Energy Network Management System

• ENMS is a supervisory system that fuses network state, device models, forecasts, and policy objectives to generate actionable setpoints across heterogeneous energy assets.
• Its design spans centralized, hierarchical, distributed, and learned paradigms to manage multi-domain energy flows under stringent operational, economic, and reliability constraints.
• Advanced formulations leverage mixed-integer programming, convex optimization, and decentralized methods to enhance cost efficiency, performance, and asset health in dynamic energy networks.

Energy Network Management System (ENMS) denotes the supervisory layer that coordinates energy production, conversion, storage, consumption, and exchange across a network of heterogeneous assets under operational, economic, and reliability constraints. Across the literature, the term spans microgrid EMS and SCADA-based supervisory control, community-level coordination of multiple microgrids, transmission control-center workflows, wireless and mobile cyber-physical resource management, event-driven flexibility coordination in distribution grids, and operator-scale orchestration of routed-energy infrastructures (Pourbabak et al., 2017, Ju, 2021, Li et al., 2018, Birgersson et al., 9 Sep 2025). This breadth suggests that ENMS is best understood functionally: it is the decision system that fuses network state, device models, forecasts, and policy objectives into admissible setpoints, schedules, prices, or routing actions over one or more time scales.

1. Scope and semantic range

The literature does not restrict ENMS to a single physical domain. In the narrowest interpretation, it is the tertiary supervisory controller of a microgrid or hybrid powertrain. In the broadest interpretation, it is the operator-scale control plane of a software-defined energy network. Between these extremes, ENMS appears wherever energy flows among multiple controllable entities must be coordinated under shared constraints.

A useful way to classify ENMS is by the managed network. In a fuel-cell/battery electric bus, the hybrid FC/battery bus can be abstracted as a small DC microgrid with multiple generation units, storage, and a common bus; in that sense, the supervisory controller for multiple fuel-cell stacks is structurally identical to a small ENMS (Shi et al., 2023). In a mobile ad hoc cloud, resource management is explicitly cross-layer: the network layer manages transmission power, link quality, and link lifetime, while the middleware layer allocates and migrates tasks using both node and network metrics, which makes it ENMS-like in a distributed wireless environment (Shah, 2019). In discrete-event distribution-grid simulation, the hierarchy of Energy Management Units (EMUs) and System Participant Models (SPMs) turns device flexibility into a networked management problem with aggregation, disaggregation, and event-triggered control (Peter et al., 31 Jul 2025).

From this perspective, ENMS is not defined by whether the managed commodity is electricity alone, whether the network is AC or DC, or whether the topology is physical or cyber-physical. It is defined by coordinated decision-making over interconnected resources.

2. Architectural patterns and system decomposition

A recurrent architectural feature is layering. In microgrids, the field layer contains distributed generators, storage, controllable loads, sensors, smart meters, and PMUs; above it sit local controllers and HEMS; above them are communication layers such as HAN, WAN, AMI, and SCADA; and above those is the central EMS/SCADA layer that performs optimization, mode management, and dispatch (Pourbabak et al., 2017). This layered structure persists in more recent ENMS formulations, but with different decompositions.

In regional multi-microgrid communities, the dominant pattern is two-level hierarchy. Each microgrid has a local EMS that solves a mixed-integer linear program for PVs, ESs, EVs, and dispatchable loads, while a community-level EMS applies a pairing algorithm to determine inter-MG exchanges and minimize transactions with the upstream grid (Ju, 2021). In transmission operations, the architecture is procedural rather than territorial: SCADA/RTUs feed state estimation, which feeds AC real-time contingency analysis, which then parameterizes real-time security-constrained economic dispatch; an enhanced version inserts corrective transmission switching and branch pseudo limits between RTCA and SCED (Li et al., 2018). In EnergyNet, the decomposition is explicitly network-theoretic: Energy Routers, EROS, and EP Servers implement local and distributed control, while ENMS acts as the operator-scale management and orchestration plane for ELANs and EWANs (Birgersson et al., 9 Sep 2025).

The same idea can be organized compactly.

Context	Managed entities	Control structure
Multi-microgrid community	PVs, ESs, EVs, dispatchable loads	Local MILP + community-level pairing algorithm
Transmission control center	SE, RTCA, RT SCED, CTS	Procedure-A / Procedure-B with branch pseudo limit
EnergyNet	Energy Routers, ELAN/EWAN, EROS, EP Server	ENMS as digital control plane
Discrete-event distribution grid	SPMs and EMUs	Hierarchical flexibility aggregation and disaggregation

These architectures differ in scale and implementation, but they share a common separation between device-level actuation, network-level coordination, and operator or market policy.

3. Formal models and optimization foundations

The most general mathematical treatment models an energy network as devices connected through terminals and nets. In the static case, terminal power is $p \in \mathbb{R}^M$, device-local power is $p_d = B_d p$, net balance is $Ap = 0$, and total cost is

$f(p) = \sum_{d=1}^D f_d(B_d p).$

The resulting static optimal power flow is

$$\begin{array}{ll} \text{minimize} & f(p) \ \text{subject to} & Ap = 0, \end{array}$$

and its dynamic extension replaces $p$ by a time-indexed matrix $p \in \mathbb{R}^{M \times T}$ while preserving the same device-sum objective and nodal balance structure (Moehle et al., 2019). This formulation is broad enough to represent generators, storage, deferrable loads, thermal loads, grid ties, converters, and transmission lines, and it is directly usable as the planning core of an ENMS.

A decentralized formulation of the same class of problems is given by optimal power scheduling over a bipartite graph of devices and nets. There, each device has an extended-real-valued objective $f_d$, each net enforces

$\sum_{t\in n} p_t(\tau) = 0,$

and the network objective is $\sum_d f_d(p_d)$. The paper develops prox-average message passing, an ADMM-based method in which devices solve local proximal problems and nets update scaled prices using average net imbalance. Under closed, convex, proper device models, the method converges and yields locational marginal prices as dual variables (Kraning et al., 2012). This gives ENMS a rigorous distributed optimization backbone rather than only a centralized one.

Mixed-integer formulations appear whenever mode logic matters. In multiple-stack fuel-cell/battery hybrids, the supervisory problem minimizes hydrogen cost plus FC and battery degradation cost over a finite horizon,

$p_d = B_d p$0

subject to power balance, FC on/off logic, battery current limits, SOC dynamics, and terminal SOC constraints, and is reformulated as a mixed-integer quadratic program (Shi et al., 2023). In privacy-preserving cooperative multi-microgrid operation, the centralized exchange problem is decomposed by Lagrangian relaxation, with dual variables updated by a subgradient method so that only prices and exchange quantities are shared, not generation costs or demands (Ceja-Espinosa et al., 2022). In wireless sensor ENMS, the control problem is cast as an infinite-horizon discounted-cost MDP, with queue lengths and battery states as the Markov state and energy-sharing decisions as continuous actions, solved by DDPG (Barat et al., 2023).

The formal landscape is therefore heterogeneous—convex OPF, MILP, MIQP, MDP, actor-critic RL—but the underlying pattern is stable: state, constraints, and multi-period coupling define the feasible region, while the ENMS objective trades operational efficiency against risk, degradation, or service quality.

4. Coordination paradigms: centralized, hierarchical, distributed, and learned

A common misconception is that ENMS is inherently centralized. The literature instead shows four recurring coordination paradigms.

Centralized ENMS remains dominant when tight coupling and hard operational logic matter. The multi-stack fuel-cell controller solves one centralized MIQP that minimizes a global cost while coordinating all stacks and the battery (Shi et al., 2023). Transmission EMS likewise remains centralized at the control-center level: Procedure-A and Procedure-B integrate state estimation, RTCA, SCED, and optionally CTS into one operational workflow (Li et al., 2018).

Hierarchical ENMS appears when privacy, modularity, or scale dominate. In the regional multi-microgrid community, local EMSs solve detailed device scheduling problems and communicate only aggregated exchange quantities and locations to the community-level EMS, which then performs pairing and grid-transaction minimization (Ju, 2021). In event-driven distribution-grid simulation, lower-level EMUs aggregate flexibility from SPMs and report it upward, while superior EMUs disaggregate setpoints downward; the hierarchy can represent household, aggregator, and DSO roles (Peter et al., 31 Jul 2025).

Distributed ENMS is justified when local autonomy or communication locality is primary. Privacy-preserving multi-microgrid EMS uses dual decomposition and local subproblems, with price-like multipliers coordinating exchange while hiding local cost and demand information (Ceja-Espinosa et al., 2022). Prox-average message passing goes further by requiring only neighbor-to-neighbor messages and synchronized iterations, with no global coordinator beyond iteration timing (Kraning et al., 2012).

Learned ENMS replaces or approximates repeated online optimization by a trained policy. In islanded PV-battery microgrids, a neural-network EMS learns the mapping from measured states to optimal droop-frequency references using MILP-OPF-generated labels, with the explicit aim of balancing SoC without solving OPF online (Gupta et al., 2022). In active network management, Gym-ANM formalizes ANM as an MDP over AC power-flow-constrained distribution systems and makes PPO and SAC comparable against MPC baselines (Henry et al., 2021). In cooperative energy-harvesting sensor networks, DDPG learns continuous energy-sharing policies for a centralized ENMS over large node sets (Barat et al., 2023).

This diversity suggests that “optimal ENMS architecture” is not universal. A plausible implication is that architecture follows the dominant bottleneck: combinatorial mode logic favors centralized mixed-integer optimization, privacy favors dual decomposition, scale favors hierarchy, and high-rate control with model mismatch favors learned surrogates.

5. Flexibility, health, quality of service, and the cyber layer

ENMS is not only about balancing power; it is increasingly about valuing flexibility, preserving asset health, and coupling physical and cyber constraints.

In health-aware propulsion networks, fuel-cell degradation is decomposed into load-change, on/off switching, idling, and high-load losses, each mapped to monetized cost terms, while battery degradation is represented by a quadratic cost in current magnitude (Shi et al., 2023). In community microgrids, ES degradation is represented by a piecewise linear cost with SOS-2 variables, and EVs are modeled with parking windows and departure energy constraints, so the ENMS internalizes asset wear and availability rather than treating storage as ideal (Ju, 2021). In islanded PV-battery microgrids, the supervisory objective is not market cost but SoC equalization, because severe SoC divergence can cause some batteries to hit their limits and lose the ability to regulate the dc-link, threatening grid-forming operation (Gupta et al., 2022).

Flexibility is increasingly abstracted into technology-independent interfaces. In discrete-event distribution-grid simulation, each SPM exposes $p_d = B_d p$1, $p_d = B_d p$2, and $p_d = B_d p$3; EMUs aggregate these envelopes upward and disaggregate target power downward, which separates device physics from supervisory logic (Peter et al., 31 Jul 2025). This same abstraction can support self-consumption maximization, aggregator-level optimization, or DSO-level grid support.

In wireless and mobile ENMS, quality of service enters the objective directly. The mobile ad hoc cloud system defines

$p_d = B_d p$4

with energy consumption

$p_d = B_d p$5

and combines transmission power control, link quality, and link lifetime prediction to make task placement and migration decisions (Shah, 2019). Here, ENMS is explicitly cyber-physical: computation, communication, and battery state are co-optimized rather than treated separately.

In distribution ANM, quality and security constraints are stated directly in power-system variables. Gym-ANM environments penalize voltage-limit violations,

$p_d = B_d p$6

and branch apparent-power overloads,

$p_d = B_d p$7

while also accounting for curtailment and storage losses in the reward (Henry et al., 2021). ENMS thus becomes the mechanism by which security margins, flexibility, and service quality are translated into operational control.

6. Communication, computation, and representative performance

Because ENMS is computationally intensive and communication-rich, its practical viability depends on algorithmic structure and software substrate.

Microgrid ENMS has long been coupled to SCADA, AMI, HAN/WAN, and web-enabled supervisory layers, with OPC-DA, OPC XML-DA, and REST gateways supporting operator interfaces and supervisory control (Pourbabak et al., 2017). More recent systems generalize this into explicit control planes. EnergyNet defines ENMS as the digital control plane that provisions, monitors, secures, and optimizes a dynamic fleet of Energy Routers, ELANs, and EWANs, sitting above EROS and EP Servers and enforcing mutual TLS, RBAC, software rollout, and predictive maintenance (Birgersson et al., 9 Sep 2025). Discrete-event ENMS adds a lean request–response flexibility protocol in which EMUs request flexibility only when cached information is outdated, controlled entities return $p_d = B_d p$8, $p_d = B_d p$9, $Ap = 0$0, and then report the next time their flexibility will change, reducing unnecessary recomputation (Peter et al., 31 Jul 2025).

On the computational side, graph-based EMS power flow addresses the “faster than real-time” requirement. Using TigerGraph and graph-parallel factorization, graph-based Newton power flow achieved 3.91 ms on IEEE 118, 16.92 ms on 1425 buses, 30.51 ms on 2643 buses, and 119.22 ms on a 10,790-bus system; on IEEE 118, this was about 37× faster than MATPOWER Newton at 145.7 ms (Shi et al., 2018). In decentralized dynamic energy management, a serial implementation solved a problem instance with over 30 million variables in 52 minutes, while the paper argues that with decentralized computing the solve time would be less than one second (Kraning et al., 2012).

Representative application studies show that ENMS design choices materially affect operational outcomes. In a fuel-cell/battery city bus with eight identical PEM stacks, MIQP (CSC) required 14.38 s versus 32258.68 s for DP over a 600 s horizon, making MIQP 2243× faster; with individual stack control, total cost fell from 10.08 to 3.56, a 64.68% reduction, with battery degradation reduced by ~4%, hydrogen consumption by ~22%, FC idling loss by ~99%, and FC load-change loss by ~41% (Shi et al., 2023). In a regional multi-microgrid community, coordination reduced total cost from \$Ap = 0$$166.737 in the 4-MG case, an improvement of 9.474$$Ap = 0$$2498.304 in the 50-MG case, an improvement of 9.129$$Ap = 0$$3498.30 in 15.06 s versus 2053.48 s in the 50-MG case, while remaining within 0.2 <p>In mobile ad hoc cloud resource management, accumulative task completion time was reduced by 10–15$$Ap = 0$46,716.93 in cooperative operation, a reduction of about 32%, while each MG improved its own cost (Ceja-Espinosa et al., 2022). In cooperative energy-harvesting sensor networks, the DDPG sharing policy reduced minimum percentage data loss to 11% for a 10-node setting, compared with 17% for centralized DQN and 43% for no sharing, and maintained about 13% average data loss even at 500 nodes (Barat et al., 2023). In ANM6-Easy, PPO and SAC both outperformed constant-forecast MPC, with best reported returns of $Ap = 0$5 and $Ap = 0$6, respectively, versus $Ap = 0$7 for $Ap = 0$8, though perfect-forecast MPC remained strongest at $Ap = 0$9 (Henry et al., 2021).

The quantitative record therefore does not support the view that ENMS is merely a bookkeeping layer. It is often the dominant determinant of achievable speed, cost, degradation, throughput, and resilience.

7. Limitations, misconceptions, and research directions

The main limitation across ENMS research is the gap between mathematically elegant formulations and the full heterogeneity of operational practice. Several papers assume perfect future load profiles, deterministic exogenous processes, or simplified network physics. The multi-stack fuel-cell study assumes perfect knowledge of the 600 s horizon, uses empirical degradation surrogates, and does not model network constraints beyond a single bus (Shi et al., 2023). The mobile ad hoc cloud system is evaluated on 20 nodes and depends on hardware capable of multiple transmission-power levels; it also shows that control overhead can offset benefits when topology offers little power-control advantage (Shah, 2019). The regional multi-microgrid EMS approximates losses using distance-based weights rather than full AC power flow, though this is a deliberate scalability choice (Ju, 2021). Gym-ANM’s ANM6-Easy is a toy environment with deterministic daily profiles and centralized control, explicitly intended as an introductory benchmark rather than a full DSO-grade ENMS (Henry et al., 2021).

A second misconception is that decentralization automatically improves scalability without qualification. The message-passing framework proves convergence for convex device models, but nonconvexity remains problematic, and many realistic devices still require convex envelopes or approximate relaxations (Kraning et al., 2012). Privacy-preserving decomposition reduces data sharing, but communication patterns, step-size tuning, and convergence speed remain operational design questions rather than solved issues (Ceja-Espinosa et al., 2022).

A third unresolved area is standardization of the control plane. EnergyNet specifies ENMS conceptually as a digital control plane with EP, EROS, Energy Routers, and operator-scale lifecycle management, but formal standardization, interoperability testing, and verified integration with legacy SCADA/DMS/EMS are still open issues (Birgersson et al., 9 Sep 2025). The same is true, in different form, for cyber-physical ENMS in mobile and ad hoc environments, where state dissemination, migration triggers, and control security are central but not yet unified (Shah, 2019).

Current research directions converge on several themes. One is hierarchical composition: local EMSs embedded in larger ENMS layers, whether across device–household–aggregator–DSO hierarchies or across router–ELAN–EWAN overlays (Peter et al., 31 Jul 2025, Birgersson et al., 9 Sep 2025). Another is explicit uncertainty handling via stochastic MPC, scenario-based OPF, or learning-based controllers that approximate optimal responses without repeated online solves (Moehle et al., 2019, Gupta et al., 2022). A third is richer health-aware and flexibility-aware objectives, in which degradation, reserve provision, curtailment, quality of service, and cyber constraints are monetized or otherwise internalized rather than treated as afterthoughts (Shi et al., 2023, Henry et al., 2021). A plausible implication is that future ENMS will not converge to a single canonical algorithm; instead, they will combine convex optimization, mixed-integer logic, graph computation, event-driven communication, and learned surrogates within the same supervisory stack.