Distributed energy scheduling method based on Internet of Things edge computing enabling

Abstract

The invention relates to the technical field of intelligent power grid and power system automation, discloses a distributed energy scheduling method based on Internet of Things edge computing enabling, and aims to solve the problem of efficient and safe cooperative regulation and control of massive heterogeneous distributed resources. The edge computing nodes sense states of power grid frequency, voltage and the like in real time, generate unified and multi-dimensional digital potential field signals in combination with strategic intentions issued by the cloud and broadcast the signals to equipment in a coverage area. An autonomous mental model is built in each distributed energy device, and the model can autonomously interpret potential field signals and independently calculate final execution power in combination with internal safety constraints such as the charge state and the temperature of the model, so that device-level safety guarantee and power grid-level self-organizing collaborative response are achieved. Furthermore, a double-closed-loop intelligent evolution mechanism is constructed at the cloud end, and prospective guidance of the future state of the system is achieved through potential field element learning.

Description

Technical Field

[0001] This invention relates to the field of smart grid and power system automation technology, specifically to a distributed energy dispatching method enabled by Internet of Things edge computing. Background Technology

[0002] With the profound transformation of the global energy structure, the penetration rate of renewable energy sources, represented by photovoltaics and wind power, in the power system is climbing at an unprecedented rate. The inherent intermittency and volatility of these power sources pose a severe challenge to the real-time power balance and safe, stable operation of the power grid. At the same time, massive amounts of dispatchable distributed energy resources (DERs), such as electric vehicles, energy storage systems, and smart home appliances, are being widely integrated into the distribution network. Their flexible power regulation capabilities offer enormous potential for absorbing fluctuating renewable energy and providing ancillary services. However, how to effectively and reliably regulate these vast, diverse, and widely distributed distributed resources to form a usable combined force to support the power grid is a core technological bottleneck currently facing the field of power system dispatch. Existing technologies, in attempting to solve this problem, typically follow several main technical paths, but all have inherent limitations.

[0003] A mainstream approach is to adopt a centralized direct control or virtual power plant (VPP) aggregation scheduling model. This model attempts to uniformly model, monitor the status, and issue commands to massive distributed resources through a central control system or VPP platform. However, as the scale of resources surges from tens of thousands to tens of millions or even hundreds of millions, this centralized architecture gradually becomes overwhelmed at the communication and computing levels. It requires establishing bidirectional high-frequency communication with each device, and the resulting communication overhead and computational burden on the central node increase exponentially, severely limiting the system's scalability. More importantly, the control latency from the central perception of disturbances to the execution of commands at the terminal often fails to meet the stringent timeliness requirements of the power grid for rapid response (such as second-level or even sub-second inertia support or primary frequency regulation), and the system's reliability also relies excessively on the health of the central node, posing a single point of failure risk.

[0004] Another technical approach relies on local autonomous control strategies, such as droop control based on frequency or voltage deviations. This method pre-embeds the control logic locally, achieving rapid local response without relying on complex communication. However, its response characteristics are rigid and singular; each device can only react simply based on its measured local information, unable to understand and respond to more complex, multi-objective system-level scheduling intentions, such as balancing frequency security with line power flow constraints and operational economy. More importantly, due to the lack of higher-level coordination mechanisms, the disorderly responses of numerous devices based on local information may even trigger localized power oscillations under certain circumstances, jeopardizing system stability and failing to achieve system-level resource optimization. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a distributed energy scheduling method empowered by IoT edge computing, aiming to solve the problem of efficient, secure, and collaborative regulation of massive heterogeneous distributed resources.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a distributed energy dispatching method empowered by IoT edge computing, the method comprising: At the edge computing node, a digital potential field representing the intended energy demand is generated based on at least one collected power grid operating parameter. The edge computing node broadcasts the digital potential field to multiple distributed energy devices within its coverage area; Each of the plurality of distributed energy devices autonomously determines its operating power based on the digital potential field it receives and its own internal state.

[0007] Preferably, the step of generating a digital potential field representing the intended energy demand specifically includes: generating at least one sub-potential field based on the at least one grid operating parameter; and weighting and superimposing the at least one sub-potential field to form the digital potential field.

[0008] Preferably, the at least one sub-potential field includes at least one of the following: a frequency-stabilized potential field generated based on the grid frequency deviation; a voltage support potential field generated based on the node voltage deviation; an economic incentive potential field generated based on the real-time electricity price signal; and a network congestion potential field generated based on the line load rate.

[0009] For example, the frequency-stabilized potential field It can be generated using the following formula: ; in, The power grid frequency is collected in real time. Nominal frequency, The amplitude coefficient of the frequency potential field. This is the frequency response sensitivity coefficient.

[0010] For example, the voltage-supported potential field It can be generated using the following formula: ; in, For position Real-time node voltage amplitude at the location, Nominal voltage, This is the voltage potential field amplitude coefficient.

[0011] Preferably, the step of generating a digital potential field representing the intended energy demand further includes: receiving an injection morphological potential field generated by a meta-learning module from the cloud; and superimposing the injection morphological potential field onto the digital potential field. The meta-learning module is used to learn and generate the injection morphological potential field aimed at achieving a preset scheduling target based on the macroscopic system operating effect emerging after the multiple distributed energy devices autonomously determine their operating power.

[0012] For example, the digital potential field The generation of can be characterized by the following formula: ; in, For the first Types of subpotential fields, Its corresponding weight coefficient, This refers to the injected morphological potential field. The meta-learning module uses algorithms such as deep reinforcement learning to map the historical macroscopic state of the system to the injected morphological potential field, in order to maximize the predefined long-term cumulative system reward, thereby achieving forward-looking guidance of system behavior.

[0013] Preferably, the step of each device autonomously determining its operating power specifically includes: the device having a built-in autonomous mental model, the autonomous mental model storing a set of behavioral parameters defining its response characteristics; the autonomous mental model generating an initial intention power based on the digital potential field and the set of behavioral parameters; and the autonomous mental model constraining the initial intention power based on the internal state of the device, thereby determining the final operating power.

[0014] To enable those skilled in the art to prepare and implement the autonomous mental model, its initialization preparation method may include the following steps: Define a model software architecture that includes basic response logic and parameterized interfaces; By performing offline training on historical power grid operation data and typical response data of similar equipment (e.g., using supervised learning or system identification methods), a set of initial behavioral parameters that can reflect the basic response characteristics of this type of equipment is obtained. (0); The model software structure and the initial behavioral parameter set are encapsulated together and implanted as firmware into the controller of the distributed energy device.

[0015] In one specific embodiment, the behavioral parameter set It explicitly includes the following: a set of response coefficients corresponding to different subpotential fields. A risk preference coefficient is used to adjust the response intensity to different guidance signals; a risk preference coefficient is used to define whether the device tends to be conservative or aggressive in the face of uncertainty; and a set of safety boundary threshold parameters related to the internal state of the device, such as maximum / minimum state of charge, maximum allowable operating temperature, etc.

[0016] Preferably, the internal state includes at least one of the device's state of charge, health status, device temperature, or user preset preferences.

[0017] For example, for a distributed energy device Its initial intended power The generation of can be characterized by the following formula: ; in, For equipment The location of The strength of the digital potential field it perceives. It is a comprehensive representation of one or more response coefficients stored in its behavioral parameter set. The final operating power is the result of the initial intended power within the feasible domain determined by the internal state of the device (such as rated power, state-of-charge safety range, etc.).

[0018] Preferably, the method further includes an individual behavior drift step: the device autonomously adjusts its set of behavior parameters based on the internal state feedback generated after its operating power is executed.

[0019] In one specific embodiment, during the individual behavior drift step, the internal state feedback upon which the device adjusts its behavior parameter set is quantified as a local cost function. An example of this local cost function is as follows: ; in, For the real-time temperature of the equipment, The first item is used to penalize actions that cause the equipment to overheat, representing the maximum permissible temperature. This is the actual power. The second term serves as a smooth reference power curve, and is used to penalize excessively drastic power fluctuations to mitigate equipment losses. and This is the penalty weight. The device calculates this cost function for its set of behavioral parameters. The gradient is calculated and small adjustments are made along the negative direction of the gradient to achieve self-optimization of individual behavior.

[0020] Preferably, the method further includes a step of solidifying group experience: aggregating the adjusted set of behavioral parameters from multiple devices in the cloud; identifying and extracting superior behavioral parameter features based on the aggregation result to generate an updated set of basic behavioral parameters; and using the updated set of basic behavioral parameters to initialize new devices or upgrade existing devices.

[0021] For example, in the step of solidifying group experience, the updated set of basic behavioral parameters By using a high-performance equipment cluster Generate by aggregating the set of behavioral parameters: ; in, For equipment In the The set of behavioral parameters for each evolutionary cycle, where Aggregate(·) is the aggregation function, such as the federated average algorithm.

[0022] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the method described above.

[0023] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0024] This invention provides a distributed energy dispatching method enabled by IoT edge computing. It has the following beneficial effects: 1. This invention offloads core computing tasks to edge nodes. Edge nodes only need to broadcast a unified digital potential field signal to their coverage area, rather than establishing individual communication with a massive number of devices, fundamentally eliminating the communication bottleneck of centralized architectures. Control decisions are made at the edge, close to the data source and actuators, ensuring sub-second response to grid disturbances and meeting the most stringent grid security and stability requirements. Simultaneously, the architecture itself possesses the horizontal scalability capability to support hundreds of millions of connected devices.

[0025] 2. In this invention, control commands are no longer mandatory power setpoints, but rather guiding potential field signals. Before making a final decision, each device's built-in autonomous mental model will forcibly integrate external guidance with its own internal constraints, such as state of charge, temperature, health, and even user presets, and quantify them into an instantaneous power feasible domain. This design ensures that any response behavior will never exceed the device's own safety boundaries, protecting both expensive assets and respecting the user's individual wishes.

[0026] 3. This invention transforms complex system-level scheduling objectives into an "environmental" signal that all devices can understand by applying a unified and concise digital potential field. Guided by this unified signal, tens of thousands of heterogeneous devices independently and in parallel make decisions that best suit their own capabilities. These seemingly simple individual decisions converge to naturally generate highly coordinated, precise, and effective collective behavior at the macro level, forming a powerful synergy to cope with power grid disturbances without any centralized, fine-tuned orchestration.

[0027] 4. Through a closed-loop learning mechanism using potential field elements deployed in the cloud, the system no longer passively responds to current events. This mechanism learns long-term operational patterns from macroscopic historical data and designs a "potential field injection pattern" that guides the system towards a better future state. For example, to prepare for a foreseeable surge in photovoltaic power generation several hours later, the system can pre-generate a potential field to guide the appropriate discharge of energy storage, thus creating space for energy absorption. This shift from "passive response" to "active guidance" significantly improves the system's ability to absorb renewable energy and its overall operational economy. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the system architecture and information interaction of the present invention; Figure 2 This is a flowchart illustrating the generation and synthesis of the digital potential field according to the present invention. Figure 3 This is a flowchart of the autonomous decision-making process for the distributed energy equipment of the present invention; Figure 4 This is a schematic diagram of the dual closed-loop intelligent evolution mechanism of the present invention. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Please see the appendix Figure 1To be continued Figure 4 This invention provides a distributed energy scheduling method empowered by IoT edge computing, which is implemented on a three-layer physical architecture consisting of a central cloud, edge computing nodes, and distributed energy devices. Each layer has clear responsibilities and efficient closed-loop information interaction, together forming an intelligent scheduling ecosystem with self-organization, self-adaptation, and self-learning capabilities.

[0031] In this embodiment, the distributed energy device is the physical foundation and final execution unit of the method of this invention. It encompasses a vast array of heterogeneous resources at the end of the power grid, such as photovoltaic inverters, power storage converters (PCS), electric vehicle charging piles, interruptible loads, and smart home appliances. Each device's local controller is embedded with and runs a lightweight "autonomous mental model." This model is the core of the device's autonomous decision-making, enabling the device to act as an independent intelligent agent, proactively sensing the environment and responding in a way that best protects its own safety and meets the grid's collaborative requirements, rather than passively waiting for precise instructions from the central controller. Deploying this model at the device level ensures extremely low decision-making latency and prioritizes the device's own physical safety constraints (such as battery temperature and charge / discharge rate) for local closed-loop verification.

[0032] Edge computing nodes, acting as crucial hubs connecting the central cloud with a massive number of devices, serve as regional "tactical coordinators." Physically, they can be deployed in locations such as edge computing gateways in regional substations and distribution networks, or 5G base stations. One of the core modules in this invention, the "multi-dimensional digital potential field generator," runs on an edge computing node. The core reason for this deployment is that the generation of the digital potential field relies on high-speed and accurate sensing of the real-time status of the regional power grid (such as local voltage and line power flow). Placing this module close to the data source at the edge avoids the network latency and bandwidth pressure caused by transmitting massive amounts of high-frequency data to the cloud, ensuring that the digital potential field can accurately and in real-time reflect the dynamic needs of the local power grid, providing time assurance for rapid response at the device end.

[0033] In this invention, the central cloud plays a top-level role as both a "strategic planner" and an "evolutionary engine." It possesses a global data perspective and powerful computing capabilities. The other two core modules of this invention, the "potential field element learning and morphology injection module" and the "mental experience solidification and aggregation module," are deployed in the central cloud. The former's task is to perform non-real-time, long-term strategy learning. It analyzes the system's emerging operational effects at the macro level and reverse-engineers the optimal guidance strategy to achieve a future scheduling goal (such as increasing the next day's renewable energy consumption rate). The latter's task is to realize the evolution of swarm intelligence. It periodically collects and aggregates the "good behaviors" spontaneously formed by massive amounts of equipment during operation and solidifies them into a new generation of basic models, thereby achieving continuous iteration and optimization of the entire ecosystem's behavioral patterns. These tasks are computationally intensive and require global data, making them best suited for execution in the cloud.

[0034] The system architecture of this invention forms a clear bidirectional information flow of "control downlink" and "data uplink". In the control downlink, the central cloud generates an "injection morphological potential field" representing the macro-strategic intent based on long-term learning. This strategy is then distributed to designated edge computing nodes. Upon receiving the strategy, the edge computing nodes combine it with multiple sub-potential fields generated from real-time collected power grid operating parameters. By dynamically weighting and superimposing, a final digital potential field with tactical guidance significance is synthesized. Its synthesis method can be characterized by the following formula:

[0035] in, For spatial location, For time, For the first The dynamic weights of the sub-potential fields are then assigned. Subsequently, the edge computing nodes apply this to the total potential field. It broadcasts to all distributed energy devices within its coverage area.

[0036] In the data uplink, after each distributed energy device completes its autonomous decision-making, it anonymizes and uploads its operational data, internal state data, and behavioral parameter set adjusted by "individual behavior drift." These data streams converge at the central cloud, providing input for the "potential field element learning and morphological injection module" to evaluate the macroscopic emergence effect, and also providing valuable samples for the "mental experience solidification and aggregation module" for group experience learning, thus forming a complete intelligent closed loop of system perception, decision-making, execution, learning, and optimization.

[0037] In this embodiment, the generation and broadcasting mechanism of the digital potential field is described in detail. This mechanism is one of the core components of the present invention. It is executed on edge computing nodes and is responsible for transforming the multi-source, heterogeneous power grid status and scheduling intentions into a unified, quantifiable environmental guidance signal that can be understood and responded to by all distributed energy devices.

[0038] The process begins with the real-time acquisition and preprocessing of panoramic status data. Edge computing nodes continuously collect key operational data from the distribution network areas under their jurisdiction through their IoT sensing networks. This data mainly includes: grid frequency, which characterizes the overall supply and demand balance of the system. Voltage amplitude at key nodes reflecting local power quality Power flow and load rate of each line, indicating network health. ; and real-time electricity price signals from the electricity market. To ensure the accuracy and consistency of subsequent calculations, edge computing nodes first perform necessary preprocessing on the collected raw data, such as synchronizing data from different sources through timestamp alignment and normalizing different physical quantities.

[0039] Based on the preprocessed data, the "multidimensional digital potential field generator" of the edge computing node will calculate multiple sub-potential fields with clear physical or economic significance. Each sub-potential field is a mathematical representation of a specific scheduling target, and together they constitute a multidimensional guiding force.

[0040] In one specific implementation of the present invention, a frequency-stabilized potential field is used to respond to system frequency deviations. It is generated. Its purpose is to guide the distributed resource cluster to provide fast virtual inertia and primary frequency modulation response. This potential field is calculated using the following formula:

[0041] in, It is the real-time collected power grid frequency. It is the nominal frequency of the power grid; This is the frequency deviation; It is an amplitude coefficient that determines the maximum intensity that the potential field can reach; It is a sensitivity coefficient that defines the steepness of the change in potential field strength with frequency deviation. A hyperbolic tangent function is used. This is because it can smoothly simulate the saturation characteristics of a real generator speed governor, meaning that the response is close to linear when the frequency deviation is small, while the response strength tends to an upper limit when the deviation is too large. The application of the negative sign ensures that when the frequency decreases, the potential field value is positive, forming an energy "attractive" potential, and conversely, a "repulsive" potential is formed.

[0042] Simultaneously, a voltage-supported potential field is used to solve the local voltage problem. It is also generated. This potential field has spatial distribution characteristics, and its intensity varies with geographical location. Its calculation formula is as follows:

[0043] In this formula, At the power grid node Location and Time The real-time voltage amplitude, This is the nominal voltage amplitude of the node; It is an amplitude coefficient used to adjust the strength of the voltage support response. This formula directly converts the per-unit voltage deviation into the potential field strength. When the local voltage is low, the potential field value at that location is positive, attracting distributed equipment in the area to generate reactive power or adjust active power to raise the voltage.

[0044] Furthermore, in order to make the operation of distributed devices economical, an economic incentive potential field is needed. It was introduced. It transforms the abstract electricity price signal into a concrete energy guidance signal, the expression of which is:

[0045] in, This is the current real-time electricity price. It is a benchmark electricity price used for normalization. It is a coefficient used to adjust the sensitivity of economic response. When the electricity price is higher than the benchmark, the potential field is positive, which encourages energy storage to discharge and photovoltaic power generation; conversely, it encourages energy storage to charge or load absorption.

[0046] To prevent and alleviate line congestion in the distribution network, this invention also generates a network congestion potential field. It creates a virtual energy "repulsive field" around the line that is about to be overloaded, and its generation method takes into account the influence of multiple lines:

[0047] In this expression, the summation term iterates through all the monitored lines. . This is the real-time load rate of line kk. It is the preset congestion alarm load rate threshold. It is a modified linear unit function that ensures that a repulsive potential is generated only when the load rate exceeds a threshold. Representing spatial points To the line The physical or electrical distance. It is a Gaussian function that makes the strength of the repulsive potential decrease smoothly with increasing distance around the line, and the rate of decrease is determined by the parameter. control. It is the intensity coefficient of the congestion potential field.

[0048] Ultimately, the edge computing nodes will integrate the aforementioned sub-potential fields, as well as the injected morphological potential fields that may be received from the cloud. By dynamically weighting, a unified total potential field is formed. The weights are not fixed but dynamically adjusted by the edge nodes based on the current operating conditions of the power grid. For example, when a severe disturbance such as an emergency frequency drop is detected, the edge node will immediately increase the weight corresponding to the frequency stabilization potential field. At the same time, appropriately reduce the weight corresponding to the economic incentive potential field. This ensures that the safety and stability of the system outweigh economic objectives. The synthesized total potential field data, after efficient encoding, is broadcast to all distributed energy devices within its coverage area via low-latency communication networks such as 5G, NB-IoT, or power line carrier.

[0049] In this embodiment, the decision-making mechanism of the distributed energy device based on its autonomous mental model is described in detail. When the digital potential field signal broadcast by the edge computing node arrives at the device, the device does not execute passively, but rather uses an "autonomous mental model" embedded in its local controller to combine the guidance of the external environment with its own internal operating state and personalized characteristics, thereby autonomously and safely calculating the final execution power.

[0050] First, the internal structure and initialization process of the autonomous mental model are explained. This model is implemented at the software level, and its core is an evolvable set of behavioral parameters. This parameter set is not a fixed set of values, but rather the key element defining the unique "personality" of the device. In a specific implementation, this behavior is the parameter set. It explicitly includes: one or more response coefficients These correspond to different types of sub-potential fields, used to quantify the device's sensitivity to specific guiding signals (such as frequency and voltage); a risk preference coefficient, used to adjust the conservative or aggressive nature of its behavior during decision-making; and a set of internal state thresholds closely related to the device's physical safety, which, for example, for energy storage devices, includes the safe upper limit of its state of charge. and lower limit and maximum allowable operating temperature Before a new device is connected to the power grid, its mental model is initialized, typically by loading a set of basic behavioral parameters trained in the cloud based on historical operational data of similar devices. This ensures that the device has reasonable, proven basic response capabilities from the outset.

[0051] Next, using a Battery Energy Storage System (BESS) as an example, we will describe in detail its complete decision-making process from sensing to execution. This process is triggered within each scheduling cycle, ensuring the device's rapid response to environmental changes.

[0052] The first step is to generate the initial intended power. When the device... At any moment Received its location Total potential field strength Then, the autonomous mental model will first generate an unconstrained, ideal power response value, i.e., the initial intention power, based on its core response characteristics. The calculation is performed by the following formula:

[0053] in, This is the set of behavioral parameters for the energy storage device. The comprehensive response coefficient represents the intensity of the device's willingness to respond. This step transforms the abstract potential field signal into a concrete power dimension for the first time, reflecting the device's initial interpretation of the external environment's guidance.

[0054] The second step is to perform rigorous quantification of internal state constraints. The initial intended power is merely an ideal value; it must be verified by the device's own physical safety and operational logic. The mental model reads the device's internal state in real time and transforms it into an instantaneous power feasible region. The main constraints include: Equipment rated power constraint: The output power must never exceed the limit specified on its hardware nameplate, i.e., it must meet the following requirements. ,in This represents the maximum charging power (usually a negative value). This represents the maximum discharge power.

[0055] State of Charge (SOC) safety constraints: To prevent battery overcharging or over-discharging, thereby ensuring its safety and lifespan, the mental model calculates the safety constraints for the next scheduling cycle Δt based on the current SOC, ensuring that the SOC does not exceed the safety boundary. The maximum permissible charge and discharge power. This power limit can be calculated using the following formula:

[0056]

[0057] in, This is the current state of charge. This is the battery's rated capacity. This ensures that the upcoming power command will not exceed the rated capacity. This can lead to SOC exceeding the limit within a certain timeframe.

[0058] In addition, it includes other protective logic such as device temperature constraints. For example, when the battery temperature is detected... When a certain preset threshold is exceeded, the mental model will dynamically tighten the rated power constraint or SOC constraint based on the built-in power derating curve, further limiting the power output.

[0059] The third step is to generate the final decision power. The mental model performs an intersection operation on all the constraints calculated in the second step to obtain a final, absolutely safe, instantaneous power feasible interval. Then, it will generate the initial intention power in the first step. Projecting onto this feasible interval can be characterized by a clip function:

[0060] This function means that if the initial intended power is within the feasible range, the final decision power is equal to that intended power; if the intended power exceeds the upper or lower limit of the range, the final decision power will be forcibly set to the nearest range boundary value. In this way, the present invention ensures that while each distributed energy device actively responds to grid coordination needs, its own safety and physical constraints remain the highest and inviolable principles. The final decision power is... The power conversion system (PCS) of the device will be sent to execute the command.

[0061] In this embodiment, the unique dual-closed-loop intelligent evolution mechanism of this invention is described in detail. This mechanism is the key to the long-term adaptive and self-improvement capabilities of this scheduling method. It consists of a top-down "potential element learning" closed loop aimed at forward guidance and a bottom-up "mental evolution" closed loop aimed at individual adaptation and group evolution.

[0062] First, we will explain the top-down potential field learning and forward-looking guidance mechanism. This mechanism operates in a central cloud environment. Its core idea is to treat the entire regional power grid, composed of edge nodes and a large number of distributed devices, as a complex, interactive environment. Through macroscopic observation and learning, it reverse-optimizes and designs strategies that can guide this "environment" to evolve towards a preset goal, i.e., "injecting morphological potential fields." This process was structured as a deep reinforcement learning problem.

[0063] In this reinforcement learning framework, the "potential field element learning and morphology injection module" in the central cloud acts as a learning agent. The observed states represent the system's macroscopic historical information. This includes high-dimensional aggregated indicators such as the mean and variance of grid frequency over a period of time, voltage qualification rate, renewable energy absorption rate, and total operating cost. Its actions do not directly control any equipment, but rather generate and distribute an injection potential field with a specific spatial and temporal dynamic. The system's reward is defined as an instantaneous reward function that quantifies the system's macroscopic performance. Here is an example:

[0064] in, It is the absolute value of the frequency deviation. It is the total operating cost of the system. It is the voltage qualification rate; , , These are the reward weights for the three objectives: frequency stability, economy, and voltage quality. It is a function that maps physical metrics to reward values, such as a Gaussian function. The goal of learning is to find an optimal policy. This strategy can transform the macroscopic state Mapped to the optimal injection potential field This maximizes long-term cumulative discount rewards.

[0065] in It is a discount factor. This optimization problem can be solved using advanced deep reinforcement learning algorithms such as Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO). Through this mechanism, the system gains "foresight" beyond real-time response, enabling it to generate and inject a guiding potential field several hours in advance for a long-term goal (such as "improving midday photovoltaic absorption"), subtly changing the state of charge of energy storage and other equipment in the region, thus achieving forward-looking deployment.

[0066] Secondly, the mechanism of bottom-up individual behavioral drift and group experience solidification is explained. This mechanism endows the system with the ability to self-correct and evolve at the micro level. It includes two interrelated links: The first stage involves "individual behavior drift" occurring at each device level. Each device's autonomous mental model is not static; it undergoes continuous fine-tuning based on its own operational experience. This fine-tuning is achieved through a local cost function defined locally on the device. Driven by this. This cost function is used to penalize behaviors that are detrimental to the device itself, and a specific embodiment of it can be expressed as:

[0067] The first item is used to penalize equipment overheating caused by excessive power or harsh environments. Exceeding the safety threshold The first item is the behavior of the battery; the second item is to suppress excessively violent or frequent charging and discharging by penalizing the square of the power, so as to slow down the degradation of battery health (SOH). and These are the respective penalty weights. The device calculates this cost function for its set of behavioral parameters. The gradient is calculated, and the parameters are updated using gradient descent:

[0068] in It is a relatively small individual learning rate. In this way, the behavior pattern of each device will "drift", gradually becoming more adapted to the specific operating conditions of its location and its own aging state.

[0069] The second stage is the "collective experience consolidation" that occurs in the central cloud. Individual behavioral drift is a small-scale, random exploration, while collective experience consolidation transforms these "advantageous mutations" from exploration into "genetic advantages" for the entire population. The cloud-based "mental experience consolidation and aggregation module" periodically collects a large set of behavioral parameters uploaded by devices after drift adjustments. It first selects a high-performing cluster of devices based on their overall performance (such as task response accuracy and self-loss rate). Then, aggregation algorithms such as Federated Averaging are used to perform a weighted average of the behavioral parameters of the cluster, thereby generating a new generation of basic mental model parameters that incorporates numerous excellent experiences. :

[0070] in, Represents the generation of evolution. It is related to equipment Weights related to data contribution or credibility. It will be used as standard firmware for the initialization of newly connected devices or pushed to existing devices for selective upgrades, thus completing a grand cycle from individual innovation to collective evolution, ensuring the long-term vitality and performance improvement of the entire scheduling system.

[0071] To more clearly illustrate the synergistic effect and complete process of the method of the present invention in actual operation, this embodiment will take a typical power grid emergency event—"the output of a regional photovoltaic cluster drops sharply due to a sudden change in weather"—as an example to explain the entire process of the present invention step by step.

[0072] At the initial moment T0 of this scenario, a regional power distribution network is operating stably, with numerous distributed photovoltaic (PV) devices providing the main power support. Suddenly, a large, rapidly moving thick cloud cover obscures the area, causing a precipitous drop in the actual power generation of the PV clusters within a very short period. This sudden and massive supply-demand gap immediately triggers a grid frequency... The sharp decline in [something] seriously threatens the safety and stability of the system.

[0073] At the very next moment, T1, the edge computing node deployed at the distribution network site in this area detected the abnormal frequency drop within milliseconds using its high-precision sensors. The node's "multidimensional digital potential field generator" was instantly activated. It first calculated a frequency-stable potential field of extremely high intensity according to the following formula. :

[0074] Due to the frequency deviation at this time It is a large negative value, and the generated This is a strong positive value, creating an energy "attraction" potential. Simultaneously, the dynamic weight controller within the edge computing node determines the current operating condition as a frequency emergency event and subsequently adjusts the weights corresponding to the frequency stable potential field. Elevate to the highest priority, while also adjusting the weight of non-safety-related potential fields such as economic incentives. It dropped to an extremely low level. A total potential field where frequency recovery is the absolute dominant factor. It is rapidly synthesized and broadcast to all distributed energy devices in the region via a low-latency communication network.

[0075] At time T2, a massive number of distributed energy devices received this digital potential field signal, which contained a strong sense of "attraction." At this point, the self-organizing and cooperative characteristics of this invention are fully demonstrated. For a battery energy storage unit with a state of charge (SOC) of 80%, its autonomous mental model calculates a large initial intended power. Upon discovering that the value was well within its safe operating range, it immediately decided to execute a high-power discharge command close to its rated power, injecting valuable energy into the grid.

[0076] Meanwhile, a charging station charging an electric vehicle also received the same potential field signal. Its mental model also generated an initial intention to discharge. However, during internal state verification, it read the owner's user preference of "must be fully charged before 7 a.m.," or detected that the vehicle's current SOC was low. Based on these higher-level internal constraints, its final decision was not to reverse discharge, but to quickly reduce the ongoing charging power to zero. By stopping the "draining" of the grid, it also contributed to alleviating the supply-demand imbalance.

[0077] Another standby energy storage device with only 20% state of charge (SOC) found its discharge capacity severely limited due to its proximity to the safe lower limit of SOC when calculating the feasible region of instantaneous power. Therefore, even when it received the same strong attracting potential field, its final decision power was safely tailored to a very small value, avoiding the risk of sacrificing its own lifespan to respond to the grid. It is these tens of thousands of diverse devices, guided by a unified potential field, that each made safe and contributing decisions best suited to its own state and capabilities. Their responses converged into a powerful and precise composite supporting power, collectively curbing further declines in grid frequency.

[0078] In the final stage T3 of the event, with the coordinated response of numerous distributed resources, the power grid's supply-demand gap was effectively filled, and the frequency began to stabilize and recover. Gradually recover to nominal value Nearby, frequency-stable potential field The intensity also naturally decays to zero, and the total potential field... Upon returning to normal operation, the entire system smoothly transitioned back to its regular operating mode. After the event, this successful collaborative response process, as a complete "real-world record," was uploaded to the central cloud. On one hand, the meta-learning module will analyze the entire process of this event to optimize its "injection of morphological potential field" strategy for dealing with similar disturbances in the future. On the other hand, the "behavioral parameter sets" of devices that responded quickly and accurately in this event will be identified as superior samples. Through a group experience solidification mechanism, their characteristics will be refined and integrated into a new generation of basic mental models.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A distributed energy dispatching method empowered by IoT edge computing, characterized in that, Includes the following steps: At the edge computing node, a digital potential field representing the intended energy demand is generated based on at least one collected power grid operating parameter. The edge computing node broadcasts the digital potential field to multiple distributed energy devices within its coverage area; Each of the plurality of distributed energy devices autonomously determines its operating power based on the digital potential field it receives and its own internal state.

2. The method according to claim 1, characterized in that, The step of generating a digital potential field representing the intended energy demand specifically includes: At least one sub-potential field is generated based on the at least one power grid operating parameter; And the at least one sub-potential field is weighted and superimposed to form the digital potential field.

3. The method according to claim 2, characterized in that, The at least one subpotential field includes at least one of the following: The frequency-stabilizing potential field generated based on the power grid frequency deviation; The voltage support potential field generated based on the node voltage deviation; The economic incentive potential field generated based on real-time electricity price signals; Network congestion potential field generated based on line load rate.

4. The method according to any one of claims 1, 2, or 3, characterized in that, The step of generating a digital potential field representing the intended energy demand also includes: Receive an injected morphological potential field generated by the meta-learning module from the cloud; The injected morphological potential field is superimposed on the digital potential field; The meta-learning module is used to learn and generate the injection morphological potential field aimed at achieving a preset scheduling target based on the macroscopic operating effect of the system that emerges after the multiple distributed energy devices autonomously determine their operating power.

5. The method according to claim 1, characterized in that, The step of each device autonomously determining its operating power specifically includes: The device has a built-in autonomous mental model, which stores a set of behavioral parameters that define its response personality; The autonomous mental model generates an initial intention power based on the digital potential field and the set of behavioral parameters; The autonomous mental model constrains the initial intention power based on the internal state of the device, thereby determining the final operating power.

6. The method according to claim 5, characterized in that, The internal state includes at least one of the device's state of charge, health status, device temperature, or user preset preferences.

7. The method according to claim 5, characterized in that, It also includes the step of individual behavior drift: the device autonomously adjusts its set of behavioral parameters based on the internal state feedback generated after it performs its operation at its operating power.

8. The method according to claim 7, characterized in that, It also includes the step of solidifying group experience: In the cloud, the adjusted set of behavioral parameters from multiple devices is aggregated; Based on the aggregation results, superior behavioral parameter features are identified and extracted to generate an updated set of basic behavioral parameters; The updated set of basic behavioral parameters is used to initialize new devices or upgrade existing devices.

9. The method according to claim 4, characterized in that, The meta-learning module maps the historical macroscopic state of the system to the injected morphological potential field through a deep reinforcement learning algorithm, so as to maximize the predefined long-term cumulative system reward.

10. The method according to claim 8, characterized in that, The step of aggregating behavioral parameter sets from multiple devices employs a federated learning aggregation method.

Images