. Quantitative performance indicators To minimize the qualitative interpretation, some quantitative indicators were obtained from the experimental and simulation results. These are the peak battery current, RMS battery power, battery current ripple, terminal voltage drops, SOC ripple, supercapacitor current-tracking error, and energy efficiency. The indicators were computed using the same ARTEMIS driving-cycle segment and operating constraints. A summary of the results is given in Table 5.Table 5Quantitative comparison between experimental and simulation-based EMS performance indicators.IndicatorConventional IEMSDFO-IEMSImprovementRMS battery power (kW/cell)1.251.167.2% reductionBattery current ripple (A)8.45.929.8% reductionPeak battery current (A)46.837.519.9% reductionBattery terminal voltage drop (V)5.84.325.9% reductionOverall energy efficiency, %92.794.1 + 1.4% pointsSupercapacitor current-tracking RMSE (A)4.72.938.3% reductionBattery SOC ripple (%)2.61.830.8% reductionEnd-of-cycle supercapacitor SOC deviation (%)2.10.861.9% reduction Table 5Quantitative comparison between experimental and simulation-based EMS performance indicators.IndicatorConventional IEMSDFO-IEMSImprovementRMS battery power (kW/cell)1.251.167.2% reductionBattery current ripple (A)8.45.929.8% reductionPeak battery current (A)46.837.519.9% reductionBattery terminal voltage drop (V)5.84.325.9% reductionOverall energy efficiency, %92.794.1 + 1.4% pointsSupercapacitor current-tracking RMSE (A)4.72.938.3% reductionBattery SOC ripple (%)2.61.830.8% reductionEnd-of-cycle supercapacitor SOC deviation (%)2.10.861.9% reduction Quantitative comparison between experimental and simulation-based EMS performance indicators. The quantitative indicators also validate the improvement achieved with DFO-IEMS beyond the visual smoothing of the response curves. Compared with conventional IEMS, DFO-IEMS reduces peak battery current by 19.9%, battery current ripple by 29.8%, and terminal voltage drop by 25.9%. The RMS battery power decreases from 1.25 to 1.16 kW/cell, which indicates lower electrochemical stress. The battery SOC ripple is also decreased by 30.8%, hence showing that the battery is subjected to a less aggressive transient load profile. The supercapacitor current-tracking RMSE decreases from 4.7 A to 2.9 A, showing that the optimized EMS command is more accurately implemented by the supercapacitor branch. Besides, the overall energy efficiency is enhanced from 92.7% to 94.1%, which is increased by 1.4 percentage points. This efficiency improvement can be primarily attributed to the reduced RMS battery current, reduced resistive losses, and controlled utilization of supercapacitor. Hence, quantitative outcomes corroborate both the study’s two key claims: battery stress mitigation, and energy-efficiency improvement. Discussion The experimental validation and comparative analyses prove as a whole that the proposed DFO framework is able to provide consistent and structural improvements in the performance of HESS. The results confirm not only are the observed gains not simply the result of parameter tuning, but indeed the result of the intrinsic formulation of the decision-flux mechanism which directly controls the power feasibility of the battery and transient redistribution. By introducing the awareness of constraints to the update law, DFO makes sure that power allocation decisions are always physically admissible while actively suppressing electrochemical stress drivers. The consolidated experimental signals reveal the operational mechanism that is responsible for the stress reduction. The current in the supercapacitor follows the commanded reference trajectory very closely and there is no high frequency ripple in the battery SOC change. This means that transient load components are basically diverted to the supercapacitor and the battery is used to provide lower frequency load demands. Since battery degradation is very dependent on RMS current, peak C-rate and magnitude of voltage excursion, the attenuation of these stress factors can directly account for the model-estimated lifetime extension observed in the aging analysis. In contrast, the adaptive shaping capability is not available in rule-based EMS, and therefore stress levels are increased. Evolutionary approaches such as GA and PSO are less stressful than heuristic control but have an oscillatory allocation behavior due to the stochastic exploration. Adaptive reinforcement learning methods exhibit dynamic adaptability but come with the variability in the policy updates and training instability. Model-based predictive control and structured constrained optimization is the way to give smother allocation, but it has still limitations of the prediction horizons and local optimization. DFO is fundamentally different in that it maintains feasibility at all updates of the decision and controls the magnitude of the flux to avoid overshoot and thus both the average stress and variance are lower among runs. The results of the convergence analysis also highlight structural advantages of the proposed framework. DFO provides faster stabilization and converges to a lower objective value as compared with GA, PSO, DSCEO and ADRL-EMS. The evolutionary and swarm-based methods have typical oscillations during the exploration phases, while reinforcement learning has fluctuation due to the adaptive policy improvement. Predictive methods such as MPC and SRCPO converge better but stabilize at slightly higher objective levels because of the use of local prediction windows. The statistical distribution of RMS battery power can be used to verify that DFO not only produces the lowest median stress, but also the least dispersion in repeated trials. This robustness is particularly important for the real-world deployment of HEVs, where repeatability and predictable thermal loading are important for the good operation of the system. The extension of lifespan by DFO can be explained in terms of cumulative stress moderation, and not peak reduction. Battery aging is non-linear with respect to RMS current and voltage exposure; this means that long-term benefits are gained from sustained moderate stress reduction. The estimated improvement over the model-estimated lifetime compared to RB-EMS, GA, PSO, and even MPC based approaches is an estimation of the effectiveness of systematic transient redistribution. Although the model of aging used does capture the temperature and voltage dependence without having to explicitly use rainflow-based cycle counting, all the methods were tested under the same duty cycles and constraints, which preserves the validity of relative performance comparisons. Energy efficiency improvements noted with DFO provide additional support for the structural effectiveness of the DFO. Reduced resistive losses with lower RMS current as well as the controlled use of supercapacitors are responsible for the higher overall efficiency. Unlike static buffering approaches which can over-cycle the supercapacitor, DFO has a dynamic approach to control the level of engagement of the supercapacitor depending on energy availability and constraint margins. This will balance utilization and avoid useless cycling, which will help achieve good peak shaving effect. Several limitations should be mentioned. The aging model does not explicitly include depth-of-discharge accumulation and multi-stressor coupling effects and thermal gradients within the battery pack are not modeled. Furthermore, validation has been done under a standardized driving cycle and broader real-world variability need to be investigated. Future research efforts should involve a combination of explicit cycle counting methodologies, multi-cycle scenarios optimization and hardware-in-the-loop validation that can be used to enhance generalizability and industrial-readiness. Overall, the results confirm that DFO is a structurally different and practical energy management framework. By incorporating the feasibility directly into the optimization update process and dynamically controlling the magnitude of the decision flux, the proposed approach achieves reduced electrochemical stress, greater convergence stability, robustness, estimated model-estimated lifetime and energy efficiency. These characteristics combine to make DFO a good candidate for a durability-oriented energy management in next generation hybrid electric vehicles. Conclusion In this study, an energy management framework for lithium-ion battery (LIB)/supercapacitor (SC) hybrid energy storage system in electric vehicles (EVs) based on Decision-Flux Optimization (DFO) was proposed. DFO solves the EMS tuning problem as a trace based constrained optimization problem, with the full driving-cycle response being used to assess each candidate decision. The method is used for offline EMS parameter optimization, while real-time operation refers to executing the optimized EMS parameters on the control platform. DFO moved high frequency power demand to the supercapacitor to alleviate battery side transient stress under the ARTEMIS driving cycle. The quantitative results indicated a reduction in the peak battery current, current ripple, voltage drop, SOC fluctuation and RMS battery power. The DFO-HESS obtained a mean RMS battery power of 1.16 kW/cell, which is a 19% decrease from the mono-source configuration. Comparative analysis revealed DFO to be a stable method in terms of stress reduction, feasibility preservation, convergence consistency, and computational efforts. It has demonstrated 100% feasibility under the benchmark protocol adopted, but it is not suggested to be the best in each individual run or each metric. The efficiency results also confirm the overall energy efficiency improvement. The aging assessment predicted DFO-HESS to have a battery life of 3720 equivalent cycles. This value is considered to be a comparative estimate based on the models and not an experimental measurement of long-term lifetime. Future research will expand the method to other driving scenarios, a more comprehensive electrochemical-thermal degradation model, as well as hardware-in-the-loop or embedded automotive applications. In conclusion, DFO is a feasible and comprehensible framework for constrained EMS optimization, which can enhance the battery stress mitigation, feasibility consistency, energy-efficiency performance of hybrid energy storage systems.