Event-Driven Edge Agent Framework for Distributed Control in Distribution Networks

Abstract

With the large-scale integration of heterogeneous energy resources and the increasing demand for flexible control, centralized control is facing challenges in terms of operational efficiency and system responsiveness when handling high-precision regulation tasks. To address this issue, this paper proposes an event-driven edge agent framework for distributed control in power distribution networks. First, based on the diverse requirements of distributed control in distribution networks, an edge agent architecture is constructed with modular components such as configuration management at its core. Second, considering the hybrid system characteristics of distribution networks, a control configuration technique based on activity-on-edge is designed, which decouples and discretizes continuous control processes through event-driven mechanisms. Furthermore, an edge-oriented automatic differentiation solver and a lightweight web application framework are developed to address the challenges of real-time optimization under resource-constrained environments. Finally, a semi-physical simulation is conducted using station-level economic dispatch as a case study to verify the effectiveness of the proposed technology. The results demonstrate that, compared to centralized control, the designed distributed agent maintains optimization accuracy while reducing event-triggering frequency by 40% and improving communication response speed by 70%, showing strong performance in operational efficiency at the edge.

1. Introduction

With the large-scale integration of distributed generation, energy storage systems, and various controllable loads [1,2,3,4], power distribution systems exhibit significant differences in dynamic response characteristics, regulation capabilities, and control interaction mechanisms [5,6,7], thereby increasing the complexity of system operation. This high degree of heterogeneity substantially raises the difficulty of unified scheduling and global coordination, posing severe challenges to traditional centralized control approaches in managing complex control tasks [8,9]. Meanwhile, the growing demand for flexible control continues to push the system toward higher requirements in regulation frequency, response latency, and edge-side execution capability [10,11,12]. In particular, as emerging applications such as electric vehicle clusters [13] and virtual power plant aggregation [14] are rapidly deployed, distribution network control tasks are becoming increasingly decentralized and dynamic, further intensifying the communication and computational burdens of centralized control architectures.

To address the above challenges, edge agents have been introduced as novel solutions for distributed control in power distribution networks [15,16,17]. By offloading computationally lightweight tasks from the central controller to devices closer to the field, the system enables localized real-time data perception and coordinated regulation with the control center, thereby improving response speed and operational flexibility [18]. Researchers have explored the engineering pathway of deploying edge agents in power systems. Reference [19] systematically elaborates on the integrated architecture of edge agents and artificial intelligence technologies in the context of the power Internet of Things. Reference [20] constructs a single-phase ground fault section location model by integrating deep learning algorithms into the edge agent framework. References [21,22] propose optimized deployment strategies for edge agent clusters in distribution networks from the perspectives of software-defined network architecture and fault handling requirements, respectively. Existing studies mainly focus on how edge agents overcome the computational bottlenecks of centralized control [23,24]. However, achieving optimal control performance under resource-constrained conditions remains a key challenge between academic research and practical deployment. Current approaches to address this issue include lightweight deployment [25] and rational allocation of computing resources [26,27,28].

Lightweight deployment encompasses both algorithm-level and model-level strategies. Current algorithm deployment often relies on pre-debugged embedding [29], which is not suitable for large-scale, dynamic deployment and online updating across massive terminal devices. At the model level, existing continuous system operation models for edge agents typically rely on gradient-based iterative mechanisms, which lead to continuous occupation of communication bandwidth. In recent years, the hybrid characteristics of power systems have been recognized [30], giving rise to hybrid system modeling approaches that abstract high-dimensional nonlinear systems as discrete event-triggered state transition systems. These methods achieve model order reduction by decoupling dynamic systems from discrete events [31]. Moreover, modeling based on discrete event-driven mechanisms initiates state evolution computation only when events are triggered, significantly reducing the computational resource consumption of edge agents. Related technologies have already been validated in microgrid energy management systems [32,33].

In addition, in multi-objective optimization control of distribution networks [34,35], time-iterative continuous control often results in high redundancy of control commands, which severely limits the real-time responsiveness of control actions and increases the computational burden on edge agents. Current research primarily focuses on resource allocation algorithms to improve resource utilization efficiency [36,37]. However, systematic solutions for optimizing continuous control mechanisms under time-step constraints remain lacking. By adopting event-driven control, the system’s dynamic evolution can be transformed into a state transition trigger mechanism, thereby avoiding repetitive time-step iterations and improving computational resource efficiency [38].

Based on the above background and challenges, this paper conducts research on distributed edge agent technology for distribution networks. First, an edge agent control architecture is designed in response to the diverse requirements of distributed control in distribution networks, enabling flexible deployment of control strategies through modules such as control configuration and intelligent computation. Second, a hybrid system model for distribution networks is established based on hybrid control theory, and an event-driven control configuration technique is proposed to discretize complex continuous control processes. The control strategy is formally modeled using AOE, where a joint dynamic mechanism of node events and edge actions is employed to achieve on-demand control. Furthermore, an automatic differentiation-based solver and a lightweight web application framework tailored for edge environments are developed to address the challenge of real-time optimization in resource-constrained settings. Finally, using station-level optimal dispatch as an application scenario, the proposed control configuration technique is applied to model and configure a distributed optimization strategy based on consensus algorithms, and the applicability of the proposed architecture and control method is validated through semi-physical simulation.

2. Edge Agent Architecture for Distributed Control

Edge agents are typically located close to distributed and flexible resources such as photovoltaics and energy storage systems, enabling localized computation and fast control near the source side by leveraging their edge computing capabilities. Through low-latency communication mechanisms, edge agents can not only execute localized control strategies but also respond to centralized optimization and dispatch, serving as a key enabler of local autonomy and global coordination in modern distribution networks. Based on this, the edge agent architecture constructed in this paper is shown in Figure 1. The main functional modules are described as follows.

1.

Configuration Module

The edge agent provides a secondary development interface to users in the form of configuration, including control strategies, communication channels, and measurement points. The control strategy adopts a control configuration technique based on hybrid control theory, enabling custom editing of edge computing strategies. Channels are used to establish communication between the edge agent and controlled devices, supporting configuration of various communication protocols. Measurement points describe the state of the controlled objects and serve as the data foundation for control strategies or communication. They are categorized into acquisition points and calculation points.

2.

Central Processing Module

The AOE scheduler verifies, parses, and executes the configuration to enable optimized operation for the designed control scenario. The edge agent performs mathematical computations such as solving optimization models and equation systems within the control strategy using an automatic differentiation-based optimization solver. This paper uses the IPOPT optimization solver, which was developed by the Computational Infrastructure for Operations Research (COIN-OR) Foundation, located in Chicago, IL, USA. A coroutine-based high-performance asynchronous execution engine supports high-concurrency task scheduling, and an embedded database is integrated to handle data storage and management under resource-constrained conditions. The asynchronous execution engine in this paper is designed based on the Tokio asynchronous runtime framework, which is developed by contributors to the Tokio project and maintained by an organization based in the USA. The database is built on RocksDB, which was developed by Facebook Inc., headquartered in Menlo Park, CA, USA.

3.

Web Application Interface Module

The edge agent employs a browser-based web application as the user interface, providing an interactive platform for user operations and offering intuitive visualization of various data, reports, and the execution statuses of control strategies. The web application interface in this paper is designed using WebAssembly, which is led by the World Wide Web Consortium (W3C) and jointly promoted by major technology companies including Mozilla, Google, Microsoft, and Apple. The standardization and development efforts are primarily carried out in the USA.

3. Edge Agent Control Configuration Technology Based on Hybrid Control

3.1. Hybrid Control Theory

As a hybrid dynamic system, the distribution network requires coordinated operation between continuous variable control and discrete logical constraints to achieve multi-objective optimization [32]. Its dynamic behavior is typically represented mathematically by high-dimensional nonlinear differential–algebraic equations, expressed as

$$\begin{array}{l}\min J(\mathit{x},\mathit{y})={\displaystyle {\int }_0^{\infty }{\displaystyle \sum _{i=1}^n{q}_i{F}_i(\mathit{x},\mathit{y}),}}\\ \begin{array}{cc}\hfill s.t. \dot{\mathit{x}}& =f(\mathit{x},\mathit{u}),\hfill \\ \hfill 0& =\phi (\mathit{x},\mathit{y}),\hfill \end{array}\end{array}$$

(1)

where x ∈ $R^n$, y ∈ $R^m$, and u ∈ $R^r$ represent the system state vector, output vector, and control vector, respectively. Traditional methods struggle to obtain satisfactory solutions within a limited time frame [39]. To address this high-dimensional constrained variational problem, edge agents adopt a hybrid control approach. The full state space of the system is partitioned based on whether the multi-objective optimal operating conditions are satisfied, resulting in a satisfactory state set and an unsatisfactory state set, defined as

$$X(T)=X_S(T)\cup X_{\overline{S}}(T)$$

(2)

where X_s(T) represents the satisfactory state set of the system at time T, that is, the state set where multi-objective optimal conditions are met; $X_{\overline{S}}\left(T\right)$ represents the unsatisfactory state set at time T; and X(T) represents the full state space of the system at time T.

When the unsatisfactory state set $X_{\overline{S}}\left(T\right)$ is an empty set, the system is considered to have achieved the desired operational objectives. At any time T, when the system operating state enters $X_{\overline{S}}\left(T\right)$, an event E is said to be triggered:

$$E=f_E(D)$$

(3)

where E represents the set of events, D represents the system operating data, and f_E represents the logical evaluation function that determines whether an event in E is generated based on D.

When an event E triggers the corresponding control strategy that drives the system operating state back to X_s(T), the system’s multi-objective optimal operation is achieved to a certain extent:

$$C=f_C(E)$$

(4)

where C represents the set of action commands executed by the edge agent, and f_C represents the transition function that generates control strategies from the event set E.

The edge agent issues control commands to operate devices within the system, leading to a change in the system state that brings it back to the satisfactory state space. At this point, the system is in a no-event operational state, expressed as

$$\dot{\mathit{x}}=f_{\mathit{x}}(\mathit{x},C)$$

(5)

$$A(C)\to E\subseteq 0$$

(6)

where x represents the state variable of the system, and f_x is the state equation that describes how the system state is changed through control. Equation (6) indicates that, through control actions, the event set E becomes an empty set.

When Equation (6) holds, it indicates that the system is operating in a satisfactory state. In other words, satisfying Equation (6) also signifies that the system has achieved multi-objective optimal operation. In summary, the system is continuously in the process of event detection, handling, and elimination. The control method based on this theory is referred to as event-driven control.

3.2. Control Configuration Technology

3.2.1. Control Configuration

As shown in Section 3.1, the hybrid control approach can be modeled using two key elements—events and action commands—and their interactions. Based on this, a control configuration technology is proposed in this paper. Control configuration is similar to “building blocks”, where control strategies are developed by selecting basic components, setting parameters, and combining them through a “configuration” process rather than traditional programming [40]. Compared with conventional programming-based development, this approach offers advantages such as lower development barriers, inherent correctness, and greater generality.

3.2.2. Control Process Model Based on AOE

In distributed scenarios, the design and execution of control strategies are essential to ensuring efficient system operation. The control process can essentially be abstracted as an orderly transition of system states across multiple state sets. During the transition from an initial satisfactory state to the next, the system typically passes through several intermediate unsatisfactory states, which require corresponding control actions for correction. As this process cannot be handled by a single rule, a clear and structured modeling method for control logic is required. To this end, this paper adopts an AOE structure to intuitively describe the control process, with the corresponding control module workflow illustrated in Figure 2. The AOE structure is fundamentally a directed acyclic graph with a single input and a single output, composed of nodes and directed edges. Its implementation details are presented as follows.

In terms of static modeling, the node set in the AOE structure corresponds to a subset of the event set E defined in Equation (3), with each node expression formulated according to the rules of f_E. Each directed edge represents a control action, i.e., a subset of the instruction set C defined in Equation (4), and its action logic and execution expression are specified based on the rules of f_C. The connectivity between nodes and edges reflects the logical dependencies and execution sequences between events and actions, and the resulting topological structure is consistent with the causal relationships among the elements in sets E and C. The terminal node serves as a validity check for the AOE-based control strategy, corresponding to Equations (5) and (6), to determine whether the current strategy can guide the system back to a satisfactory operational state, i.e., whether the intended control objective has been achieved.

In terms of dynamic execution, the control process starts from the initial node. The system continuously evaluates the event trigger expression defined at this node based on real-time measurement data. Once the expression is satisfied, the node is considered triggered, corresponding to the event generation described in Section 3.1. The system then follows the conditional logic specified at the node to select the corresponding outgoing edge and executes the control action defined on that edge. The execution of the edge action is typically accompanied by updates to system state variables or the issuance of control commands, thereby driving the system toward the satisfactory state space. Upon completion of the action, the process transitions to the next node pointed to by the edge, where the evaluation cycle continues until the next triggering condition is met. Starting from the AOE initial node, this process continues to drive the system until the end node is reached, thus completing the entire execution of the control strategy. The migration at each step is determined by the triggering of node events, thereby implementing event-driven control strategy execution.

Taking the AOE-based control flow illustrated in Figure 2 as an example, the edge agent first collects real-time operating data from nearby devices. Once the system is determined to have deviated from its target and entered an unsatisfactory state, the corresponding AOE control strategy is immediately activated to guide the system back to a satisfactory condition. The control process begins at the initial node, where the predefined logical expression is evaluated to determine whether the triggering condition is satisfied. If the condition holds true, the system proceeds along edge 1–2 to execute the corresponding control action and then transitions to the next node, node 2. This intermediate node typically represents a temporary state, which may be used to assess control effects or evaluate downstream branch logic. If the path condition is met, the system continues along edge 2–3 to execute the next control action and proceeds accordingly until the terminal node is reached. The terminal node is associated with a termination event, which is used to verify whether the control task has been successfully completed—for example, whether the command was issued correctly, key actions were executed as expected, or optimization tasks were fulfilled—thus completing the full closed-loop execution of the control strategy.

3.3. Control Configuration Implementation

When editing control strategies using edge agents, the AOE nodes, edges, and their attributes are configured according to the defined standard format. The AOE scheduler will automatically validate, parse, solve, and activate the user-defined AOE for execution.

First, based on the fundamental concepts and properties of the AOE data structure, the following rules and constraints are applied:

• A node with an in-degree of zero is called the AOE start node, and a node with an out-degree of zero is called the end node. There is exactly one start node in each AOE, and there are one or more end nodes.
• When a node has multiple out-degrees, the multiple outgoing edges are numbered in the order of configuration, such as edge 1, edge 2, and so on.
• Each node has an in-degree greater than or equal to zero. When a node has multiple incoming edges, the node can be reached when any of the incoming edges is triggered and its associated action is completed. At this point, the event represented by the node is in a waiting-trigger state.
• Multiple expressions can be listed on a single edge. Once an edge is entered, all expressions on that edge are executed simultaneously.

Next, in terms of the specific configuration format for the basic unit nodes and edges of the AOE, their definitions are as described in Section 3.2.2. The specific expressions for configuration vary depending on the type and strategy design, and due to space limitations, only the basic forms of both are presented here without further elaboration.

In terms of node types, the behavior of each node in the AOE control flow determines how its outgoing edges are handled. Based on different control logic requirements, two basic execution modes are commonly defined: the first is sequential execution, in which all outgoing edges are executed in order once the node is triggered; the second is conditional branching, where one edge is selected from multiple candidates based on the current system state or a specific logical expression, enabling decision-making, switching, and conditional transitions within the control strategy.

Accordingly, the event types of nodes in this paper are categorized into two major types: condition-execution nodes, which trigger all subsequent actions and emphasize execution completeness; and branch-selection nodes, which determine the next control path based on logical conditions and emphasize control divergence. Complex control processes can be constructed by combining these two types of nodes, providing strong generality and formal consistency in the expression of structured control logic.

In terms of edge action types, for various computational requirements in distribution networks, such as solving nonlinear equation systems in distributed power flow calculations or solving optimal economic dispatch models formulated as MILP or NLP, the control configuration provides directly callable properties for computing edge actions. Meanwhile, for control commands involving remote switching and setpoint adjustment by the agent, the control configuration also supports assignment edge actions to set measurement point values. Based on this, complex control strategies can be developed and implemented to meet the control requirements of distributed scenarios.

It should be noted that the solver invoked in the control process is an open-source implementation, which is solely responsible for executing mathematical computations and cannot directly determine the feasibility of the constructed optimization model. When the optimization problem is mathematically solvable and successfully converges, the solver returns a valid optimal solution. Conversely, if the model is infeasible or fails to converge during computation, the solving process is considered unsuccessful. Therefore, it is typically necessary to define branching logic at the node following the corresponding computation edge. If the solution is successful, the system retrieves the optimization results and issues control commands accordingly. If the solution fails, the edge agent automatically enters an exception handling process, triggers a failure alert, and executes a predefined backup control strategy to ensure the continuity and safety of device operation.

Finally, in the overall configuration of the control strategy, the activation of the AOE and the execution of the entire network’s behavior from the start node are defined as the AOE triggers. Based on the requirements of practical engineering applications, the AOE trigger is divided into three types, as shown in

Table 1. Trigger types of AOE network.

Trigger Type	Name	Definition
EventDriven	Event-driven	Enable the AOE network when the first node event condition is met
TimeDriven	Time-driven	Enable the AOE network at the set time
SimpleRepeat	Simple repetition drive	Enable and loop the AOE network at the set time

.

4. Automatic Differentiation-Based Optimization Solving Technology for the Edge Side

As described in Section 3.3, the actions in the control configuration involve computational requirements such as NLP and MILP, which often require large-scale derivative calculations including Jacobian and Hessian matrices of multivariable functions. To address this, this paper applies automatic differentiation algorithms to optimize the solving process within edge agents. By decoupling control modeling from the solving engine through control configuration and automatic differentiation, the approach enables solver reuse, thereby enhancing the operational efficiency and reliability of edge agents. At the same time, it allows users to focus solely on the design of control strategies, simplifying mathematical computation and reducing implementation complexity.

4.1. Model Solving Based on Automatic Differentiation

Automatic differentiation technology is divided into forward mode and reverse mode [41]. By employing a hierarchical composite function and layer-by-layer derivative solving strategy, the computational complexity is effectively reduced and the final expressions are simplified. As a result, the overall storage space required for computation is significantly reduced compared to symbolic differentiation methods [42].

For the function y = f(x), it can be represented as

$$\left\{\begin{array}{l}y=f_i(w_i)\\ w_i=f_{i-1}(w_{i-1})\\ ······\\ w_1=f_0(w_0)\\ w_0=x\end{array}\right.$$

(7)

In the forward mode, the computation proceeds by first evaluating dw₀/dx using Equation (7),and then recursively deriving dw_k/dx up to dw_i/dx, from which dy/dx can be obtained.

$${\displaystyle \frac{\mathrm{d}{w}_1}{\mathrm{d}x}}={\displaystyle \frac{\mathrm{d}{w}_1}{\mathrm{d}{w}_0}}{\displaystyle \frac{\mathrm{d}{w}_0}{\mathrm{d}x}},{\displaystyle \frac{\mathrm{d}{w}_2}{\mathrm{d}x}}={\displaystyle \frac{\mathrm{d}{w}_2}{\mathrm{d}{w}_1}}{\displaystyle \frac{\mathrm{d}{w}_1}{\mathrm{d}x}},\dots ,{\displaystyle \frac{\mathrm{d}{w}_i}{\mathrm{d}x}}={\displaystyle \frac{\mathrm{d}{w}_i}{\mathrm{d}{w}_{i-1}}}{\displaystyle \frac{\mathrm{d}{w}_{i-1}}{\mathrm{d}x}},{\displaystyle \frac{\mathrm{dy}}{\mathrm{d}x}}={\displaystyle \frac{\mathrm{dy}}{\mathrm{d}{w}_i}}{\displaystyle \frac{\mathrm{d}{w}_i}{\mathrm{d}x}}$$

(8)

The computation process of the reverse mode is opposite to that of the forward mode. It begins with calculating dy/dw_i using Equation (7) and then proceeds downward to derive dy/dw_k and eventually dy/dw₀, from which dy/dx can be obtained.

$${\displaystyle \frac{\mathrm{dy}}{\mathrm{d}{w}_{i-1}}}={\displaystyle \frac{\mathrm{dy}}{\mathrm{d}{w}_i}}{\displaystyle \frac{\mathrm{d}{w}_i}{\mathrm{d}{w}_{i-1}}},{\displaystyle \frac{\mathrm{dy}}{\mathrm{d}{w}_{i-2}}}={\displaystyle \frac{\mathrm{dy}}{\mathrm{d}{w}_{i-1}}}{\displaystyle \frac{\mathrm{d}{w}_{i-1}}{\mathrm{d}{w}_{i-2}}},\dots ,{\displaystyle \frac{\mathrm{dy}}{\mathrm{d}{w}_0}}={\displaystyle \frac{\mathrm{dy}}{\mathrm{d}{w}_1}}{\displaystyle \frac{\mathrm{d}{w}_1}{\mathrm{d}{w}_0}},{\displaystyle \frac{\mathrm{dy}}{\mathrm{d}x}}={\displaystyle \frac{\mathrm{dy}}{\mathrm{d}{w}_0}}{\displaystyle \frac{\mathrm{d}{w}_0}{\mathrm{d}x}}$$

(9)

From a mathematical perspective, both the forward and reverse modes of automatic differentiation are derived from the chain rule and thus provide theoretically equivalent accuracy in gradient computation. Meanwhile, by comparing Equations (8) and (9), it can be observed that in the forward mode, each input variable corresponds to a distinct intermediate derivative path, requiring multiple propagation steps to obtain the complete gradient. In contrast, the reverse mode shares a unified propagation path for all intermediate derivatives, allowing all partial derivatives with respect to input variables to be obtained after computing a single backward pass [43]. In high-dimensional gradient computation tasks, the reverse mode can significantly reduce overall computational load, making it more suitable for edge agents that demand fast response and operate under constrained resources. Therefore, this study adopts the reverse mode to improve computational speed.

4.2. Case Study Validation

Taking the optimal power flow problem in a distribution network as an example, the validation of the edge agent’s model-solving capability based on automatic differentiation is performed. Mathematically, the optimal power flow calculation is a typical nonlinear optimization problem, and the mathematical model is as follows:

$$\left\{\begin{array}{l}\min F(\mathit{x},\mathit{u})\\ s.t. G(\mathit{x},\mathit{u})=0,H(\mathit{x},\mathit{u})\le 0\end{array}\right.$$

(10)

where F represents the objective function; G represents the equality constraints; H represents the inequality constraints; x is the state variable; and u is the control variable.

The standard case 9 test case from MATPOWER 8.0 is selected, with the objective function set to minimizing system operating cost, i.e., minimizing the active power output of generating units under fixed fuel prices. This paper uses MATPOWER 8.0, which was developed by the Power Systems Engineering Research Center (PSERC) at Cornell University, Ithaca, NY, USA.

The AOE control strategy network for optimal power flow calculation consists of four nodes: start node 1, computation node 2, assignment node 3, and end node 4. The action executed on edge 2–3 is a nonlinear programming task for solving the optimal power flow, while edge 3–4 performs an assignment operation used to display the optimal power flow calculation results.

The results obtained by executing the AOE network are shown in Appendix A,

Table A1. Operation results of the optimal power flow calculation case.

Serial Number	Point Number	Name	Another Name	Current Value
1	1001	Point1	V1	1.099999920539086
2	1002	Point2	V2	1.0973546514586485
3	1003	Point3	V3	1.0866203282032385
4	1004	Point4	V4	1.094221462595771
5	1005	Point5	V5	1.084448462899668
6	1006	Point6	V6	1.1000000109999306
7	1007	Point7	V7	1.0894894953643341
8	1008	Point8	V8	1.1000000109967762
9	1009	Point9	V9	1.0717554151378195
10	1010	Point10	THETA1	0.0000000000000000
11	1011	Point11	THETA2	0.0854101800932149
12	1012	Point12	THETA3	0.05671518263242919
13	1013	Point13	THETA4	−0.04298613947999122
14	1014	Point14	THETA5	−0.06949871348651646
15	1015	Point15	THETA6	0.010522383653749938
16	1016	Point16	THETA7	−0.020878902993962594
17	1017	Point17	THETA8	0.015806264325692185
18	1018	Point18	THETA9	−0.0805511204216152
19	1019	Point19	PG1	0.8979870754479929
20	1020	Point20	PG2	1.3432060080841293
21	1021	Point21	PG3	0.9418738048899336
22	1022	Point22	QG1	0.1296559578027687
23	1023	Point23	QG2	0.0003187190823897824
24	1024	Point24	QG3	−0.22634181614061613
25	1025	Point25	Docal_POINT	0

. In this table, voltage angles are expressed in radians, and power values are in per unit (p.u.), which must be multiplied by the base value of 100 MW to obtain physical units. The results computed by MATPOWER 8.0 are shown in Appendix A, Figure A1, where the voltage angles are expressed in degrees. A comparison shows that the two sets of results are consistent, thereby verifying the computational accuracy of the edge agent based on automatic differentiation.

5. Web Application Technology for Edge Agents

The deployment of edge agents requires deep coupling with the underlying hardware architecture, and their design must follow the principles of high performance and lightweight implementation to adapt to heterogeneous edge computing resources and the constraints of real-time response and power sensitivity. To this end, the edge agent adopts a microservice architecture for backend development and uses a web application frontend developed in WebAssembly. Both frontend and backend are implemented in Rust to enable unified full-stack programming and improve code reuse, thereby enhancing system operational efficiency.

5.1. Database Technology for Edge Agents

Embedded database technology is used for data storage and management, with access control handled by the application itself, eliminating the need for a separately deployed database server. Embedded databases are customizable, lightweight, and integrated within the process, making them suitable for embedded systems. In this paper, the open-source arbitrary binary K-V (key-value, key-value pair) data storage engine RocksDB is adopted as the embedded database implementation for the edge agent.

The data structure and management method of RocksDB are introduced as follows. RocksDB has three types of data structures: (1) mentable, a memory data structure, where all write requests enter; (2) logfile, an ordered write storage structure, where write requests enter selectively; and (3) sstfile, when the mentable is full, data enter the sstfile and are stored, and the related logfile is safely deleted. The data in the sstfile are sorted, so that they can be quickly searched based on the key. In terms of data management, RocksDB stores data as binary key-value pairs and supports sharding of a database instance by column families. RocksDB organizes all data in sequence and can specify compression algorithms to sort the key-value pairs. Common operations include retrieving a value by key (Get), inserting a key-value pair (Put), deleting a key-value pair (Delete), and iteration (Iter).

Compared with traditional storage solutions that use high-capacity servers to store raw text and other files, embedded databases offer higher reliability and stronger low-level control capabilities. They can operate stably and reliably over long periods without human intervention and are capable of directly interacting with underlying hardware devices. The use of embedded databases eliminates the overhead of client-server architecture, resulting in more compact code, faster execution speed, and improved performance when running in embedded environments.

5.2. Asynchronous Non-Blocking Communication Technology

In various edge computing scenarios, reliable acquisition of measurement data from devices is a prerequisite for controlling them. Distribution networks contain a wide range of complex device types, requiring the edge agent to support parallel data acquisition from different devices using multiple communication protocols and to standardize and store the collected data.

The edge agent application adopts asynchronous non-blocking technology to implement communication gateways for the power Internet of Things, supporting multiple protocols including IEC 104, Modbus, DL/T645, and MQTT. Taking the non-blocking implementation of the IEC 104 protocol as an example, the protocol is designed with a three-layer structure: TCP/IP layer, application layer, and user process.

Based on the Tokio asynchronous framework, a non-blocking I/O network communication framework is developed, covering client connection, server listening and connection, and read/write method design, enabling data exchange over lower-layer TCP/IP connections. Asynchronous channels are used to support information sharing and command transmission between the application layer and the user process. In the application layer, asynchronous tasks are used to implement IEC 104 protocol functions such as link transmission rules, message parsing and assembly, and application functionalities. The overall structure and data flow of IEC 104 are shown in Figure 3.

In the asynchronous non-blocking communication framework described above, the underlying non-blocking network communication effectively conserves hardware resources and enhances the parallel processing capability of multi-channel communication. The user process standardizes the collected data from each channel, converting them into a unified measurement point format, which enables consistent computation and logical analysis across data from different devices using a generalized approach. By decoupling data acquisition from data processing, each stage operates in parallel, improving overall data flow and processing speed.

Consistency and performance tests were conducted on the communication system developed using this framework. As shown in Figure 4, under high communication load, the delay rate of the proposed method is significantly lower than that of traditional solutions. This verifies that the developed communication application can complete data transmission in accordance with protocol specifications. Based on asynchronous non-blocking technology, the edge agent can still ensure the accuracy and real-time performance of communication data on resource-constrained embedded devices.

6. Station-Level Optimal Dispatch in Distribution Networks Based on Distributed Edge Agents

This section verifies the distributed control technology of edge agents using the economic dispatch of distribution network stations. First, a station-level economic optimization dispatch model based on a discrete consensus algorithm is designed. Then, a digital twin of the actual system is simulated, and the distributed control performance of the edge agents is analyzed.

6.1. Economic Dispatch Optimization Model for Stations

To achieve economic optimality in a station, the total generation cost within the station’s jurisdiction should be minimized. The generation cost C_i of the distributed generation unit at bus i in the station can be approximated by a quadratic function [44], as shown below.

$$C_i={\displaystyle \frac{1}{2}}{a}_i{P_i}^2+b_i{P}_i+c_i$$

(11)

where a_i, b_i, and c_i represent the cost coefficients of the distributed generation unit, and P_i represents the actual active power output of the unit.

Given that the bus voltage remains stable under actual station operating conditions and the deviation is negligible, the actual active power output P_i (in p.u.) of the distributed generation unit is approximately equivalent to its output current I_i (in p.u.) [45]. Accordingly, Equation (11) can be rewritten as

$$C_i={\displaystyle \frac{1}{2}}{a}_i{I_i}^2+b_i{I}_i+c_i$$

(12)

Based on Equation (12), an economic dispatch optimization model for the station is constructed. Considering the power transfer requirements among the buses within the station, in addition to satisfying the station power balance equation and the output current constraints of the distributed generation units, corresponding voltage constraints are formulated. The arithmetic mean of bus voltages is used as the reference value. The objective function and constraints are defined as follows:

$$\begin{array}{l}f=\min {\displaystyle \sum _{i=1}^n{C}_i}\\ \mathrm{s}.\mathrm{t}.\left\{\begin{array}{l}{\displaystyle \sum _{i=1}^n{I}_i}={\displaystyle \sum _{i=1}^n{L}_i}\\ 0\le I_i\le I_i^{\max }\\ {\displaystyle \frac{{\displaystyle \sum _{i=1}^n{U}_i}}{n}}=U_{\mathrm{mean}}\end{array}\right.\end{array}$$

(13)

where L_i represents the load current injected at bus I; $I_i^{\max }$ represents the maximum output current of the distributed generation unit at bus I; U_i represents the measured voltage at bus i; and U_mean represents the reference value of the bus voltage.

6.2. Mathematical Design of Control Strategy

Based on Lagrangian duality theory, the optimization functional of the station economic dispatch model is constructed. Through gradient conditions and the first-order optimality conditions of the Karush–Kuhn–Tucker (KKT) framework, it is analytically derived that the total generation cost of the station reaches its minimum when the marginal costs at all buses are equal. The corresponding expression is

$${\displaystyle \frac{\partial C_1}{\partial I_1}}={\displaystyle \frac{\partial C_2}{\partial I_2}}=\dots ={\displaystyle \frac{\partial C_n}{\partial I_n}}$$

(14)

Define λ_i = ∂C_i/∂I_i as the marginal cost at bus i. Considering the property of multi-agent systems to achieve state consensus through distributed information exchange, a distributed consensus computation framework is constructed using marginal cost as the consensus state variable. Under this framework, each agent iteratively updates its marginal cost until convergence is achieved. For bus i, the control steps are as follows.

Step 1: Initialization. When the economic dispatch task is initiated, the edge agent acquires bus information and calculates the marginal cost λ_i[k] of bus i at time step k as well as the marginal cost deviation μ_i[k] with respect to its topological neighboring buses. The calculations are as follows:

$${\lambda }_i\left[k\right]=a_i{I}_i\left[k\right]+b_i$$

(15)

$${\mu }_i\left[k\right]={\displaystyle \sum _{j=1,j\in L_i}^n{g}_{ij}({\lambda }_i\left[k\right]}-{\lambda }_j\left[k\right])$$

(16)

where g_ij represents the element in the i-th row and j-th column of the nodal conductance matrix G. L_i represents the set of topological neighboring buses of bus i.

Step 2: Determine iteration status. If the event trigger condition defined in Equation (17) is satisfied [46], proceed to Step 3; otherwise, maintain the current state until the next time step and terminate the task.

$$\left\{\begin{array}{l}error=e_i\left[k\right]-{\lambda }_i\left[k\right]\\ erro{r}^2>{\displaystyle \frac{\sigma \epsilon \left[1-2\epsilon \mathit{G}(i,i)\right]}{2\mathit{G}(i,i)}}{\mu }_i{\left[k\right]}^2\end{array}\right.$$

(17)

where e_i[k] represents the event-triggered marginal cost stored at bus i at time step k; and σ and ε represent error coefficients, where 0 < σ < 1. To ensure that Equation (17) remains non-negative, it must also satisfy 0 < ε < 1/G(i,i).

Step 3: Update the event-triggered marginal cost, calculate the voltage adjustment value for bus i at time step k + 1, and issue the control command to complete the task.

$$e_i\left[k+1\right]={\lambda }_i\left[k\right]$$

(18)

$$V_i\left[k+1\right]=V_i\left[k\right]-q{\mu }_i\left[k\right]$$

(19)

where q represents the step-size coefficient. To ensure control stability, it must satisfy 0 < q < 1/λ_Anλ²_Gn [47], where λ_An is the maximum value among all cost coefficients a_i, and λ_Gn represents the largest eigenvalue of the conductance matrix G.

6.3. Strategy Description Based on Control Configuration Technology

The operational states and dynamic responses in the control steps of Section 6.2 are defined as events and actions, respectively. Combined with control configuration technology, the control steps can be described using an AOE structure, as shown in Figure 5. Taking bus 1 as an example, the configuration expressions for each event and action are provided in Appendix B, thereby completing the control strategy design.

Specifically, node 1 is the start node with an EventDriven trigger type. When the agent receives the start command, Start_command1, the strategy is initiated, and edge 1–2 is executed to calculate the marginal cost and deviation based on Equations (15) and (16). Node 2 is a switch-type event node; if the condition is satisfied, the process proceeds to node 4; otherwise, it jumps to termination node 3 to exit the process. Then, actions on edges 2–4 and 4–5 are executed, corresponding to Step 3 in Section 6.2. After completion, the process reaches end node 5, closing the AOE-based control loop.

6.4. Simulation Analysis

6.4.1. Station Simulation Environment Construction

A digital twin model of a five-bus station system in a distribution network is constructed in Simulink to simulate real-time economic dispatch at the station level. The topology of the five-bus station is shown in Figure 6, and the corresponding parameters are listed in Appendix C. The simulation model is built using Simulink 10.2 in MATLAB 2020b, and the simulation is run on a computer equipped with an Intel(R) i7-8700 processor and 16 GB of RAM.

Figure 7 shows the hardware-in-the-loop simulation architecture between the edge agent and Simulink. The edge agent is deployed on hardware and is responsible for collecting model data, executing the AOE control strategy, and sending control commands to the Simulink model. In Simulink, Modbus communication with each edge agent is implemented using the S-Function module, which serves as the control module in the model, linking the input signals from data acquisition and the output signals for control commands. The hardware device used in this study is equipped with an ARMV7 CPU, 2 GB of RAM, and a 4 GB hard drive.

6.4.2. Economic Dispatch Results and Analysis

Figure 8 shows the simulation results of station-level economic dispatch, where the dispatch task is initiated at 1 s. It can be observed that after each load fluctuation, the edge agents dynamically adjust the output power of the distributed generation units by modifying the marginal cost. The marginal costs quickly converge to a common value, thereby achieving economic optimality. Additionally, when the load at bus 1 exceeds 1 p.u., the edge agents are able to control other buses to support bus 1, ensuring that the DG output at bus 1 does not exceed its upper limit. Throughout the control process, the average voltage across all five buses remains at the rated level, demonstrating the effectiveness of the proposed distributed control algorithm.

6.4.3. Comparison of Dispatch Trigger Frequency

During the operation of control systems, frequent communication and repetitive computation often lead to waste of computational resources and device energy, which has become a major bottleneck for the large-scale deployment of edge agents. In the context of distribution networks, if edge agents communicate too frequently, resource exhaustion may occur, posing a serious threat to the stable operation of the system. Figure 9 shows the number of event triggers for each edge agent during the distributed control process at the edge.

When event-driven control is adopted, the number of triggers for the five edge agents is 2119, 2147, 1907, 1942, and 2006, respectively. Without event-driven control, each edge agent is triggered 3251 times. The comparison shows that the trigger counts at each node are reduced by 34.82%, 33.96%, 41.34%, 40.26%, and 40.26%, respectively. When Equation (17) is not satisfied, the edge agent does not trigger the task, significantly reducing the number of activations. This demonstrates that edge-based distributed control effectively reduces unnecessary task triggering, improves the efficiency of concurrent task execution, and achieves better real-time control performance.

6.4.4. Dispatch Efficiency Comparison

Using traditional centralized control as a baseline, the efficiency of the proposed edge-based distributed control is compared, as shown in

Table 2. Comparison of time consumption across control method stages.

Control Method	Consumed Time/ms
	Data Upload	Control Computation	Command Issuance
Edge-based distributed control	6.1	5.0	6.3
Centralized control	25.4	3.0	24.9

. It can be observed that, compared to centralized control, the edge-based distributed approach significantly reduces communication latency by leveraging the geographically proximal deployment of edge agents, thereby shortening the data upload and command transmission paths. Although distributed control introduces additional parallel processing overhead due to multi-node coordination, its localized decision-making mechanism avoids the data aggregation bottleneck inherent in centralized systems. Overall, the edge-based distributed control method responds more quickly to changes in distribution network conditions and demonstrates superior global responsiveness in dynamic scenarios.

Furthermore, an economic dispatch control strategy for a distribution network station with a larger number of nodes is constructed and computed. As the number of nodes increases, the computation times required for strategy execution using the traditional centralized control method and the proposed edge-based distributed control method are shown in

Table 3. Comparison of calculation times of two method strategies.

Control Method	Computation Time/ms
	5 Nodes	14 Nodes	39 Nodes
Edge-based distributed control	5.0	5.0	5.0
Centralized control	3.0	12.0	34.0

.

The computation time increases when using the centralized control method, while it remains nearly constant when using the edge-based distributed control method. This is because, in edge-based distributed control, each edge agent only communicates with directly connected agents, resulting in smaller computation scales and immunity to increasing network complexity. In contrast, with the centralized control method, as the number of nodes increases, both the communication network and the computation scale expand, leading to longer strategy computation times.

7. Discussion

In recent years, edge agent technology has been gradually applied in distributed control of power distribution networks, but several key challenges remain. First, most existing edge control schemes are designed for specific scenarios and thus struggle to accommodate the broad adaptability required for managing heterogeneous multi-source resources. Second, many methods lack a structured representation of control logic, making it difficult to support modular configuration and reuse. Moreover, in typical emerging scenarios, edge nodes must respond to frequent state changes, which can lead to communication congestion and compromise system stability under resource-constrained conditions.

To address these challenges, this paper proposes a control configuration method oriented toward modular deployment. Specifically, an activity-on-edge model is employed to abstract control logic, explicitly representing the execution sequence and triggering relationships of control flows. By incorporating hybrid control theory, discrete events and continuous dynamics are unified within a structured framework. This approach improves the reusability and adaptability of control logic across diverse application scenarios.

At the execution level, a web-based edge agent architecture is adopted, integrating event-driven control with automatic differentiation-based optimization. The event-driven mechanism activates control only when predefined conditions are met, effectively improving responsiveness and resource efficiency in dynamic environments. Automatic differentiation avoids symbolic expression expansion and enables fast computation of high-dimensional gradients. By introducing web application technology, the frontend and backend of the system are decoupled to improve module reuse and execution efficiency. In addition, asynchronous communication enables non-blocking data exchange between agents, ensuring timely information delivery and supporting the real-time responsiveness required in distributed control systems.

Despite these advantages, the proposed method still has some limitations. The current framework assumes relatively stable edge hardware and communication conditions, while in real-world deployments, device heterogeneity and unstable communication links may affect coordination among agents. Furthermore, the event-driven mechanism is sensitive to parameter settings, and inappropriate parameter selection may lead to delayed responses or frequent triggering, negatively affecting overall system performance.

Looking ahead, the proposed method offers strong scalability and integration potential. The edge agent architecture can be further integrated with IoT-based sensing platforms or extended to coordinate flexible loads such as electric vehicles with dynamic charging behavior. Future work will focus on validating the proposed method in more complex and diverse distribution network scenarios to further improve and refine its performance.

8. Conclusions

This paper proposes a distributed edge agent technology for distribution networks with several key contributions. First, a highly adaptable edge agent control architecture is designed based on hybrid control theory, considering the diverse requirements of distributed control in distribution networks. Furthermore, a control configuration technology suitable for edge agents is proposed, enabling standardized configuration of distribution network tasks. This facilitates fast and lightweight deployment, as well as flexible updates, of event-driven control strategies. In addition, by taking economic optimization dispatch at the station level as a representative edge computing application, the proposed edge-based distributed control method is compared with traditional centralized control. The results validate the effectiveness of the proposed approach, which demonstrates advantages such as low communication latency and reduced computation scale, making it suitable for large-scale distribution network control scenarios