A power supply power energy intelligent scheduling method and system based on an internet of things

Abstract

This invention relates to the field of power dispatching technology, and more particularly to a power dispatching method and system based on the Internet of Things (IoT). This method collects real-time data on photovoltaic, energy storage, and load data from industrial parks, constructs an energy coordination topology using a graph neural network, and reveals the coupling relationships between nodes. Based on this, a coordination dispatching strategy is generated, and spatiotemporal simulation is performed on a digital twin platform to detect power flow conflicts and mark constrained regions. Deep reinforcement learning is used to optimize conflict events, trace strategy defects, reconstruct the reward function, and output control parameter compensation values. Finally, the constrained regions and compensation parameters are fused to generate an instruction set and issue it for execution. This method achieves refined perception and intelligent closed-loop control of microgrids, effectively improving energy utilization efficiency and system operational stability.

Description

Technical Field

[0001] This invention relates to the field of power dispatching technology, and in particular to a power intelligent dispatching method and system based on the Internet of Things. Background Technology

[0002] Industrial parks, as key areas of energy consumption and carbon emissions, increasingly urgently require intelligent and low-carbon operation of their internal microgrid systems. Industrial parks typically integrate various energy elements such as photovoltaic power generation, energy storage systems, production loads, and auxiliary facilities, characterized by strong fluctuations on both the source and load sides, complex equipment coupling relationships, and high difficulty in real-time scheduling. Traditional power dispatching methods are mostly based on centralized optimization models, relying on static grid parameters and idealized prediction data. These methods struggle to effectively address the uncertainties in distributed energy output and the dynamic changes in load demand, leading to delayed dispatching decisions, low energy utilization, and even safety issues such as line overload and voltage exceeding limits.

[0003] In recent years, the widespread application of Internet of Things (IoT) technology has made it possible to achieve refined sensing and real-time control of microgrids. By deploying multimodal sensors and edge computing nodes, high-precision, low-latency acquisition of key parameters such as photovoltaic output, energy storage status, and load changes can be achieved, forming multi-source heterogeneous data streams. However, existing scheduling systems still have significant bottlenecks in data fusion and intelligent decision-making: on the one hand, there is a lack of deep modeling of the spatial correlation and behavioral coordination among distributed energy sources, making it difficult to uncover the implicit coupling relationships between nodes; on the other hand, the scheduling strategy generation process is disconnected from the dynamic response of the physical power grid, and there is a lack of a forward-looking verification mechanism for the effectiveness of control commands, causing the optimization strategy to be susceptible to failure due to factors such as equipment response delays and communication interference during actual execution.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a power intelligent scheduling method and system based on the Internet of Things, aiming to solve the technical problems of existing power scheduling methods being unable to effectively cope with the strong volatility and complex equipment coupling of distributed energy in industrial park microgrids, and having weak data fusion capabilities, insufficient collaborative relationship modeling, poor security and low energy efficiency caused by the disconnect between scheduling decisions and physical responses.

[0006] To achieve the above objectives, the present invention provides a power supply intelligent scheduling method based on the Internet of Things, the method comprising:

[0007] Real-time acquisition of photovoltaic power output data, energy storage system status data, and load demand data in industrial park microgrid scenarios, and output of raw data stream;

[0008] The raw data stream is input into a pre-trained graph neural network, and an energy collaborative topology graph is constructed by analyzing the spatial topological relationships, output correlations and energy transaction record characteristics of distributed energy nodes.

[0009] Based on the energy collaboration topology diagram, a dynamic electricity price response mechanism is triggered to generate collaborative scheduling strategies for distributed energy nodes with power coupling relationships;

[0010] A digital twin of the industrial park microgrid is constructed on the digital twin platform. The collaborative scheduling strategy is imported into the digital twin environment for spatiotemporal pre-simulation, and power flow conflict points are detected in real time and the parameter set of grid constraint conflict area is marked.

[0011] By optimizing conflict events in pre-simulation using deep reinforcement learning algorithms, the source of control strategy defects is traced and the reward function space is reconstructed, and the compensation value of control parameters is output.

[0012] By integrating the parameter set of conflicting areas in the power grid with the compensation values ​​of control parameters, a set of distributed energy control instructions is generated and executed.

[0013] Optionally, the load demand data includes industrial production load curves and public utility electricity consumption records.

[0014] Optionally, the specific steps for constructing the energy synergy topology map are as follows:

[0015] Analyze the distributed energy node information in the raw data stream to extract node spatial coordinates, output fluctuation coefficients, and energy trading frequency;

[0016] The frequency of energy transactions is smoothed by an index to generate transaction correlation coefficients, and the spatial coordinates are calculated by inverse power grid impedance to generate electrical distance weights.

[0017] The power output fluctuation coefficient is used as the initial weight of the graph edge. The transaction correlation coefficient and the electrical distance weight are integrated to generate a dynamic energy weight matrix. The energy node features are aggregated through a graph attention network to output an energy collaborative topology graph.

[0018] Optionally, the specific steps for generating a cooperative scheduling strategy for distributed energy nodes with power coupling relationships are as follows:

[0019] When the output of any photovoltaic node in the energy collaboration topology graph fluctuates abnormally, the state of all energy storage systems connected to the fluctuating node is traversed to obtain the state of charge value and the charging and discharging power limit.

[0020] If the state of charge of the energy storage system is lower than the threshold, a power compensation command will be generated to forcibly allocate the output of the associated photovoltaic node to the energy storage system;

[0021] If the state of charge of the energy storage system is higher than the threshold, calculate the maximum discharge power of the corresponding energy storage system and generate a virtual output curve that includes power ramp-up rate constraints.

[0022] Based on the dynamic weight values ​​of the connecting edges in the energy coordination topology graph, power adjustment coefficients are assigned to energy nodes constrained by virtual output curves, and a coordination scheduling strategy is output.

[0023] Optionally, the specific steps for real-time detection of power flow conflict points and marking of the power grid constraint conflict region parameter set are as follows:

[0024] A dynamic digital twin is constructed by loading the industrial park's power grid topology model and real-time measurement data stream onto a digital twin platform.

[0025] The collaborative scheduling strategy is imported into the dynamic digital twin, and a power flow spatiotemporal simulation is performed to calculate the line overload rate, node voltage offset, and frequency fluctuation amplitude.

[0026] Activate the power system simulation engine to simulate short-circuit fault effects and voltage sag processes, and mark the parameter set of grid constraint conflict areas that exceed the safety threshold.

[0027] Optionally, the compensation value for the output control parameter is determined through the following steps:

[0028] The parameter set of the power grid constraint conflict area is combined with the parameters of the digital twin environment to generate a state-action pair sample set, which is used as input data for the deep reinforcement learning algorithm;

[0029] The value function of the state-action pair sample set is calculated by the deep deterministic policy gradient algorithm, and the reward function corresponding to the optimal control policy is optimized.

[0030] The study analyzes the weight allocation defects of the reward function, identifies the voltage regulation weights and frequency response penalty coefficients that cause conflicts, and reconstructs the spatial structure of the reward function.

[0031] Inject device aging simulation noise and communication delay interference into the state-action pair sample set to generate adversarial disturbance samples, and train the policy network model to output control parameter compensation values.

[0032] Optionally, the specific steps for generating the distributed energy control instruction set and executing the instruction issuance are as follows:

[0033] Based on the control parameter compensation values, calculate the PWM modulation duty cycle of the energy storage converter and generate converter control commands;

[0034] Based on the parameter set of grid-constrained conflict regions, the output curve of the photovoltaic inverter is replanned using the particle swarm optimization algorithm, and the output power reference value sequence is generated.

[0035] The converter control commands and power reference value sequences are aligned with timestamps to generate structured command data packets. The converter control commands are sent to the energy storage controller via industrial Ethernet, while the power reference value sequences are transmitted to the photovoltaic inverter via a wireless sensor network.

[0036] Furthermore, to achieve the above objectives, the present invention also provides an Internet of Things-based intelligent power dispatching system, the system comprising:

[0037] The data acquisition module is used to collect photovoltaic power output data, energy storage system status data and load demand data in real time in the microgrid scenario of the industrial park, and output raw data stream;

[0038] The topology modeling module is used to input the raw data stream into a pre-trained graph neural network and construct an energy collaboration topology graph by analyzing the spatial topological relationships, output correlations, and energy transaction record characteristics of distributed energy nodes.

[0039] The collaborative decision-making module is used to trigger a dynamic electricity price response mechanism based on the energy collaboration topology diagram and generate collaborative scheduling strategies for distributed energy nodes with power coupling relationships.

[0040] The simulation pre-simulation module is used to build a digital twin of the industrial park microgrid on the digital twin platform, import the collaborative scheduling strategy into the digital twin environment for spatiotemporal pre-simulation, detect power flow conflict points in real time, and mark the parameter set of grid constraint conflict areas.

[0041] The reinforcement optimization module is used to optimize conflict events in the pre-simulation using deep reinforcement learning algorithms, trace the source of control strategy defects and reconstruct the reward function space, and output control parameter compensation values.

[0042] The instruction execution module is used to integrate the parameter set of the grid constraint conflict area with the control parameter compensation value, generate the distributed energy control instruction set, and execute and issue the instructions.

[0043] Furthermore, to achieve the above objectives, the present invention also provides an Internet of Things (IoT)-based intelligent power scheduling device, the device comprising: a memory, a processor, and an IoT-based intelligent power scheduling program stored in the memory and executable on the processor, the IoT-based intelligent power scheduling program being configured to implement the steps of the IoT-based intelligent power scheduling method as described above.

[0044] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an Internet of Things (IoT)-based intelligent power scheduling program, wherein when the IoT-based intelligent power scheduling program is executed by a processor, it implements the steps of the IoT-based intelligent power scheduling method as described above.

[0045] This invention provides an IoT-based intelligent power dispatching method. By integrating multimodal IoT sensing and edge computing capabilities, the method achieves high-precision, low-latency data acquisition of photovoltaic output, energy storage status, and load demand in industrial park microgrids, improving system observability. It utilizes graph neural networks to construct an energy collaborative topology graph, deeply exploring spatial correlations, output correlations, and transaction behavior characteristics among distributed energy nodes, enhancing the ability to model complex coupling relationships. Combined with a digital twin platform, it performs spatiotemporal pre-simulation of collaborative dispatching strategies, enabling early identification of power flow conflicts and grid constraint exceedance risks in a virtual environment, significantly improving the safety and foresight of dispatching decisions. Through deep reinforcement learning, it performs source analysis of simulated conflict events and dynamically optimizes the reward function, achieving autonomous evolution and parameter compensation of the control strategy, enhancing the system's adaptability to uncertain disturbances. Finally, it integrates physical constraint information and intelligent compensation parameters to generate closed-loop control commands, achieving precise matching between dispatching strategies and equipment responses. The overall solution forms a closed-loop intelligent scheduling system encompassing perception, modeling, simulation, optimization, and execution, effectively improving the energy efficiency, operational stability, and intelligence level of industrial park microgrids. It has good practical value and promising prospects for widespread application. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating an embodiment of the Internet of Things-based intelligent power scheduling method of the present invention.

[0047] Figure 2 This is a structural block diagram of an embodiment of the Internet of Things-based intelligent power dispatching system of the present invention.

[0048] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0049] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0050] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the IoT-based intelligent power scheduling method of the present invention, which presents an embodiment of the IoT-based intelligent power scheduling method of the present invention.

[0051] In one embodiment, the IoT-based intelligent power dispatching method includes:

[0052] Step S100: Real-time acquisition of photovoltaic output data, energy storage system status data, and load demand data in the industrial park microgrid scenario, and output of raw data stream.

[0053] Photovoltaic output data can be the actual power output of photovoltaic power generation units at a specific point in time or within a specific time period. It can reflect the real-time power generation capacity of distributed power sources and be used for scheduling decisions and state estimation. In an exemplary embodiment, photovoltaic output data can be collected and uploaded through current and voltage sensors and communication modules deployed at photovoltaic inverters or combiner boxes. Energy storage system status data can be a multi-dimensional set of parameters describing the current operating status of energy storage devices, including state of charge (SOC), charging and discharging power, temperature, and health status. It can be used to support feasibility assessments and capacity planning for energy storage participation in scheduling. For example, energy storage system status data can be collected by a battery management system (BMS) through internal sensors and uploaded through an edge gateway. Load demand data can be the actual or predicted power demand of various electrical devices or areas within an industrial park at a specific time. It can be used as an important input for scheduling objectives to balance supply and demand. Furthermore, load demand data can be collected in real time through smart meters, load monitoring terminals, or SCADA systems.

[0054] The raw data stream can be an unprocessed time-series data set synchronously collected and transmitted by multimodal IoT sensing nodes, which can be used to provide a high-fidelity, low-latency input source for subsequent modeling and analysis. In a specific embodiment, the raw data stream can be divided into photovoltaic output stream, energy storage status stream, and load demand stream according to data source; and into second-level stream, minute-level stream, and event-triggered stream according to time granularity. Real-time collection of photovoltaic output data, energy storage system status data, and load demand data in an industrial park microgrid scenario, and output of the raw data stream, can be achieved by synchronously collecting and aggregating data from multimodal sensors and edge computing nodes deployed in various energy units of the park. Furthermore, this operation can be achieved by uploading multi-source sensors after nanosecond-level alignment using a time synchronization protocol (such as PTP), or by locally caching and compressing the data on edge nodes and then pushing it in batches to the central platform using MQTT or CoAP protocols, thereby achieving high-precision, low-latency status perception and improving system observability.

[0055] Step S200: Input the raw data stream into the pre-trained graph neural network, and construct an energy collaborative topology graph by analyzing the spatial topological relationships, output correlations and energy transaction record characteristics of distributed energy nodes.

[0056] Graph neural networks (GNNs) can be deep learning models capable of processing graph-structured data. They aggregate neighboring node information through message passing mechanisms to learn node embedding representations, and can be used for end-to-end modeling of complex coupling relationships between distributed energy nodes. In an exemplary embodiment, the GNN can be pre-trained offline based on historical operating data and topology, and deployed on an edge or cloud inference engine. Distributed energy nodes can be physical or logical units in a microgrid with independent power generation, energy storage, or controllable load capabilities. They can serve as basic building blocks of the energy coordination topology graph, carrying energy and information interaction. Furthermore, distributed energy nodes can include functional types such as photovoltaic nodes, energy storage nodes, and flexible load nodes. Spatial topology relationships can be the adjacency and hierarchical relationships of distributed energy nodes in the physical power grid connection structure. They can be used to reflect electrical connection constraints and power flow propagation paths, forming the basis for constructing the graph structure. In a specific embodiment, spatial topology relationships can serve as the input edge structure of the GNN, jointly determining the message passing weights between nodes along with output correlation.

[0057] Output correlation can be the statistical dependence or causal relationship of output power changes of different distributed energy nodes over time. It can be used to reveal implicit coupling relationships that exist in non-physical connections but with behavioral coordination. For example, output correlation can include positive correlation, negative correlation, and lagged correlation. Energy transaction record features can be patterned information reflected in the historical electricity transaction behavior between nodes, such as transaction frequency, electricity volume, and price sensitivity. They can be used to help identify the tendency of coordinated behavior driven by the economy and enhance the market adaptability of strategies. Furthermore, energy transaction record features can include behavioral dimensions such as transaction activity features, price response features, and contract performance features. The energy coordination topology graph can be a weighted directed graph with distributed energy nodes as vertices, integrating spatial topology, output correlation, and transaction behavior features. It can be used to uniformly represent the multidimensional coupling relationship between physical connections, dynamic behavior, and economic interaction. In an exemplary embodiment, the energy coordination topology graph can be generated by a graph neural network through feature extraction and graph structure reconstruction of the original data stream, and used as the input basis for the dynamic electricity price response mechanism to drive the generation of coordinated scheduling strategies.

[0058] By inputting the raw data stream into a pre-trained graph neural network, and analyzing the spatial topological relationships, output correlations, and energy transaction record characteristics of distributed energy nodes, an energy collaborative topology graph is constructed. This can be achieved by the graph neural network receiving node features and edge structure information, generating node embeddings, and reconstructing graph weights through multi-layer message passing. Furthermore, this operation can be implemented by dynamically assigning weights to different feature dimensions using a graph attention mechanism to highlight key coupling paths, or by introducing a graph contrastive learning framework to enhance the generalization ability of the topology graph under unsupervised conditions. This can reveal implicit coupling relationships that are difficult to capture using traditional methods, thereby enhancing the modeling capabilities of complex systems.

[0059] Step S300: Based on the energy collaboration topology diagram, a dynamic electricity price response mechanism is triggered to generate a collaborative scheduling strategy for distributed energy nodes with power coupling relationships.

[0060] The dynamic electricity price response mechanism can be a rule engine that automatically adjusts dispatch instructions based on real-time electricity price signals and node coordination relationships. It can be used to transform economic incentives into physical control actions, promoting source-load-storage coordinated optimization. For example, the dynamic electricity price response mechanism can employ response logics such as threshold triggering, gradient adjustment, and game-theoretic equilibrium. Power coupling relationships can be the mutual influence of power output among multiple distributed energy nodes due to electrical connections or control strategies. This can be used to define the set of nodes requiring coordinated dispatch, avoiding global imbalances caused by local optimization. In a specific embodiment, power coupling relationships can be identified and quantified by an energy coordination topology diagram, serving as the target of the dynamic electricity price response mechanism. Coordinated dispatch strategies can be joint power allocation and control schemes formulated for node groups with power coupling relationships. These can be used to achieve multi-entity coordinated operation, improving overall energy efficiency and stability. Furthermore, coordinated dispatch strategies can include optimization objectives such as economic optimization, safety priority, and minimum carbon emissions.

[0061] Based on the dynamic electricity price response mechanism triggered by the energy collaboration topology diagram, a collaborative scheduling strategy can be generated for distributed energy nodes with power coupling relationships. This can be achieved by identifying strongly coupled node groups based on the topology diagram, calculating the joint optimization objective in conjunction with real-time electricity price signals, and solving for the scheduling scheme. Furthermore, this operation can be implemented by employing distributed optimization algorithms, such as ADMM, to iteratively solve for collaborative strategies among nodes, or by constructing a Stackelberg game model based on game theory to solve for equilibrium scheduling schemes. This allows for the generation of scheduling strategies with physical collaborative logic, improving the rationality and executability of the strategies.

[0062] Step S400: Construct a digital twin of the industrial park microgrid on the digital twin platform, import the collaborative scheduling strategy into the digital twin environment for spatiotemporal pre-simulation, detect power flow conflict points in real time, and mark the parameter set of grid constraint conflict areas.

[0063] The digital twin platform can be a virtualized runtime environment integrating a physical model, a data-driven model, and a simulation engine. It can be used to support closed-loop verification and risk prediction of scheduling strategies in virtual space. For example, the digital twin platform can adopt architectures such as cloud-native twin platforms, edge twin platforms, and hybrid twin platforms. The digital twin of an industrial park microgrid can be a high-fidelity dynamic mirror model of the actual industrial park microgrid built in the digital twin platform. It can be used to reproduce the electromagnetic transients, electromechanical transients, and control logic behavior of the real system. In a specific embodiment, the digital twin of the industrial park microgrid can be modeled online synchronously based on equipment parameters, topology, control strategies, and real-time data streams. Spatiotemporal pre-simulation can dynamically extrapolate the execution process of scheduling strategies over a future period within the digital twin. It can be used to expose potential operational conflicts in advance, avoiding direct policy issuance that could lead to physical system overruns. Furthermore, spatiotemporal pre-simulation can include simulation granularities such as electromagnetic transient simulation, electromechanical transient simulation, and steady-state power flow simulation.

[0064] Power flow conflict points can be the specific locations and times where line overload, transformer overcapacity, or node voltage exceeding limits occur during pre-simulation. They can be used to identify areas where scheduling strategies are infeasible, serving as optimization feedback signals. For example, power flow conflict points can include conflict types such as thermal stability conflict points, voltage stability conflict points, and protection coordination conflict points. The grid constraint conflict region parameter set can be a set of electrical parameters describing the region where the power flow conflict point is located, including line impedance, node voltage limits, and equipment capacity. It can be used to quantify physical constraint boundaries and guide control parameter compensation. In an exemplary embodiment, the grid constraint conflict region parameter set can be fused with control parameter compensation values ​​to generate the final control command, ensuring that the strategy meets physical feasibility requirements.

[0065] A digital twin of an industrial park microgrid is constructed on a digital twin platform. The collaborative scheduling strategy is imported into the digital twin environment for spatiotemporal simulation, real-time detection of power flow conflict points, and marking of grid constraint conflict region parameter sets. This can be achieved by applying the scheduling strategy as input excitation to the digital twin, running electromagnetic-electromechanical hybrid simulation, and monitoring constraint exceedances. Furthermore, this operation can be implemented by using hardware-in-the-loop (HIL) to connect to the actual controller to verify the impact of communication and execution delays, or by using a reduced-order model to accelerate simulation and shorten simulation time while maintaining accuracy. This allows for the early identification of operational risks in a virtual environment, enabling proactive verification of scheduling decisions.

[0066] Step S500: Optimize the conflict events in the pre-simulation using a deep reinforcement learning algorithm, trace the source of control strategy defect features and reconstruct the reward function space, and output the control parameter compensation value.

[0067] The deep reinforcement learning algorithm can be an adaptive decision-making algorithm combining deep neural networks and a reinforcement learning framework. Through trial and error optimization strategies, it can be used to perform attribution analysis on simulated conflict events and dynamically adjust the policy generation mechanism. For example, deep reinforcement learning algorithms can employ Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC) algorithms. Conflict events can be specific abnormal situations in spatiotemporal pre-simulation that violate power grid operation constraints, and can be used as negative reward signals for deep reinforcement learning to drive policy evolution. Furthermore, conflict events can include cause types such as equipment response delay conflicts, communication packet loss conflicts, and prediction bias conflicts. Control policy defect features can be inherent deficiencies in the scheduling policy that lead to conflict events, such as response lag, gain mismatch, and coupling neglect, and can be used to provide attribution basis for reward function reconstruction, improving policy robustness. In a specific embodiment, control policy defect features can include defect types such as timing mismatch features, amplitude overshoot features, and lack of coordination features.

[0068] The reward function space can be a set of multi-dimensional objective functions used in deep reinforcement learning to evaluate the merits of policies. It can guide agents to learn scheduling behaviors that balance safety, economy, and stability. In an exemplary embodiment, the reward function space can be dynamically reconstructed driven by control policy defect features, influencing the generation direction of control parameter compensation values. Control parameter compensation values ​​can be incremental adjustments calculated to correct scheduling policy defects, used to correct the original control commands, and can improve the adaptability and accuracy of the policy in actual execution. Furthermore, the control parameter compensation values ​​can be fused with the parameter set of the grid constraint conflict region to generate the final control command, forming a closed-loop feedback.

[0069] By optimizing conflict events in pre-simulation using deep reinforcement learning algorithms, tracing the source of control policy defects and reconstructing the reward function space, and outputting compensation values ​​for control parameters, this can be achieved by using conflict events as negative samples to train the reinforcement learning agent, analyzing the reasons for policy failure, adjusting the reward function, and generating compensation parameters. Furthermore, this operation can be further refined by employing a meta-learning framework to quickly adapt to new types of conflicts to improve generalization compensation capabilities, or by combining causal inference models to identify the root causes of policy defects to avoid bias in compensation direction. This enables autonomous evolution of the control policy and parameter compensation, enhancing the system's adaptability to uncertain disturbances.

[0070] Step S600: Integrate the parameter set of the grid constraint conflict area with the compensation value of the control parameter to generate a distributed energy control instruction set and execute the instruction issuance.

[0071] The distributed energy control instruction set can be a set of instructions for specific power setpoints, switching states, or operating modes for each distributed energy node. It can be used to directly drive physical devices to execute scheduling strategies, achieving closed-loop control. For example, the distributed energy control instruction set can include active power instructions, reactive power instructions, and start / stop control instructions. By integrating the parameter set of grid constraint conflict areas with control parameter compensation values, generating the distributed energy control instruction set, and issuing the instructions, the physical constraint boundaries and intelligent compensation quantities can be jointly encoded to generate final control instructions that meet safety and performance requirements. Furthermore, this operation can be achieved by using a model predictive control (MPC) framework for online rolling optimization of the instruction sequence, or by using digital signatures and encrypted channels to ensure the security and integrity of the instruction issuance. This allows for precise matching of scheduling strategies and device responses, ensuring the effectiveness of closed-loop control.

[0072] Taking the microgrid dispatching of an industrial park with a high proportion of photovoltaic access as an example, the IoT-based intelligent power dispatching method in this embodiment can be as follows: An industrial park has deployed rooftop photovoltaics, energy storage power stations, and multiple production line loads. At 10:00 AM, a sudden change in cloud cover caused a sharp drop in photovoltaic output by 30%, and at the same time, a production line started high-power equipment. The IoT sensing layer collects three types of data in real time and forms a raw data stream; the graph neural network identifies a strong power coupling relationship between the production line and adjacent energy storage nodes; the dynamic electricity price response mechanism triggers energy storage discharge to mitigate load impact; the collaborative dispatching strategy is imported into a digital twin for a 5-minute pre-simulation, and the simulation finds that a 10kV feeder will be overloaded; deep reinforcement learning analyzes that the conflict originates from the energy storage response delay, reconstructs the reward function, and outputs compensation values; finally, the control command integrates the line thermal stability parameters and compensation values, instructing the energy storage to start 0.5 seconds in advance and operate with current limiting, successfully avoiding physical system over-limit.

[0073] In one embodiment, the load demand data includes industrial production load curves and utility electricity consumption records.

[0074] The industrial production load curve can describe the power consumption trajectory of production equipment within an industrial park over time, reflecting the energy demand patterns of the production process. It can be used to characterize highly volatile, time-dependent load behavior, supporting precise scheduling and coordinated energy storage response. In an exemplary embodiment, the industrial production load curve can be obtained from total power consumption data through a non-intrusive load monitoring algorithm built into the edge computing node. Furthermore, the industrial production load curve can include, but is not limited to, one or more of the following: continuous process load curve, intermittent start-stop load curve, and multi-shift cyclic load curve.

[0075] Public facility electricity consumption records can be a collection of electricity consumption data generated during the operation of auxiliary systems such as lighting, air conditioning, elevators, and security systems within the park. These records can be used to characterize basic, low-fluctuation, but wide-coverage electricity consumption behaviors, and are used for baseload modeling and energy efficiency management. For example, public facility electricity consumption records can be obtained by tagging and categorizing smart meter data through rule matching based on equipment ledgers and an electricity consumption feature database. In a specific embodiment, public facility electricity consumption records may include, but are not limited to, environmental regulation electricity consumption records, security monitoring electricity consumption records, and transportation service electricity consumption records.

[0076] Incorporating industrial production load curves and public facility electricity consumption records into load demand data can be achieved during the initial data stream construction phase by classifying, labeling, and structurally storing the collected load data according to the type of electricity user. Furthermore, this can be accomplished through load decomposition algorithms (such as non-intrusive load monitoring) built into edge computing nodes to automatically identify and separate production and public facility electricity consumption components, or by using rule matching based on equipment ledgers and electricity consumption feature databases to tag and categorize smart meter data before uploading. This approach can improve the semantic granularity and modeling accuracy of load demand data, enhancing the subsequent ability of graph neural networks to identify source-load coupling relationships.

[0077] Taking the refined load scheduling of a multi-production-line industrial park as an example, the IoT-based intelligent power scheduling method in this embodiment can be as follows: An electronics manufacturing park includes three 24-hour shift production lines and supporting office buildings, parking lots, and lighting systems. The edge gateway identifies the peak load curve (industrial production load curve) formed by the start and stop of the stamping machine through non-intrusive load monitoring, and at the same time separates the nighttime off-peak operation record of the air conditioning system (public facility electricity consumption record). The two types of data are labeled separately and then merged into the original data stream; the graph neural network identifies the strong coupling relationship between the stamping machine and the energy storage system, while the air conditioning load has a weak positive correlation with the midday output of photovoltaic power; in the digital twin simulation, if only the total load data is used, it will mask the power gap at the moment of the stamping machine start-up, while the subdivided data allows the simulation to accurately reproduce this transient process, thereby triggering early energy storage discharge compensation and avoiding voltage dips.

[0078] In one embodiment, an energy synergy topology graph is constructed, and the specific steps are as follows:

[0079] The distributed energy node information in the raw data stream is analyzed to extract the node spatial coordinates, output fluctuation coefficients, and energy trading frequency.

[0080] The distributed energy node information can be a set of structured attributes of each distributed energy unit contained in the original data stream, which can be used as the original feature input for constructing an energy collaborative topology graph. Furthermore, the distributed energy node information can include, but is not limited to, one or more of the following: location information, operating status information, and transaction behavior information. Node spatial coordinates can be two-dimensional or three-dimensional geographic or electrical location identifiers of distributed energy nodes in the physical layout of the industrial park microgrid, which can be used to calculate the electrical distance between nodes and reflect the tightness of physical connections. In an exemplary embodiment, node spatial coordinates can be obtained through equipment ledger data or a GIS system and mapped to the power grid topology model. The output fluctuation coefficient can be a dimensionless statistical indicator characterizing the degree of drastic change in the output power of a distributed energy node over time, which can be used to quantify the uncertainty of node operation and as the initial weight of graph edges to reflect dynamic coupling potential. For example, the output fluctuation coefficient can use the standard deviation normalization coefficient, the coefficient of variation, the sliding window volatility, etc. The energy trading frequency can be the number of times a distributed energy node participates in electricity trading per unit time, which can be used to reflect the node's market activity and to derive the transaction correlation coefficient. In one specific embodiment, energy trading frequency may include day-ahead market frequency, real-time market frequency, peer-to-peer trading frequency, etc.

[0081] Parsing distributed energy node information in the raw data stream and extracting node spatial coordinates, output fluctuation coefficients, and energy trading frequencies can be achieved by performing structured parsing of the raw data stream and extracting three types of key attributes according to predefined fields. Furthermore, this operation can be implemented by using a streaming ETL pipeline to perform real-time field mapping and feature extraction on the raw data stream, or by using a local preprocessing module at the edge nodes to perform initial feature screening before uploading to the central modeling unit. This achieves the technical effect of standardizing the transformation of multi-source heterogeneous data into the features required for graph modeling.

[0082] The frequency of energy transactions is smoothed by an index to generate transaction correlation coefficients, and the spatial coordinates are calculated by inversely proportional to the power grid impedance to generate electrical distance weights.

[0083] The transaction correlation coefficient can be a continuous value representing the strength of long-term transaction interactions between nodes, obtained by exponentially smoothing the frequency of energy transactions. It can be used to transform discrete transaction behaviors into continuous correlation weights, enhancing the graph structure's ability to express economic synergy. In this embodiment, the transaction correlation coefficient can be filtered using a first-order or higher-order exponential smoothing algorithm on the historical transaction frequency sequence. The electrical distance weight can be a normalized weight reflecting the strength of electrical coupling, calculated based on the node's spatial coordinates and grid impedance parameters. It can be used to transform physical location into electrical proximity, correcting graph edge weights. Furthermore, the electrical distance weight can be obtained by summing the line impedances along the paths between nodes, taking the reciprocal, and then normalizing it.

[0084] Applying exponential smoothing to energy trading frequency to generate trading correlation coefficients can be achieved by filtering historical trading frequency sequences using an exponential smoothing algorithm to generate smoothed continuous correlation strength values. Furthermore, this operation can be implemented by dynamically adjusting the filtering strength based on trading volatility using an adaptive smoothing factor, or by introducing double exponential smoothing to simultaneously capture changes in trading levels and trends. This can suppress random noise in trading behavior and highlight long-term synergistic trends. Calculating electrical distance weights by inversely proportionalizing grid impedance to spatial coordinates can be done by calculating the equivalent impedance between nodes based on grid topology and line parameters, and taking its reciprocal as the electrical proximity. Further, this operation can be achieved by using a power flow tracing algorithm to determine the actual power path before calculating the impedance, or by combining short-circuit capacity indicators to normalize and correct the electrical distance. This can transform geometric distance into coupling weights with electrophysical significance.

[0085] The power output fluctuation coefficient is used as the initial weight of the graph edge. The transaction correlation coefficient and the electrical distance weight are integrated to generate a dynamic energy weight matrix. The energy node features are aggregated through a graph attention network to output an energy collaborative topology graph.

[0086] The initial weights of the graph edges can be the initial association strength values ​​of the edges in the graph structure, set based on the output fluctuation coefficient. This can provide a benchmark for subsequent multi-source weight fusion, reflecting the coupling tendency at the operational behavior level. Furthermore, the initial weights of the graph edges can be one of the foundational components of the dynamic energy weight matrix, fused with the transaction association coefficient and electrical distance weight. The dynamic energy weight matrix can be a variable graph adjacency matrix generated by fusing the output fluctuation coefficient, transaction association coefficient, and electrical distance weight, characterizing the multidimensional coupling strength between nodes. This can be used to provide edge weights with physical-behavioral-market semantics for the graph attention network, improving the topology graph's expressive power. In this embodiment, the dynamic energy weight matrix can be generated by nonlinearly fusing the three types of weights through a weighted linear combination or a multilayer perceptron. Energy node features can be multidimensional attribute vectors corresponding to each distributed energy node, including output characteristics, state parameters, and behavioral patterns. These can be used as message passing content for the graph attention network to generate node embedding representations. For example, energy node features can include operational features, topological features, and transaction features.

[0087] Using the power output fluctuation coefficient as the initial weight of the graph edges, a dynamic energy weight matrix is ​​generated by fusing the transaction correlation coefficient and electrical distance weight. This can be achieved by using the power output fluctuation coefficient as a base and weighting it with the transaction correlation coefficient and electrical distance weight to form the final edge weights. Furthermore, this operation can be implemented by automatically allocating the contribution ratio of the three types of weights using a learnable fusion gating mechanism, or by dynamically selecting weight combinations for different scenarios using an attention mechanism. This achieves the technical effect of constructing a multi-dimensional graph structure that combines operational volatility, market interaction, and electrical coupling. By aggregating energy node features through a graph attention network, an energy collaboration topology graph is output. This can be achieved by the graph attention network calculating the attention coefficients of neighboring nodes based on the dynamic energy weight matrix, and then weighting and aggregating features to generate node embeddings, ultimately outputting a weighted graph structure. Furthermore, this operation can be achieved by using a multi-head graph attention mechanism to capture different types of coupling relationships, or by introducing residual connections to retain original node features to avoid information dilution. This achieves the technical effect of adaptive focusing on highly correlated neighbors and enhancing the ability to identify implicit collaborative relationships.

[0088] Taking a multi-source heterogeneous coupling modeling scenario as an example, the IoT-based intelligent power dispatching method in this embodiment can be as follows: In a certain park, photovoltaic A and energy storage B are far apart but frequently engage in point-to-point transactions, while photovoltaic C and D are adjacent but their output fluctuations are highly synchronized. The system extracts the spatial coordinates, output fluctuation coefficients (A: 0.45, B: 0.12, C: 0.38, D: 0.37), and transaction frequencies of each node from the original data stream; exponential smoothing of the A and B transaction frequencies yields a transaction correlation coefficient of 0.82; the electrical distance weights of A and B are calculated to be 0.35 (due to high line impedance) and CD to be 0.92; using the output fluctuation coefficients as initial edge weights, a dynamic energy weight matrix is ​​generated after fusion, in which the edge weights of CD are significantly increased, while the edge weights of A and B are retained due to high transaction correlation; the graph attention network accordingly assigns higher attention to CD, while retaining the A and B collaborative path, and the final output energy collaborative topology map accurately reflects the two types of relationships: physical proximity collaboration and market-driven collaboration.

[0089] In one embodiment, a cooperative scheduling strategy is generated for distributed energy nodes with power coupling relationships. The specific steps are as follows:

[0090] When the output of any photovoltaic node in the energy collaboration topology graph fluctuates abnormally, the state of all energy storage systems connected to the fluctuating node is traversed to obtain the state of charge value and the charging and discharging power limit.

[0091] Among them, abnormal fluctuations in photovoltaic (PV) node output can be unplanned and drastic changes in the actual output power of a PV node deviating from the expected value within a short period of time, and can be used as an event signal to trigger the generation of collaborative scheduling strategies. In an exemplary embodiment, abnormal fluctuations in PV node output can include, but are not limited to, one or more of the sudden increase and sudden decrease types, and can also be classified by cause such as meteorological disturbance or equipment failure. The energy storage system status can be a set of operating parameters related to a specific energy storage unit, specifically referring to the associated energy storage status used to respond to PV fluctuations, and can be used to determine the capability boundary and response mode of energy storage to participate in collaborative scheduling. Furthermore, the energy storage system status can include, but is not limited to, response dimensions such as state of charge, power limit, and dynamic characteristics. The state of charge value can be the ratio of the current remaining electricity of the energy storage system to its rated capacity, expressed as a percentage, and can be used as a key threshold basis for judging whether the energy storage has charging or discharging capabilities. In a specific embodiment, the state of charge value can be calculated in real time by the battery management system and reported through the communication interface.

[0092] Charging and discharging power limits can be the maximum allowed charging or discharging power of an energy storage system under its current state. These limits can be used to constrain the amplitude of power compensation commands and ensure safe operation of the equipment. For example, charging and discharging power limits can be categorized in several ways, including but not limited to maximum charging power limits and maximum discharging power limits. When any photovoltaic node in the energy collaboration topology graph experiences abnormal power fluctuations, the states of all energy storage systems connected to the fluctuating node are traversed to obtain the state of charge (SOC) value and charging / discharging power limits. This can be achieved by monitoring the photovoltaic power output deviation and then retrieving associated energy storage units along the graph's adjacency relationships after the deviation exceeds a preset threshold, and reading their key state parameters. Furthermore, this operation can be accelerated by using graph database indexes to query adjacent energy storage nodes, or by deploying a lightweight graph traversal engine at the edge to achieve millisecond-level response times. This enables rapid resource location based on graph topology, avoiding the latency caused by global scanning.

[0093] If the state of charge of the energy storage system is lower than the threshold, a power compensation command will be generated to forcibly allocate the output of the associated photovoltaic node to the energy storage system;

[0094] The power compensation command can be a control command that forcibly redirects excess output from associated photovoltaic (PV) nodes to energy storage when the state of charge (SOC) of the energy storage is below a threshold. This can be used to avoid curtailment of solar power and improve energy storage utilization, achieving local source-storage balance. In one specific embodiment, the power compensation command can apply to the output of associated PV nodes and is subject to charging and discharging power limits. The output of associated PV nodes can be the output power of other PV nodes that have a connection edge with the currently abnormally fluctuating PV node in the energy collaboration topology graph. This output can be used as the object of power redistribution and can be forcibly scheduled to energy storage in low SOC scenarios. For example, the output of associated PV nodes can include, but is not limited to, connection strength classifications such as strong-association output and weak-association output.

[0095] If the state of charge (SOC) of the energy storage system falls below a threshold, a power compensation command is generated to forcibly allocate the output of the associated photovoltaic (PV) nodes to the energy storage system. This can be achieved by generating an active power transfer command with the energy storage system as the sink when the SOC falls below a set lower limit, prioritizing the absorption of fluctuating PV output. Furthermore, this operation can employ a proportional allocation strategy to distribute excess PV power to multiple low-SOC energy storage systems according to their connection weights, or dynamically adjust the allocation ratio based on line capacity constraints to avoid secondary overload. This can prevent curtailment of solar power, improve energy storage charging efficiency, and enhance local energy self-balancing capabilities.

[0096] If the state of charge of the energy storage system is higher than the threshold, calculate the maximum discharge power of the corresponding energy storage system and generate a virtual output curve that includes power ramp-up rate constraints.

[0097] The maximum discharge power can be the instantaneous maximum output power that the energy storage system can provide under high charge conditions, and can be used as an upper limit benchmark for constructing the virtual output curve. In an exemplary embodiment, the maximum discharge power can be dynamically calculated based on the current SOC, temperature, and BMS health assessment. The power ramp-up rate constraint can be the maximum allowable rate of power change of the energy storage system per unit time, and can be used to limit the slope of the virtual output curve to ensure that the command conforms to the dynamic response capability of the equipment. For example, the power ramp-up rate constraint can include, but is not limited to, categories such as charging ramp-up rate and discharging ramp-up rate. The virtual output curve can be a time-series output trajectory that simulates the power support it participates in, generated based on the maximum discharge power of the energy storage and the ramp-up rate constraint. It can be used to convert the energy storage capacity into equivalent power generation behavior, making it easier to incorporate into a unified scheduling framework. In a specific embodiment, the virtual output curve can construct a feasible trajectory in the power-time domain through piecewise linear interpolation or spline functions.

[0098] If the state of charge (SOC) of the energy storage system exceeds a threshold, the maximum discharge power of the corresponding energy storage system is calculated, generating a virtual output curve that includes power ramp-up constraints. This can be based on the physical limits of high SOC energy storage to construct an equivalent power generation trajectory that conforms to the dynamic characteristics of the equipment. Furthermore, this operation can employ model predictive control to continuously optimize the virtual output curve to match future fluctuation trends, or introduce a safety margin to conservatively estimate the maximum discharge power to address SOC estimation errors. This transforms energy storage capacity into a dispatchable virtual power source, improving system flexibility.

[0099] Based on the dynamic weight values ​​of the connecting edges in the energy coordination topology graph, power adjustment coefficients are assigned to energy nodes constrained by virtual output curves, and a coordination scheduling strategy is output.

[0100] The dynamic weight value of the connection edge can be the real-time edge weight representing the coupling strength between two nodes in the energy collaboration topology graph. Integrating multi-dimensional information on electrical, behavioral, and transactional aspects, it can be used as the basis for allocating power regulation coefficients, reflecting collaboration priority. In an exemplary embodiment, the dynamic weight value of the connection edge can directly determine the magnitude of the regulation coefficient obtained by the energy node constrained by the virtual output curve. The energy node constrained by the virtual output curve can be a distributed energy node that needs to adjust its operating state according to the virtual output curve in a high SOC scenario. It can be used as the main body for power regulation execution, undertaking collaborative support tasks. For example, the energy node constrained by the virtual output curve can include, but is not limited to, roles such as load nodes, other energy storage nodes, and controllable photovoltaic inverter nodes.

[0101] The power regulation coefficient can be a normalized regulation ratio factor allocated to each energy node based on dynamic weights. It can be used to quantify the responsibility share of each node in collaborative scheduling, achieving weighted allocation based on coupling strength. In one specific embodiment, the power regulation coefficient can be generated by Softmax normalization of the dynamic weights of the connection edges. Based on the dynamic weight values ​​of the connection edges in the energy collaborative topology graph, power regulation coefficients are allocated to energy nodes constrained by virtual output curves, and a collaborative scheduling strategy is output. This can be achieved by weighting the virtual output demand according to the dynamic weights between each node and energy storage, generating the final scheduling instruction set. Furthermore, this operation can be further refined by using weighted least squares to solve for the coefficient allocation that satisfies the total regulation amount and minimizes the deviation, or by introducing a priority mask to shield nodes without regulation capabilities (such as uncontrollable loads). This allows for differentiated allocation of regulation tasks based on coupling strength, improving collaborative efficiency and fairness.

[0102] Taking a scenario of collaborative support triggered by a sudden drop in photovoltaic power as an example, the IoT-based intelligent power dispatching method in this embodiment can be as follows: In an industrial park, rooftop photovoltaic A experiences a sudden drop in output from 800kW to 300kW within 2 seconds due to cloud cover. After detecting the anomaly, the system, based on the energy collaboration topology, finds strong connections between A and energy storage B (SOC = 85%) and energy storage C (SOC = 40%). After traversing the states, since B's SOC is higher than the threshold (70%), its maximum discharge power is calculated to be 600kW, with a ramp rate of 100kW / s, generating a virtual output curve that linearly increases from 0 to 500kW within 0-5 seconds. C, with its SOC below the threshold, receives 200kW of excess output from the adjacent photovoltaic D for charging. Subsequently, based on the dynamic weights of 0.78 for the AB side and 0.65 for the AC side, the system allocates power adjustment coefficients of 0.62 and 0.38 to the load E and energy storage F participating in the support, respectively, requiring them to reduce loads by 124kW and 76kW to make up for the remaining shortfall. The final coordinated scheduling strategy was issued within 300ms, which prevented the bus voltage from dropping.

[0103] In one embodiment, power flow conflict points are detected in real time and the parameter set of grid constraint conflict regions is marked. The specific steps are as follows:

[0104] A dynamic digital twin is constructed by loading the industrial park's power grid topology model and real-time measurement data stream onto a digital twin platform.

[0105] The collaborative scheduling strategy is imported into the dynamic digital twin, and a power flow spatiotemporal simulation is performed to calculate the line overload rate, node voltage offset, and frequency fluctuation amplitude.

[0106] Activate the power system simulation engine to simulate short-circuit fault effects and voltage sag processes, and mark the parameter set of grid constraint conflict areas that exceed the safety threshold.

[0107] The industrial park power grid topology model can be a structured digital model describing the electrical equipment, line parameters, and node attributes in the park's microgrid. It can serve as the basic framework for constructing a dynamic digital twin, providing physical consistency constraints. In this embodiment, the industrial park power grid topology model can be imported based on SCADA system ledger data or a CIM model and updated periodically. The real-time measurement data stream can be multi-source time-series data continuously uploaded by the IoT sensing layer, reflecting the current power grid operating status. It can be used to drive the synchronous evolution of the digital twin and the physical system, improving simulation fidelity. For example, the real-time measurement data stream can include, but is not limited to, one or more of voltage and current phasor streams, switch state streams, and frequency measurement streams. The dynamic digital twin can be a high-fidelity virtual microgrid mirror with online synchronization capabilities, constructed by integrating a static power grid topology model and real-time measurement data stream. It can be used to provide a physically consistent virtual verification environment for collaborative scheduling strategies. In an exemplary embodiment, the dynamic digital twin can be dynamically updated by embedding real-time measurements into an electromagnetic-electromechanical hybrid simulation model using a data assimilation algorithm.

[0108] By loading an industrial park power grid topology model and real-time measurement data streams onto a digital twin platform, a dynamic digital twin can be constructed. This can be achieved by fusing a static topology model with real-time data streams through data assimilation or state estimation methods, forming a dynamically evolving virtual mirror. Furthermore, this operation can be implemented by using extended Kalman filtering to perform state estimation on real-time measurements and then injecting it into the simulation model, or by utilizing an edge-cloud collaborative architecture to preprocess data at the edge before uploading it to the cloud-based digital twin platform to construct a global model. This allows for high-fidelity synchronization between the digital twin and the physical microgrid, providing a reliable environment for policy verification.

[0109] Power flow spatiotemporal simulation can be a time-continuous and spatially distributed power flow extrapolation of the execution process of a coordinated scheduling strategy within a dynamic digital twin. It can be used to evaluate the operational safety of the strategy under steady-state and quasi-steady-state conditions. In a specific embodiment, power flow spatiotemporal simulation can include, but is not limited to, one or more of time-domain power flow simulation, frequency-domain response simulation, and spatial distribution simulation. Line overload rate can be the ratio of the actual transmitted power of a line to its thermal stability limit, used to quantify thermal overload risk and can be used as one of the core indicators for determining whether a line exceeds its limits. For example, line overload rate can include, but is not limited to, one or more of instantaneous overload rate, moving average overload rate, and peak overload rate. Node voltage offset can be the absolute or relative value of the deviation between the actual voltage amplitude of a node and its rated voltage, and can be used to reflect the level of power quality and voltage stability. Further, node voltage offset can include, but is not limited to, one or more of single-phase offset and three-phase unbalance offset.

[0110] Frequency fluctuation amplitude can be the maximum instantaneous deviation of the system frequency from the nominal value (e.g., 50Hz), and can be used to characterize the microgrid's inertia support capacity and supply-demand balance. In an exemplary embodiment, frequency fluctuation amplitude can include, but is not limited to, one or more of the following: second-level fluctuations, minute-level drifts, and event-triggered mutations. Introducing the collaborative scheduling strategy into a dynamic digital twin and performing power flow spatiotemporal simulation to calculate line overload rate, node voltage offset, and frequency fluctuation amplitude can be achieved by inputting scheduling commands as external excitations into the dynamic digital twin, with the runtime simulation engine outputting key safety indicators. Furthermore, this operation can be implemented by using quasi-static timing simulation to balance accuracy and computational efficiency, or by combining a distributed simulation framework to calculate multi-region power flow indicators in parallel, thereby quantitatively evaluating the safety and power quality impact of the strategy under steady-state operation. Short-circuit fault effects can simulate electromagnetic transient phenomena such as current surges and voltage drops caused by metallic or high-resistance grounding short-circuit faults in the power grid, and can be used to verify the robustness of the scheduling strategy and the effectiveness of protection coordination under extreme faults. For example, the effects of a short circuit fault may include, but are not limited to, one or more of the following: three-phase short circuit, two-phase short circuit, and single-phase grounding.

[0111] Voltage sag processes can be short-term significant voltage drops and recovery processes caused by large load startup, fault clearing, or distributed generation disconnection. They can be used to assess the continuity of power supply to sensitive loads and reactive power support capabilities. In one embodiment, voltage sag processes can include, but are not limited to, one or more of fault-type sags, startup-type sags, and switching-type sags. Safety thresholds can be the allowable limits of various electrical parameters specified in power system operation procedures, and can be used as a criterion for automatically identifying conflict events. Furthermore, safety thresholds can include, but are not limited to, one or more of thermal stability thresholds, voltage compliance thresholds, and frequency tolerance thresholds. Activating the power system simulation engine to simulate short-circuit fault effects and voltage sag processes, and marking the parameter set of grid constraint conflict areas exceeding safety thresholds, can be achieved by superimposing typical transient disturbance conditions on basic power flow simulation, monitoring system response, and automatically identifying over-limit areas. Furthermore, this operation can be achieved by randomly injecting multiple fault scenarios using the Monte Carlo method to assess the overall risk probability, or by selecting high-incidence fault types based on a historical fault database for targeted stress testing. This proactively exposes the vulnerabilities of the strategy under abnormal conditions, improving the robustness of the dispatching scheme.

[0112] Taking fault ride-through verification under high photovoltaic penetration as an example, the IoT-based intelligent power dispatching method in this embodiment can be as follows: A microgrid in a certain park reaches peak photovoltaic output at noon. The coordinated dispatch strategy requires energy storage to support the nighttime load with a 200kW discharge. The system imports this strategy into a dynamic digital twin, loading real-time photovoltaic output, load curves, and grid topology. The power flow spatiotemporal simulation shows that the 10kV feeder overload rate reaches 105%, and the voltage deviation at node A is +8%. Further activating the power system simulation engine, a single-phase ground fault is simulated on bus B. The simulation shows that the voltage temporarily drops to 0.65 pu and lasts for 300 ms, exceeding the sensitive equipment tolerance threshold (0.7 pu, 200 ms). The system automatically marks the feeder and nodes A and B as grid constraint conflict areas, outputting a conflict area parameter set including line impedance, node rated voltage, and protection settings, for subsequent deep reinforcement learning module strategy correction.

[0113] In one embodiment, the output control parameter compensation value is determined through the following steps:

[0114] The parameter set of the power grid constraint conflict area is combined with the parameters of the digital twin environment to generate a state-action pair sample set, which is used as input data for the deep reinforcement learning algorithm;

[0115] The value function of the state-action pair sample set is calculated by the deep deterministic policy gradient algorithm, and the reward function corresponding to the optimal control policy is optimized.

[0116] The study analyzes the weight allocation defects of the reward function, identifies the voltage regulation weights and frequency response penalty coefficients that cause conflicts, and reconstructs the spatial structure of the reward function.

[0117] Inject device aging simulation noise and communication delay interference into the state-action pair sample set to generate adversarial disturbance samples, and train the policy network model to output control parameter compensation values.

[0118] The digital twin environment parameters can be the system operating state and environmental context information recorded by the digital twin during simulation, including topology, equipment model parameters, and external disturbance settings. These parameters can be combined with the conflict region parameter set to construct a complete state description reflecting the scheduling failure scenario. In an exemplary embodiment, the digital twin environment parameters can be extracted from the spatiotemporal pre-simulation log as the basic elements for state construction. For example, the digital twin environment parameters can include, but are not limited to, one or more of equipment model parameters, environmental boundary conditions, and simulation configuration parameters. The state-action pair sample set can be a training sample set consisting of a state vector composed of the power grid constraint conflict region parameter set and the digital twin environment parameters, and its corresponding original scheduling actions (control commands). This can be used as input to a deep reinforcement learning algorithm to learn the mapping relationship between policy failure and state features. Furthermore, the state-action pair sample set can be automatically constructed by extracting the system state at the time of conflict and the executed scheduling actions from the spatiotemporal pre-simulation log. In one specific embodiment, the state-action pair sample set can be constructed by using a sliding window to extract state sequences at multiple time steps before and after a conflict; or by sampling the samples based on their importance, prioritizing the retention of high-risk conflict scenarios to improve training efficiency. This provides high-quality training data that reflects real failure scenarios for reinforcement learning, avoiding the disconnect between pure simulation data and reality.

[0119] Deep deterministic policy gradient algorithms (DDPGs) can be deep reinforcement learning algorithms applicable to continuous action spaces. Through an Actor-Critic architecture, they jointly optimize the policy network and value network, and can be used to learn the optimal control policy from a set of state-action pairs and infer its implicit reward function. Examples of DDPGs include, but are not limited to, dual-delay DDPG (TD3), ensemble DDPG, and priority experience replay DDPG. The value function can be a function that evaluates the long-term cumulative reward of performing an action in a given state, and can be used as the output of the Critic network in the DDPG algorithm to support policy gradient calculation. In a specific embodiment, the value function can include, but is not limited to, the state value function V(s) and the action value function Q(s, a). The optimal control policy can be a scheduling decision mapping that maximizes the expected cumulative reward under the current reward function, and can be used as a reference benchmark for reward function reconstruction to identify policy defects. Examples of optimal control policies include, but are not limited to, deterministic policies and stochastic policies.

[0120] The reward function can be a scalar feedback signal used in reinforcement learning to quantify the merits of a strategy. It is composed of a weighted combination of multi-objective constraints and can be used to guide an agent to learn safe, stable, and efficient scheduling behaviors. In an exemplary embodiment, the reward function may include, but is not limited to, voltage safety terms, frequency stability terms, and economic operation terms. A state-action pair sample set is generated by combining the power grid constraint conflict region parameter set with the digital twin environment parameters. This sample set serves as input data for the deep reinforcement learning algorithm. It can be derived from the system state at the moment of conflict occurrence extracted from the digital twin simulation log, including conflict region parameters, environmental parameters, and corresponding scheduling actions, thus constructing the sample pairs. Furthermore, this operation can be performed by using a sliding window to extract state sequences at multiple time steps before and after the conflict, constructing time-series state-action pairs; or by performing importance sampling on the samples, prioritizing high-risk conflict scenarios to improve training efficiency. This provides high-quality training data that reflects real failure scenarios for reinforcement learning, avoiding a disconnect between pure simulation data and reality.

[0121] The value function of state-action pairs on a sample set is calculated using a deep deterministic policy gradient algorithm, optimizing the reward function corresponding to the optimal control policy. This can be achieved by fitting the action value function using the Critic network of DDPG and then inferring the implicit reward function through inverse reinforcement learning or policy gradient analysis. Furthermore, this operation can recover the reward function from the expert policy (i.e., the original scheduling policy) using maximum entropy inverse reinforcement learning; or by identifying the contribution of each state dimension to action selection through policy gradient sign analysis, indirectly inferring the reward weight. This reveals the internal structure of the current policy optimization objective and provides a basis for defect diagnosis. The voltage regulation weight can be an adjustable coefficient in the reward function used to measure the intensity of voltage exceedance penalty. It can be used to control the policy's emphasis on voltage safety; too low a weight can easily lead to voltage exceedance. In a specific embodiment, the voltage regulation weight can be a key component of the reward function's spatial structure, and its defects directly affect the direction of the control parameter compensation value.

[0122] The frequency response penalty coefficient can be a parameter in the reward function used to quantify the severity of the frequency deviation penalty. It can influence the policy's sensitivity to system inertia support and supply-demand balance. For example, the frequency response penalty coefficient can include, but is not limited to, the steady-state frequency penalty coefficient and the dynamic frequency change rate penalty coefficient. The reward function space structure can be the overall architecture of the composition, weight configuration, and nonlinear transformation methods of each sub-item in the reward function. It can be used to determine the multidimensional balance relationship of the policy optimization objective and is a fundamental driving factor of policy performance. In an exemplary embodiment, the reward function space structure can be reconstructed by analyzing the failure modes of existing policies in conflict scenarios.

[0123] Analyzing the weight allocation defects in the reward function, identifying the voltage regulation weights and frequency response penalty coefficients that cause conflicts, and reconstructing the spatial structure of the reward function can be achieved by comparing the weights of the actual exceedance indicators in the conflict scenario with the corresponding terms in the reward function, locating sub-terms with insufficient or unbalanced weights, and redesigning the function form and weight ratios. Furthermore, this operation can be further optimized by introducing multi-objective Pareto front analysis to automatically search for the optimal weight combination; or by using a learnable reward network to replace the fixed weighting form, achieving dynamic weight adjustment. This allows the revised strategy to optimize the objective, making it more accurately reflect multi-dimensional operational constraints and fundamentally avoiding similar conflicts.

[0124] Equipment aging simulation noise can be random disturbances injected into the training data to simulate equipment performance degradation effects such as increased internal resistance of energy storage and decreased photovoltaic efficiency. This can be used to improve the robustness of the policy network to uncertainties in equipment state. For example, equipment aging simulation noise can include, but is not limited to, battery aging noise, inverter efficiency degradation noise, and line impedance drift noise. Communication delay interference can simulate time delay disturbances caused by network congestion and packet loss during control command issuance or state upload. This can be used to enhance the policy's adaptability to timing misalignments and avoid control failures caused by command lag. In a specific embodiment, communication delay interference can include, but is not limited to, fixed delay, random jitter delay, and sudden long delay. Adversarial disturbance samples can be enhanced training samples generated by superimposing equipment aging noise and communication delay interference on the original state-action pairs. This can be used to train a more robust policy network and improve its generalization ability in real complex environments. Furthermore, adversarial disturbance samples can be generated by injecting physically consistent disturbance models into the state vector and action execution sequence.

[0125] The policy network model can be a neural network used in deep reinforcement learning to output control actions. Here, it specifically refers to the Actor network trained with adversarial samples, which can be used to directly generate control parameter compensation values ​​with perturbation compensation capabilities. For example, the policy network model can include, but is not limited to, fully connected policy networks, graph neural network policy networks, and temporal convolutional network policy networks. By injecting device aging simulation noise and communication delay interference into the state-action pair sample set, adversarial perturbation samples are generated. The policy network model is then trained to output control parameter compensation values. This can be achieved by superimposing physically consistent perturbations into the state input and action execution path of the original samples, training the policy network to still output effective compensation actions under perturbation. Furthermore, this operation can use Wasserstein distance to constrain the perturbation amplitude, ensuring that the noise conforms to the real distribution; or by combining domain randomization techniques to cover various aging and delay combinations during training, thereby improving the robustness and adaptability of the policy in real-world, non-ideal environments and ensuring that the compensation values ​​are effective in actual deployment.

[0126] Taking a voltage over-limit compensation scenario considering equipment aging as an example, the IoT-based intelligent power dispatching method in this embodiment can be as follows: In a digital twin simulation of a microgrid in a certain park, the bus voltage drops to 0.88 per unit due to a sudden drop in photovoltaic power, marking it as a conflict area. The system extracts the grid constraint conflict area parameter set at this moment, including line impedance, node load, and digital twin environmental parameters (weather, temperature), and constructs state-action pair samples; by back-deriving the original reward function through the DDPG algorithm, it is found that the voltage regulation weight is only 0.3, far lower than the frequency term of 0.8; the reward function is reconstructed, the voltage weight is increased to 0.7, and a secondary penalty term is introduced; then, aging noise with a 15% increase in simulated battery internal resistance and a 200ms communication delay are injected into the samples to generate adversarial disturbance samples; after training, the policy network outputs control parameter compensation values ​​including +120kVar reactive power compensation and energy storage activation 0.3 seconds in advance under the same scenario, successfully maintaining the voltage above 0.93 per unit.

[0127] In one embodiment, a distributed energy control instruction set is generated and the instructions are issued. The specific steps are as follows:

[0128] Based on the control parameter compensation values, the PWM modulation duty cycle of the energy storage converter is calculated, and converter control commands are generated.

[0129] The energy storage converter can be a power electronic device connecting the energy storage battery and the power grid, used to control charging and discharging power and reactive power output. The energy storage converter can also be used as a physical device to execute converter control commands, achieving precise regulation of active and reactive power. In an exemplary embodiment, the energy storage converter can be one or more of, but not limited to, two-level converters, three-level converters, and modular multilevel converters. The PWM modulation duty cycle can be the ratio of the high-level duration to the period in the pulse width modulation signal, used to control the output voltage amplitude of the converter. The PWM modulation duty cycle directly determines the active and reactive power output capability of the energy storage converter and is a key parameter for underlying control. Furthermore, the PWM modulation duty cycle can be generated by mapping the control parameter compensation value through coordinate transformation and PI adjustment.

[0130] The converter control command can be a set of underlying drive signals containing parameters such as PWM duty cycle, switching frequency, and phase angle. The converter control command can be used to directly act on the power devices of the energy storage converter to achieve rapid power response. For example, the converter control command can include, but is not limited to, one or more of active power control commands, reactive power control commands, and start-stop control commands. Based on the control parameter compensation value, the PWM modulation duty cycle of the energy storage converter is calculated, and the converter control command is generated. This can be achieved by mapping the control parameter compensation value to a voltage reference value in the dq coordinate system, and then generating the duty cycle signal through inverse Park transform and PWM modulation. Furthermore, this operation can be achieved by using space vector PWM (SVPWM) to improve the DC bus voltage utilization rate, or by introducing a dead-time compensation algorithm to reduce the impact of nonlinear distortion on the duty cycle, thereby enabling precise dynamic adjustment of the energy storage converter's output power to respond to system compensation requirements.

[0131] Based on the parameter set of the grid-constrained conflict region, the output curve of the photovoltaic inverter is replanned using the particle swarm optimization algorithm, and the output power reference value sequence is generated.

[0132] The particle swarm optimization (PSO) algorithm can be a global optimization algorithm based on swarm intelligence, searching for the optimal solution by simulating the foraging behavior of bird flocks. PSO can be used to reconstruct the photovoltaic (PV) output curve under grid constraints, balancing safety and economy. In a specific embodiment, the PSO algorithm may include, but is not limited to, one or more of adaptive inertial weighted PSO, chaotic PSO, and multi-objective PSO. The output curve can be the planned active power output trajectory of the PV inverter over a future period. The output curve can be used as a concrete manifestation of the scheduling strategy on the PV side, and must meet line capacity and voltage constraints. For example, the output curve may include, but is not limited to, one or more of second-level smooth curves, minute-level stepped curves, and event-triggered abrupt change curves. The power reference value sequence can be the optimized output curve discretized into a time-stamped sequence of power setpoints. The power reference value sequence can be used for the PV inverter controller to track and execute, ensuring that the actual output meets the requirements of coordinated scheduling. Furthermore, the power reference value sequence and the converter control commands can be aligned with timestamps to form a distributed energy control command set.

[0133] Based on the parameter set of grid-constrained conflict regions, a particle swarm optimization (PSO) algorithm is used to replan the output curve of the photovoltaic (PV) inverter, outputting a power reference value sequence. This can be achieved by minimizing the curtailment rate and voltage deviation, solving for the optimal output trajectory within the PSO framework that satisfies constraints such as line capacity and node voltage. Furthermore, this operation can be implemented by employing constraint processing techniques, such as the penalty function method, to transform hard constraints into optimization objectives, or by introducing a rolling time-domain optimization mechanism to achieve online dynamic replanning. This results in a physically feasible and safe PV dispatch scheme, avoiding secondary limit violations.

[0134] The converter control commands and power reference value sequences are aligned with timestamps to generate structured command data packets. The converter control commands are sent to the energy storage controller via industrial Ethernet, while the power reference value sequences are transmitted to the photovoltaic inverter via a wireless sensor network.

[0135] The timestamp can be a precise time stamp identifying the moment the instruction takes effect, typically using UTC or PTP synchronization time. The timestamp ensures strict synchronization of control instructions from multiple devices across time, preventing misalignment. In an exemplary embodiment, the timestamp can be one or more of millisecond-level, microsecond-level, and nanosecond-level timestamps. The structured instruction data packet can be a communication unit that encapsulates converter control instructions and power reference value sequences in a unified format and appends metadata. The structured instruction data packet supports reliable transmission and parsing of heterogeneous instructions, ensuring consistency in multi-device collaborative execution. Furthermore, the structured instruction data packet can be encapsulated in JSON or Protobuf format, containing fields such as device ID, instruction type, timestamp, and payload.

[0136] Industrial Ethernet can be a high-reliability, low-latency wired communication network for industrial control scenarios. It can be used to transmit converter control commands with high real-time requirements, ensuring deterministic response. For example, Industrial Ethernet may include, but is not limited to, one or more of PROFINET, EtherCAT, and ModbusTCP. The energy storage controller can be a local control unit that receives and parses converter control commands, driving the energy storage converter to perform power regulation. It can act as a command execution terminal, converting digital commands into physical actions. In one specific embodiment, the energy storage controller may include, but is not limited to, one or more of local controllers and centralized controllers. The wireless sensor network can be a communication network composed of low-power wireless nodes, suitable for flexible access of distributed devices. It can be used to transmit power reference value sequences to distributed photovoltaic inverters, reducing wiring costs. Furthermore, the wireless sensor network may include, but is not limited to, one or more of LoRaWAN, ZigBee, and NB-IoT.

[0137] Aligning converter control commands and power reference value sequences by timestamp to generate structured command data packets can be achieved by interpolating or aligning the two types of commands according to a unified time base and encapsulating them into a standard communication format with metadata. Furthermore, this operation can employ a time window caching mechanism to align asynchronously arriving commands, or use message queue middleware to ensure command order and integrity, thereby ensuring coordinated action of multiple devices in the time dimension and improving the overall system response consistency. Sending converter control commands to the energy storage controller via industrial Ethernet, while simultaneously transmitting the power reference value sequence to the photovoltaic inverter via a wireless sensor network, can be achieved by selecting differentiated transmission channels based on device communication characteristics and issuing corresponding commands separately. Furthermore, this operation can be achieved by enabling Time-Sensitive Networking (TSN) on the industrial Ethernet to ensure microsecond-level determinism, or by using forward error correction coding in the wireless sensor network to improve anti-interference capabilities, thus balancing the real-time nature of critical control commands with the deployment flexibility of distributed devices.

[0138] Taking the collaborative control execution under a hybrid communication architecture as an example, the IoT-based intelligent power dispatching method in this embodiment can be used to address a situation where a 10kV feeder in a park's microgrid experiences a sudden load surge in the afternoon, leading to overload. The system generates control parameter compensation values ​​of +150kVar reactive power support and -200kW active power reduction. Based on this, the PWM duty cycle of the energy storage converter is calculated, generating converter control commands containing dq-axis voltage instructions. Simultaneously, based on the overload line parameters, a particle swarm optimization algorithm is used to adjust the output curves of three adjacent photovoltaic inverters for the next 5 minutes from 800kW / unit to 650kW / unit, outputting a time series with a 1-second interval. The two types of commands are aligned with PTP synchronization timestamps and encapsulated into Protobuf format data packets. The converter commands are sent to the energy storage controller via EtherCAT industrial Ethernet within 2ms, and the photovoltaic power reference value sequence is delivered to each inverter via LoRaWAN wireless network within 500ms. Finally, the system completes power flow adjustment within 3 seconds, reducing the feeder load to 95% of its rated value.

[0139] In addition, refer to Figure 2 To achieve the above objectives, the present invention also provides an intelligent power dispatching system based on the Internet of Things, the system comprising:

[0140] Data acquisition module 10 is used to collect photovoltaic output data, energy storage system status data and load demand data in the microgrid scenario of the industrial park in real time, and output raw data stream;

[0141] The topology modeling module 20 is used to input the raw data stream into a pre-trained graph neural network and construct an energy collaboration topology graph by analyzing the spatial topological relationships, output correlations and energy transaction record characteristics of distributed energy nodes.

[0142] The collaborative decision-making module 30 is used to trigger a dynamic electricity price response mechanism based on the energy collaborative topology diagram and generate collaborative scheduling strategies for distributed energy nodes with power coupling relationships.

[0143] The simulation pre-simulation module 40 is used to build a digital twin of the industrial park microgrid on the digital twin platform, import the collaborative scheduling strategy into the digital twin environment for spatiotemporal pre-simulation, detect power flow conflict points in real time, and mark the parameter set of grid constraint conflict area.

[0144] The reinforcement optimization module 50 is used to optimize conflict events in the pre-simulation through deep reinforcement learning algorithms, trace the source of control strategy defect features and reconstruct the reward function space, and output control parameter compensation values.

[0145] The instruction execution module 60 is used to integrate the parameter set of the grid constraint conflict area with the control parameter compensation value, generate a distributed energy control instruction set, and execute and issue the instructions.

[0146] Other embodiments or specific implementations of the IoT-based intelligent power dispatching system described in this invention can be found in the above-described method embodiments, and will not be repeated here.

[0147] Furthermore, to achieve the above objectives, the present invention also provides an Internet of Things (IoT)-based intelligent power scheduling device, the device comprising: a memory, a processor, and an IoT-based intelligent power scheduling program stored in the memory and executable on the processor, the IoT-based intelligent power scheduling program being configured to implement the steps of the IoT-based intelligent power scheduling method as described above.

[0148] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an Internet of Things (IoT)-based intelligent power scheduling program, wherein when the IoT-based intelligent power scheduling program is executed by a processor, it implements the steps of the IoT-based intelligent power scheduling method as described above.

[0149] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A power supply intelligent scheduling method based on the Internet of Things, characterized in that, The method includes: Real-time acquisition of photovoltaic power output data, energy storage system status data, and load demand data in industrial park microgrid scenarios, and output of raw data stream; The raw data stream is input into a pre-trained graph neural network, and an energy collaborative topology graph is constructed by analyzing the spatial topological relationships, output correlations and energy transaction record characteristics of distributed energy nodes. Based on the energy collaboration topology diagram, a dynamic electricity price response mechanism is triggered to generate collaborative scheduling strategies for distributed energy nodes with power coupling relationships; A digital twin of the industrial park microgrid is constructed on the digital twin platform. The collaborative scheduling strategy is imported into the digital twin environment for spatiotemporal pre-simulation, and power flow conflict points are detected in real time and the parameter set of grid constraint conflict area is marked. By optimizing conflict events in pre-simulation using deep reinforcement learning algorithms, the source of control strategy defects is traced and the reward function space is reconstructed, and the compensation value of control parameters is output. By integrating the parameter set of conflicting areas in the power grid with the compensation values ​​of control parameters, a set of distributed energy control instructions is generated and executed.

2. The power supply intelligent scheduling method based on the Internet of Things as described in claim 1, characterized in that, The load demand data includes industrial production load curves and electricity consumption records for public utilities.

3. The power supply intelligent scheduling method based on the Internet of Things as described in claim 1, characterized in that, The specific steps for constructing the energy synergy topology diagram are as follows: Analyze the distributed energy node information in the raw data stream to extract node spatial coordinates, output fluctuation coefficients, and energy trading frequency; The frequency of energy transactions is smoothed by an index to generate transaction correlation coefficients, and the spatial coordinates are calculated by inverse power grid impedance to generate electrical distance weights. The power output fluctuation coefficient is used as the initial weight of the graph edge. The transaction correlation coefficient and the electrical distance weight are integrated to generate a dynamic energy weight matrix. The energy node features are aggregated through a graph attention network to output an energy collaborative topology graph.

4. The IoT-based intelligent power dispatching method as described in claim 3, characterized in that, The specific steps for generating a collaborative scheduling strategy for distributed energy nodes with power coupling relationships are as follows: When the output of any photovoltaic node in the energy collaboration topology graph fluctuates abnormally, the state of all energy storage systems connected to the fluctuating node is traversed to obtain the state of charge value and the charging and discharging power limit. If the state of charge of the energy storage system is lower than the threshold, a power compensation command will be generated to forcibly allocate the output of the associated photovoltaic node to the energy storage system; If the state of charge of the energy storage system is higher than the threshold, calculate the maximum discharge power of the corresponding energy storage system and generate a virtual output curve that includes power ramp-up rate constraints. Based on the dynamic weight values ​​of the connecting edges in the energy coordination topology graph, power adjustment coefficients are assigned to energy nodes constrained by virtual output curves, and a coordinated scheduling strategy is output.

5. The IoT-based intelligent power dispatching method as described in claim 4, characterized in that, The specific steps for real-time detection of power flow conflict points and marking of power grid constraint conflict region parameter sets are as follows: A dynamic digital twin is constructed by loading the industrial park's power grid topology model and real-time measurement data stream onto a digital twin platform. The collaborative scheduling strategy is imported into the dynamic digital twin, and a power flow spatiotemporal simulation is performed to calculate the line overload rate, node voltage offset, and frequency fluctuation amplitude. Activate the power system simulation engine to simulate short-circuit fault effects and voltage sag processes, and mark the parameter set of grid constraint conflict areas that exceed the safety threshold.

6. The IoT-based intelligent power dispatching method as described in claim 5, characterized in that, The specific steps for compensating the output control parameter value are as follows: The parameter set of the power grid constraint conflict area is combined with the parameters of the digital twin environment to generate a state-action pair sample set, which is used as input data for the deep reinforcement learning algorithm; The value function of the state-action pair sample set is calculated by the deep deterministic policy gradient algorithm, and the reward function corresponding to the optimal control policy is optimized. The study analyzes the weight allocation defects of the reward function, identifies the voltage regulation weights and frequency response penalty coefficients that cause conflicts, and reconstructs the spatial structure of the reward function. Inject device aging simulation noise and communication delay interference into the state-action pair sample set to generate adversarial disturbance samples, and train the policy network model to output control parameter compensation values.

7. The IoT-based intelligent power dispatching method as described in claim 6, characterized in that, The specific steps for generating the distributed energy control instruction set and executing the instruction issuance are as follows: Based on the control parameter compensation values, calculate the PWM modulation duty cycle of the energy storage converter and generate converter control commands; Based on the parameter set of grid-constrained conflict regions, the output curve of the photovoltaic inverter is replanned using the particle swarm optimization algorithm, and the output power reference value sequence is generated. The converter control commands and power reference value sequences are aligned by timestamp to generate a structured command data packet. The converter control commands are sent to the energy storage controller via industrial Ethernet, while the power reference value sequence is transmitted to the photovoltaic inverter via a wireless sensor network.

8. A power supply intelligent dispatching system based on the Internet of Things, characterized in that, The system includes: The data acquisition module is used to collect photovoltaic power output data, energy storage system status data and load demand data in real time in the microgrid scenario of the industrial park, and output raw data stream; The topology modeling module is used to input the raw data stream into a pre-trained graph neural network and construct an energy collaboration topology graph by analyzing the spatial topological relationships, output correlations, and energy transaction record characteristics of distributed energy nodes. The collaborative decision-making module is used to trigger a dynamic electricity price response mechanism based on the energy collaboration topology diagram and generate collaborative scheduling strategies for distributed energy nodes with power coupling relationships. The simulation pre-simulation module is used to build a digital twin of the industrial park microgrid on the digital twin platform, import the collaborative scheduling strategy into the digital twin environment for spatiotemporal pre-simulation, detect power flow conflict points in real time, and mark the parameter set of grid constraint conflict areas. The reinforcement optimization module is used to optimize conflict events in the pre-simulation using deep reinforcement learning algorithms, trace the source of control strategy defects and reconstruct the reward function space, and output control parameter compensation values. The instruction execution module is used to integrate the parameter set of the grid constraint conflict area with the control parameter compensation value, generate the distributed energy control instruction set, and execute and issue the instructions.

9. A power supply intelligent dispatching device based on the Internet of Things, characterized in that, The device includes: a memory, a processor, and an Internet of Things (IoT)-based intelligent power scheduling program stored in the memory and executable on the processor, the IoT-based intelligent power scheduling program being configured to implement the steps of the IoT-based intelligent power scheduling method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an Internet of Things (IoT)-based intelligent power scheduling program, which, when executed by a processor, implements the steps of the IoT-based intelligent power scheduling method as described in any one of claims 1 to 7.

Images