A cloud-edge-end-along four-layer distributed energy collaborative optimization architecture and method based on swarm intelligence

Abstract

This invention relates to a four-layer distributed energy collaborative optimization architecture and method based on swarm intelligence, belonging to the field of distributed energy collaborative optimization technology. The technical solution includes a top-down collaborative four-layer structure: cloud-side, edge-side, end-side, and edge-side. This four-layer structure forms a hierarchical intelligent operation mode of individual autonomous decision-making on the edge side, regional collaborative optimization on the end side, mutual assistance and regulation on the edge side, and global demand guidance on the cloud side. Intelligent agents integrating customized swarm intelligence algorithms are constructed on the edge, end, and edge sides. The edge side realizes individual perception and closed-loop control of distributed energy, the end side completes regional multi-objective optimization, and the edge side realizes inter-group mutual assistance and dynamic computing power allocation. This invention provides technical support for the large-scale development of distributed energy, improves the renewable energy absorption rate and operational economy, and solves problems such as low collaborative efficiency, poor adaptability, unreasonable computing power allocation, and insufficient revenue potential in existing technologies.

Description

Technical Field

[0001] This invention relates to a four-layer distributed energy collaborative optimization architecture and method based on swarm intelligence (cloud-edge-device-edge), belonging to the field of distributed energy collaborative optimization technology. Specifically, it relates to a four-layer distributed energy collaborative optimization architecture based on swarm intelligence (cloud-edge-device-edge), applicable to collaborative management, optimized scheduling, and multi-market linkage scenarios of various distributed energy sources such as photovoltaics, wind power, energy storage, and adjustable loads, ensuring national energy security and stable operation of the power system. Background Technology

[0002] Currently, the global energy transition has entered a critical stage, and distributed energy, as the core carrier of this transition, is exhibiting a massive, decentralized, and heterogeneous development trend. However, existing distributed energy regulation technologies mostly adopt centralized control strategies, relying excessively on centralized command from a central brain, which has many limitations: First, it is difficult to adapt to the massive, decentralized access of distributed energy sources, placing enormous pressure on the computing power of centralized scheduling, resulting in high response latency and an inability to achieve real-time coordination of distributed energy sources; second, the output of distributed power sources is random and fluctuating, and load demand is uncertain, making it difficult for centralized control to accurately balance the core contradiction between output fluctuations and grid security constraints, easily leading to problems such as insufficient renewable energy consumption and unstable grid operation; third, the existing collaborative architecture lacks a hierarchical computing power allocation mechanism, resulting in serious waste of computing resources, and the underlying equipment often has vendor barriers, poor adaptability, weak scalability, and high commissioning and maintenance costs; fourth, the multi-objective optimization capability is insufficient, failing to take into account multiple needs such as renewable energy consumption, carbon emission reduction, operational economics, and energy utilization, making it difficult to explore the full-scenario benefit potential.

[0003] Swarm intelligence theory, with its core advantages of self-organization, self-learning, and autonomous negotiation, offers a new approach to solving distributed multi-agent collaborative problems. In existing technologies, some distributed energy collaborative architectures attempt to incorporate swarm intelligence algorithms, but these are mostly single- or two-tier architectures, failing to form a complete hierarchical collaborative system. Effective information exchange and decision-making linkage between different levels are lacking, hindering the hierarchical distribution of computing power and resource aggregation optimization. Furthermore, the algorithms lack adaptability to real-world engineering scenarios, failing to address the heterogeneity and volatility characteristics of distributed energy sources through lightweight optimization, and thus failing to meet the plug-and-play and real-time response requirements of underlying devices.

[0004] Therefore, there is an urgent need in this field for a novel distributed energy collaborative optimization architecture that integrates swarm intelligence theory, breaks through the limitations of traditional centralized control, and constructs an intelligent system with hierarchical collaboration, computing power adaptation, equipment compatibility, and multi-objective optimization. This system would solve problems such as low collaborative efficiency, poor adaptability, unreasonable computing power allocation, and insufficient profit potential in existing technologies, and provide technical support for the large-scale development of distributed energy. Summary of the Invention

[0005] This invention proposes a cloud-edge-device-edge four-layer distributed energy collaborative optimization architecture and method based on swarm intelligence. It constructs an intelligent system with hierarchical collaboration, computing power adaptation, device compatibility, and multi-objective optimization, providing technical support for the large-scale development of distributed energy, improving the renewable energy consumption rate and operational economy, adapting to the needs of massive heterogeneous distributed energy access, and solving problems such as low collaborative efficiency, poor adaptability, unreasonable computing power allocation, and insufficient revenue potential in existing technologies.

[0006] The technical solution of this invention is:

[0007] A four-layer distributed energy collaborative optimization architecture based on swarm intelligence (cloud-edge-device-edge) comprises a top-down, collaborative structure of cloud-side, edge-side, device-side, and edge-side. This four-layer structure forms a hierarchical intelligent operation mode of "individual autonomous decision-making on the edge-side, regional collaborative optimization on the device-side, swarm mutual assistance and regulation on the edge-side, and global demand guidance on the cloud-side." Intelligent agents integrating customized swarm intelligence algorithms are constructed on the edge, device, and device sides. The edge side achieves individual perception and closed-loop control of distributed energy resources, the device side completes regional multi-objective optimization, and the edge side achieves inter-swarm mutual assistance and dynamic computing power allocation. The cloud side is built on a basic platform, reusing its original data output interface to provide global demand data for the edge-side intelligent agents. This global demand data includes quantitative data on global grid security constraints, quantitative data on regional load forecasting, and quantitative data on electricity market response.

[0008] Furthermore, the basic platform includes a power grid dispatch center, a power trading center, etc.

[0009] Furthermore, the side-by-side, end-by-end, and edge-by-edge are each independently constructed as intelligent agents integrating swarm intelligence algorithms. Each intelligent agent is configured with appropriate hardware modules and algorithm models according to the differences in hierarchical functions, and has the ability to make autonomous decisions, interact with information, and execute instructions. The swarm intelligence algorithm is an intelligent optimization algorithm adapted to distributed multi-agent collaborative scenarios. The swarm intelligence algorithm supports the hierarchical collaborative control of intelligent agents at each level through self-organization, self-learning, and autonomous negotiation.

[0010] Furthermore, the edge-aware control intelligent agent is bound one-to-one with the distributed energy individual. As a dedicated intelligent agent between the distributed energy physical device and the edge-side intelligent agent, it is responsible for completing the real-time status acquisition, autonomous closed-loop control and cross-level information interaction of the distributed energy individual.

[0011] The end-side is a distributed energy zone aggregation intelligent agent, which aggregates distributed energy individuals and their corresponding side-side intelligent agents to form a distributed energy zone by dividing them according to voltage level or geographical area. The core function of the end-side distributed energy zone aggregation intelligent agent is to complete the information interaction, autonomous negotiation and autonomous decision-making of multiple side-side intelligent agents in the zone, and realize the mutual assistance between distributed energy zones and the multi-objective optimization and control within the zone.

[0012] The edge side is a distributed energy cluster aggregation intelligent agent, which aggregates distributed energy areas and their corresponding edge-side intelligent agents to form a distributed energy cluster by dividing them according to voltage level or geographical region. The edge-side distributed energy cluster aggregation intelligent agent connects to the cloud side to obtain and parse global demand data, autonomously generates cluster-level response plans and output optimization plans, completes autonomous information interaction and decision command implementation between edge-side intelligent agents in the cluster, and realizes mutual assistance between distributed energy clusters and secondary optimization and control within the cluster under the premise of meeting the global demand of the cloud side and the power grid security constraints.

[0013] Furthermore, the swarm intelligence algorithm is an intelligent optimization algorithm adapted to distributed multi-agent collaborative scenarios, including but not limited to one of particle swarm optimization, ant colony optimization, and genetic algorithm, or an operator-level / model-level fusion algorithm of the above algorithms. The operator-level fusion algorithm combines and optimizes the core operators of each algorithm, while the model-level fusion algorithm nests or parallels the overall model of each algorithm for optimization.

[0014] Furthermore, the edge-sensing control agent along the side is a lightweight, adaptable agent deployed at the local control end of the distributed energy unit, integrating a high-precision edge sensing module, a multi-protocol conversion module, and a high-precision acquisition and control module. The high-precision edge sensing module is used to collect the electrical operating parameters and environmental sensing parameters of the distributed energy unit. The electrical operating parameters include voltage, current, active power, reactive power, and energy storage state of charge; the environmental sensing parameters include photovoltaic irradiance, ambient temperature, wind speed, and light intensity. The multi-protocol conversion module integrates an industrial-standard hardware communication interface and has a built-in customized and optimized protocol parsing engine. The system automatically identifies the common industrial communication protocols of various distributed energy converters, extracts and converts the core fields of communication data packets, and enables parallel parsing and priority sorting of multiple protocols, thereby improving the efficiency and accuracy of protocol parsing. The high-precision acquisition and control module is the execution unit of the edge-sensing control agent, used to realize the real-time acquisition of individual operating parameters and status parameters of distributed energy sources and the execution of autonomous closed-loop control. The cross-level information interaction between the edge-sensing control agent on the edge and the distributed energy zone aggregation agent on the end side specifically includes standardized operating parameters acquired on the edge, equipment status feedback data uploaded to the end side, and control commands obtained from the end side.

[0015] Furthermore, the distributed energy entities include photovoltaic, energy storage, adjustable loads, and wind power equipment, and can be extended to other types of distributed energy equipment.

[0016] A distributed energy collaborative optimization method based on the above-mentioned cloud-edge-device-edge four-layer architecture, wherein the edge-aware control agent of the edge is equipped with a lightweight swarm intelligence algorithm, which is a collaborative optimization algorithm based on the fusion of simplified particle swarm algorithm and neural network;

[0017] The lightweight adaptation method of the lightweight swarm intelligence algorithm is as follows: nodes with iteration errors less than a preset threshold (the preset threshold is set according to 1%-3% of the rated output of individual distributed energy resources) are used as redundant iteration nodes and pruned, and only the iteration nodes of local optimal solutions are retained; at the same time, the multi-dimensional parameters collected by the high-precision edge perception module are compressed by principal component analysis, core feature dimensions are selected, and redundant feature dimensions are eliminated.

[0018] The neural network adopts a feedforward neural network structure, with the number of neurons in its input layer being consistent with the number of core feature dimensions obtained after principal component analysis, and the number of neurons in its output layer being 1. The neural network is used to make short-term predictions of the output trend of individual distributed energy sources on a time scale of 5-15 minutes, and to obtain the predicted output value.

[0019] The predicted output value of the neural network is used as the initial input parameter of the lightweight swarm intelligence algorithm; and after each iteration of the lightweight swarm intelligence algorithm completes a local optimum, the algorithm corrects the iteration parameters by adjusting the deviation between the predicted output value and the actual collected value; the deviation is calculated as follows:

[0020]

[0021] in: for The relative deviation in force output at any given moment; To contribute to neural network prediction; To contribute to actual data collection; The rated output of the equipment is set. The specific correction rule is as follows: when the deviation is >10%, the algorithm iteration step size is increased linearly according to the deviation ratio; when the deviation is ≤10%, the algorithm learning rate is fine-tuned to 0.01-0.05.

[0022] Furthermore, the autonomous decision-making process of the distributed energy zone aggregation agent is based on a multi-objective optimization model. The constraints of the multi-objective optimization model include upper and lower limits of distributed energy output, grid voltage / power constraints, and energy storage SOC constraints. The objective functions include four objectives: maximizing renewable energy consumption, minimizing carbon emissions, optimizing operational economy, and maximizing energy utilization. Each objective function is configured with dynamic weight coefficients calculated based on the analytic hierarchy process (AHP). These dynamic weight coefficients are dynamically adjusted according to the renewable energy output ratio of the distributed energy zone, the real-time electricity market trading price, and the carbon quota trading price.

[0023] The adjustment formula for the dynamic weighting coefficient is as follows:

[0024]

[0025] in, The dynamic weighting coefficients of the single objective function are as follows: x corresponds to the four objectives mentioned above; α is the weighting coefficient corresponding to the proportion of new energy output, corresponding to the objective of maximizing new energy consumption; β is the weighting coefficient corresponding to the electricity market transaction price, corresponding to the objective of optimal operating economy; γ is the weighting coefficient corresponding to the carbon quota transaction price, corresponding to the objective of minimizing carbon emissions; δ is the weighting coefficient corresponding to the comprehensive energy utilization coefficient, corresponding to the objective of maximizing energy utilization rate; and satisfies α+β+γ+δ=1; Pn' is the linear normalized value of the proportion of regional new energy output, Pe' is the linear normalized value of the real-time electricity market transaction price, Pc' is the linear normalized value of the carbon quota transaction price, and Pu' is the linear normalized value of the regional energy utilization rate; the values ​​of Pn', Pe', Pc', and Pu' are all in the range of 0-1; the linear normalized value is calculated as follows: linear normalized value = (actual value of the parameter - minimum value of the parameter) / (maximum value of the parameter - minimum value of the parameter); the maximum and minimum values ​​of each parameter are taken as the historical statistical extreme values ​​of the past 30 natural days.

[0026] Furthermore, the analytic hierarchy process (AHP) index system includes four primary indicators: new energy output volatility, market electricity price fluctuation rate, carbon quota trading return rate, and comprehensive energy utilization rate. The weight values ​​of the four primary indicators are calculated based on the AHP, and these weight values ​​are directly converted into specific values ​​of α, β, γ, and δ.

[0027] The objective function of the multi-objective optimization model is solved iteratively by a multi-objective collaborative swarm intelligence algorithm mounted on the edge agent. The algorithm iteration terminates when the objective function converges to a preset accuracy range with the total active power deviation of the distributed energy zone as the calculation dimension, or when the number of algorithm iterations reaches a preset maximum number of iterations. The total active power deviation is calculated as follows: Total active power deviation = |Total actual active power output of the zone - Total planned active power output of the zone| / Total scheduled active power output of the zone. The preset maximum number of iterations is set according to the scale of the distributed energy zone; the more distributed energy units there are, the larger the number of iterations.

[0028] The multi-objective collaborative swarm intelligence algorithm carried by the distributed energy zone aggregation intelligent agent on the end side is an operator layer fusion algorithm based on non-dominated sorting genetic algorithm and ant colony algorithm, and integrates a convolutional neural network (CNN) feature extraction model for data feature extraction.

[0029] The feature extraction model is used to extract features from the interaction data of multiple side agents within the area collected by the edge agent. The core feature dimensions extracted include output deviation, adjustment response time, and resource utilization rate, and each feature is a quantitative feature. Among them, the output deviation is calculated according to the above method, the adjustment response time = decision command issuance time - equipment response completion time, and the resource utilization rate = actual energy utilization / total available energy × 100%. The quantitative feature dimension results extracted by the feature extraction model are used as the initial input parameters of the fusion algorithm.

[0030] The fusion of the non-dominated sorting genetic algorithm and the ant colony algorithm at the operator layer specifically involves fusing the selection and crossover operators of the non-dominated sorting genetic algorithm with the path search and pheromone update operators of the ant colony algorithm. The fusion rule is as follows: based on the three core dimensions of quantified features extracted by the feature extraction model, an initial population is generated through the selection and crossover operators. The path search operator performs local optimization based on the initial population, and the pheromone update operator updates the pheromone according to the local optimization results. The population iteration and path optimization are achieved through iterative loops, and the optimal solution, i.e., the multi-objective optimization and control instruction within the region, is output as the decision instruction for the edge agent.

[0031] Furthermore, the edge-side distributed energy cluster aggregation intelligent agent is equipped with a load sensing module and a computing power allocation module; the load sensing module provides real-time quantitative data input to the computing power allocation module, the computing power allocation module generates computing power allocation scheduling data, and the edge-side intelligent agent dynamically adjusts and optimizes the computing strategy through the computing power allocation scheduling data; the load sensing module and the computing power allocation module are the core functional modules for the edge-side intelligent agent to realize secondary optimization and control within the cluster;

[0032] The load sensing module is used to collect two types of quantitative data in real time: the overall operating load parameters of the distributed energy group and the quantitative indicators of computing power demand for optimization tasks of each end-side intelligent agent within the group. The overall operating load parameters include the total active power, total reactive power, and group load rate, where the group load rate = total active power / rated total active power of the group × 100%. The quantitative indicators of computing power demand include the amount of task data, the number of algorithm iterations, and the computing power occupancy rate.

[0033] The computing power allocation module dynamically allocates computing power to each edge agent within the group based on a two-factor weighting algorithm of "load ratio - task complexity"; the calculation formula for the computing power allocation ratio is as follows: ;in, This represents the computing power allocation ratio for a single edge agent, with a value ranging from 0 to 1. The proportion of the distributed energy zone load corresponding to this edge agent to the total group load is calculated as follows: =Total active power of the district / Total active power of the group × 100%; When the total active power of the group is 0 , The number of endpoint agents within the group; The proportion of the optimization task complexity of this edge agent to the total task complexity of the group; This is the load proportion weighting coefficient. Let be the task complexity weight coefficient, and satisfy . The The value is dynamically adjusted based on the power grid operating conditions; the power grid operating conditions are quantitatively defined according to the group-level load rate and adjusted accordingly for different operating conditions. The value of makes the allocation of computing power more in line with the needs of power grid operation;

[0034] The task complexity is quantified using a weighted summation method. The calculation formula is: Task complexity = 0.4 × Task data volume ratio + 0.3 × Algorithm iteration count ratio + 0.3 × Computing power utilization ratio. Each ratio is the ratio of the indicator corresponding to a single edge agent to the sum of the indicators corresponding to all edge agents in the group, and all are normalized, with a value range of 0-1.

[0035] The computing power allocation module also performs a secondary dynamic adjustment of the computing power allocation ratio of each end-side agent in the group based on the grid constraint verification results of the edge-side agents: for end-side agents whose optimization results have over-constraint items, the corresponding ratio is increased based on the original computing power allocation ratio; for end-side agents whose optimization results have no over-constraint items, the corresponding ratio is decreased based on the original computing power allocation ratio; after adjustment, the sum of the computing power allocation ratios of all end-side agents in the group is still 1; the over-constraint items are items whose edge-side optimization results exceed the cloud-side global grid security constraint quantification threshold, and the global grid security constraint quantification threshold is clearly defined by the cloud-side global demand data.

[0036] Furthermore, the swarm intelligence algorithm carried by the distributed energy swarm aggregation intelligent agent on the side is a grid constraint-adaptive enhancement optimization algorithm based on ant colony algorithm fusion feature extraction model and deep reinforcement learning model.

[0037] The input parameters of the algorithm are the group-level operating load parameters collected by the load sensing module and the computing power allocation ratio results of the computing power allocation module. The number of algorithm iterations is dynamically and adaptively adjusted according to the computing power allocation ratio. The higher the computing power allocation ratio, the larger the number of iterations. The number of iterations on the side algorithm is matched with the number of iterations on the end side within the group.

[0038] The feature extraction model is constructed using a random forest model, which is used to extract features from inter-group interaction data collected by edge agents and global power grid constraint data sent from the cloud; the core of the extraction is 1214.

[0039] The quantitative characteristics include inter-group output deviation, grid constraint margin, and load forecast deviation;

[0040] The deep reinforcement learning model uses the core quantitative features extracted by the feature extraction model as the model input and maximizes the inter-group collaborative decision matching degree as the model reward function. After offline training and online optimization, it outputs the inter-group collaborative decision weight coefficient, i.e., the group decision factor. The offline training is based on the historical data of inter-group interactions and power grid constraints over the past 30 natural days, and the online optimization dynamically adjusts the model parameters based on real-time operation data. The inter-group collaborative decision matching degree = inter-group decision instruction fit degree × inter-group resource complementarity degree, with a value range of 0-1. The group decision factor is the priority weight of inter-group collaborative decision.

[0041] The group decision factor is directly integrated into the pheromone update rule of the ant colony algorithm; the pheromone update formula of the ant colony algorithm is: τ ij (t+1)=(1-ρ)×τ ij (t)+ρ×Q ij ×λ; where τ ij (t+1) represents the pheromone concentration of path ij at time t+1, τ ij(t) represents the initial pheromone concentration at time t along path ij, ρ is the pheromone evaporation coefficient, and Q... ij Q is the fitness value of path ij, which is linearly positively correlated with the inter-group collaborative decision-making matching degree. The higher the matching degree, the better Q. ij The larger the value, the more λ becomes the group decision factor, used to adjust the priority of collaborative decision-making among groups.

[0042] The cloud side refers to existing platforms such as dispatch centers and trading centers, which will not be modified and will directly interface with the "edge side" to provide global demand data for edge-side intelligent agents. The global demand data includes global power grid security constraint quantitative data, regional load forecast quantitative data, and power market response quantitative data.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] (1) The present invention provides a four-layer collaborative optimization architecture based on swarm intelligence, namely “cloud-edge-end-edge”, which forms a hierarchical operation mode of “individual autonomous decision-making on the edge side - regional collaborative optimization on the end side - mutual assistance and regulation on the edge side - global demand guidance on the cloud side”. It realizes the collaborative optimization of massive, decentralized and heterogeneous distributed energy, solves the core contradiction between the output fluctuation of distributed power sources, random load fluctuation and grid security constraints, provides key technical support for energy transformation, and has important strategic significance for ensuring national energy security and stability.

[0045] (2) Each level of this invention is constructed as an intelligent agent integrating swarm intelligence algorithms, which has autonomy, adaptability and swarm interaction. It can complete complex collaborative tasks through self-organization, self-learning and autonomous negotiation. The algorithms along the side, end side and edge side are all customized and optimized for the hierarchical functions, such as the lightweight algorithm along the side, the multi-objective fusion algorithm at the end side and the power grid constraint adaptation algorithm at the edge side, which realizes the precise adaptation of the algorithm to the engineering scenario and improves the efficiency and accuracy of collaborative optimization.

[0046] (3) The present invention integrates a multi-protocol conversion module along the side agent, realizing plug-and-play for complex underlying devices of multiple manufacturers and types of converters, solving the problems of high barriers and poor adaptability of existing equipment manufacturers. At the same time, it has the advantages of lightweight equipment, low debugging and maintenance costs, and strong scalability, and can flexibly adapt to the access needs of various distributed energy entities. The end agent constructs a multi-objective optimization model, and dynamically adjusts the weight of the objective function by combining the analytic hierarchy process, taking into account multiple needs such as new energy consumption, carbon emission reduction, operation economy, and energy utilization rate, and realizing multi-objective collaborative optimization of the distributed energy area. The edge agent is equipped with a load perception and computing power allocation module, and realizes dynamic allocation of computing power based on the "load ratio-task complexity" two-factor algorithm, which improves the utilization rate of computing power resources and ensures the real-time and accuracy of optimization decisions.

[0047] (4) This invention aggregates fragmented distributed energy resources into a distributed energy cluster that is schedulable, measurable, tradable, autonomous, and mutually supportive through the hierarchical distribution of computing power and the optimization of resource aggregation. This achieves "distributed access and clustered management and control", strengthens the overall support capability of the distributed energy network, and adapts to the trend of large-scale and diversified development of distributed energy. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall structure of the "cloud-edge-device-edge" four-layer distributed energy collaborative optimization architecture based on swarm intelligence, according to an embodiment of the present invention.

[0049] Figure 2 This is a schematic diagram of the hardware module structure of the side-mounted intelligent agent according to an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the iterative correction process of the lightweight swarm intelligence algorithm for side-by-side agents in an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram of the solution process for the edge-side intelligent agent multi-objective optimization model according to an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram illustrating the process of allocating computing power to the side-side intelligent agent according to an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of the iterative process of the side agent swarm intelligence algorithm in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0055] In this embodiment, as shown in the figure, a four-layer distributed energy collaborative optimization architecture based on swarm intelligence—"cloud-edge-device-edge"—is provided. This architecture includes a four-layer hierarchical structure of cloud-side, edge-side, device-side, and edge-side components working in synergy. This four-layer structure forms a hierarchical intelligent operation mode of "individual autonomous decision-making on the edge-side, regional collaborative optimization on the device-group mutual assistance and regulation on the edge-side, and global demand guidance on the cloud-side." Based on this four-layer distributed energy collaborative optimization architecture, a distributed energy collaborative optimization method is provided. The details are as follows:

[0056] 1. Implementation of the side agent

[0057] The edge-sensing and control agents are bound one-to-one with individual distributed energy sources. These distributed energy sources utilize photovoltaic modules, energy storage devices, adjustable loads, and wind power equipment. The edge-sensing agents are deployed at the local control terminals of each distributed energy source, integrating high-precision edge sensing modules, multi-protocol conversion modules, and high-precision data acquisition and control modules. Figure 2 As shown.

[0058] The high-precision edge sensing module collects the electrical operating parameters and environmental sensing parameters of distributed energy units in real time: electrical operating parameters include voltage (acquisition accuracy ±0.5%), current (acquisition accuracy ±0.5%), active power, reactive power, and energy storage state of charge (acquisition accuracy ±1%); environmental sensing parameters include photovoltaic irradiance (acquisition range 0-2000W / m², accuracy ±5%), ambient temperature (acquisition range -40℃-85℃, accuracy ±0.5℃), wind speed (acquisition range 0-60m / s, accuracy ±0.1m / s), and light intensity.

[0059] The multi-protocol conversion module integrates industrial standardized hardware communication interfaces such as RS485, Ethernet, and CAN. It has a built-in customized and optimized protocol parsing engine that can automatically identify industrial common communication protocols of various distributed energy converters such as Modbus, IEC104, and DL / T860. It can complete the extraction and format conversion of core fields of communication data packets, realize multi-protocol parallel parsing and priority sorting, parsing latency ≤100ms, and parsing accuracy ≥99.5%.

[0060] The high-precision acquisition and control module is the core execution unit of the edge-side intelligent agent. Based on the standardized operating parameters output by the multi-protocol conversion module, it completes real-time synchronous acquisition of the operating status of distributed energy individuals. The acquisition frequency is 50Hz, and the control command execution response delay is ≤50ms. The module has built-in closed-loop control logic and can directly receive locally optimal decision commands output by the lightweight swarm intelligence algorithm to realize autonomous closed-loop control of distributed energy individuals, such as output adjustment, start-stop control, and state switching. At the same time, the module collects the status data of distributed energy individuals after their actions, such as actual output value, equipment operating status code, and fault information, and uploads them to the edge-side intelligent agent after standardized format conversion to complete the information interaction for closed-loop control.

[0061] The side-mounted intelligent agent is equipped with a lightweight swarm intelligence algorithm, which is a collaborative optimization algorithm based on the fusion of a simplified particle swarm algorithm and a feedforward neural network. The preset threshold is set at 2% of the rated output of the distributed energy individual. Principal component analysis is performed on the multi-dimensional parameters collected by the high-precision edge perception module to compress the data and select four core feature dimensions: active power, photovoltaic irradiance, energy storage SOC, and ambient temperature, while eliminating redundant feature dimensions. The feedforward neural network has 4 input layer neurons and 1 output layer neuron, which is used to make short-term predictions of the output trend of the distributed energy individual with a time scale of 10 minutes to obtain the output prediction value.

[0062] The algorithm iterative correction process is as follows: Figure 3 As shown: The predicted output value of the neural network is used as the initial input parameter of the lightweight swarm intelligence algorithm. The initial iteration step size is set to 0.05, the initial learning rate is 0.03, and the maximum number of iterations is 100. After each iteration to a local optimum, the algorithm calculates the deviation between the predicted output value and the actual collected value. When the deviation is >10%, the algorithm iteration step size is increased linearly according to the deviation ratio, with the step size adjustment range being 0.01-0.1, while keeping the learning rate unchanged to accelerate the iteration convergence speed. When the deviation is ≤10%, the algorithm learning rate is fine-tuned to 0.03. To maintain iterative stability and avoid over-iteration, when the deviation of three consecutive iterations is ≤10%, the algorithm is considered to have converged and the iteration is terminated. The local optimum of this iteration is output as the final decision instruction. If the convergence condition is not met after 100 iterations, the iteration is terminated immediately, the local optimum of the previous iteration is used as the temporary decision instruction, and the iteration anomaly information is uploaded to the edge agent. Through this correction rule and convergence output judgment, the accuracy, convergence speed and engineering feasibility of the algorithm iteration are ensured, and the algorithm convergence time is ≤500ms.

[0063] 2. Implementation of edge-side intelligent agents

[0064] The end-side is a distributed energy zone aggregation agent. According to the voltage level, 10-50 distributed energy individuals and their corresponding side-side agents are aggregated to form a distributed energy zone. The core function of the end-side agent is to complete the information interaction, autonomous negotiation and autonomous decision-making of multiple side-side agents in the zone, and realize the mutual assistance between distributed energy zones and the multi-objective optimization and control within the zone.

[0065] The autonomous decision-making process of the edge-side intelligent agent is based on a multi-objective optimization model. The constraints include: upper and lower limits of distributed energy output (the upper limit of photovoltaic output is 100% of the rated output, and the lower limit is 0; the upper limit of energy storage charging and discharging power is 100% of the rated power, and the lower limit is -100%; the adjustable load adjustment range is 50%-100% of the rated load), grid voltage constraints (0.95Un≤voltage≤1.05Un, where Un is the rated voltage), power constraints, and energy storage SOC constraints (20%≤SOC≤80%). The objective function includes four major objectives: maximizing the absorption of new energy, minimizing carbon emissions, optimizing operational economy, and maximizing energy utilization.

[0066] The dynamic weight coefficients of each objective function were calculated using the Analytic Hierarchy Process (AHP). The AHP index system includes four primary indicators: the volatility of new energy output, the fluctuation rate of market electricity prices, the return on carbon quota trading, and the comprehensive energy utilization rate. The calculated values ​​are α=0.35, β=0.3, γ=0.2, and δ=0.15, which satisfy α+β+γ+δ=1. The values ​​of Pn', Pe', Pc', and Pu' were calculated by linearly normalizing the historical extreme values ​​of the past 30 natural days, and the values ​​range from 0 to 1.

[0067] The multi-objective collaborative swarm intelligence algorithm carried by the edge agent is an operator-layer fusion algorithm based on the non-dominated sorting genetic algorithm (NSGA-III) and the ant colony algorithm, and integrates a convolutional neural network (CNN) feature extraction model. The feature extraction model extracts features from the interaction data of multiple edge agents within the region, obtaining three core quantitative features: output deviation, adjustment response time, and resource utilization. The solution process of the fusion algorithm is as follows: Figure 4 As shown, using the three core features as initial inputs, an initial population of 100 is generated through the selection and crossover operators of NSGA-III. The path search operator of the ant colony algorithm performs local optimization based on the initial population, and the pheromone update operator updates the pheromone according to the local optimization results. The algorithm terminates when the total active power deviation converges to within ±2% or the number of iterations reaches 200. The optimal solution is output as the decision instruction of the edge agent.

[0068] 3. Implementation of Side Agents

[0069] The edge-side system aggregates 5-100 distributed energy zones to form a distributed energy cluster. The edge-side agents are equipped with a load sensing module and a computing power allocation module. The load sensing module collects the overall operating load parameters of the distributed energy cluster in real time (total active power, total reactive power, and cluster load rate) and the quantitative indicators of computing power requirements (task data volume, number of algorithm iterations, and computing power occupancy rate) of the optimization tasks of each edge-side agent in the cluster. The computing power allocation module dynamically allocates computing power to each edge-side agent in the cluster based on a two-factor weight algorithm of "load ratio - task complexity".

[0070] The formula for calculating the computing power allocation ratio is as follows: , Dynamic adjustments are made based on the power grid operating conditions: when the light load condition, i.e., the group-level load rate, is ≤30%. When the load is heavy, i.e., 30% < group load rate ≤ 70%, When the full load condition, i.e., the group-level load rate, is >70%, ; This represents the proportion of the distributed energy zone load corresponding to this end-side intelligent agent to the total group load. When the total active power at the group level is 0... The number of endpoint agents within the group; The task complexity is the proportion of the optimization task complexity of this edge agent to the total task complexity of the group. Task complexity = 0.4 × task data volume proportion + 0.3 × algorithm iteration number proportion + 0.3 × computing power utilization rate proportion.

[0071] The computing power allocation process is as follows: Figure 5 As shown: The load sensing module collects data and transmits it to the computing power allocation module, which calculates the load ratio of each end-side intelligent agent. With task complexity Determined in conjunction with the power grid operating conditions Calculate the computing power allocation ratio η; at the same time, make a secondary adjustment based on the grid constraint verification results. For edge agents with over-constraint terms in the optimization results, increase the computing power allocation ratio by 5%-10% on the basis of the original ratio. For edge agents without over-constraint terms, decrease the ratio by 5%-10%. After the adjustment, the sum of the computing power allocation ratios of all edge agents in the group is still 1.

[0072] The swarm intelligence algorithm mounted on the edge agent is a grid constraint-adaptive enhancement optimization algorithm based on ant colony optimization, fused with a random forest feature extraction model and a deep reinforcement learning model. The input parameters of the algorithm are the swarm-level operating load parameters collected by the load sensing module and the computing power allocation ratio results from the computing power allocation module. The random forest feature extraction model extracts features from inter-swarm interaction data and global grid constraint data sent from the cloud side to obtain three core quantitative features: inter-swarm output deviation, grid constraint margin, and load prediction deviation. The deep reinforcement learning model takes the three core features as input, uses the maximization of inter-swarm collaborative decision matching degree as the reward function, performs offline training based on historical data of nearly 30 natural days, and dynamically adjusts the model parameters online according to real-time operating data to output the swarm decision factor λ.

[0073] The pheromone update formula for the ant colony algorithm is τ. ij (t+1)=(1-ρ)×τ ij (t)+ρ×Q ij ×λ, where ρ=0.1 (pheromone evaporation coefficient), Q ijQ is linearly positively correlated with the matching degree of inter-group collaborative decision-making. When the matching degree is 1, Q ij =100; when the matching degree is 0, Q =100; ij =0, λ is the group decision factor, with a value range of 0.5-1.5; the algorithm iteration process is as follows: Figure 6 As shown, the number of algorithm iterations is dynamically adjusted according to the computing power allocation ratio. When the computing power allocation ratio is 10%, the number of iterations is 50; when the computing power allocation ratio is 50%, the number of iterations is 100; and when the computing power allocation ratio is 100%, the number of iterations is 200. The number of iterations on the edge side is matched with the number of iterations on the end side within the group to ensure the accuracy and real-time performance of inter-group collaborative decision-making.

[0074] 4. Implementation on the cloud side

[0075] The cloud side reuses existing power grid dispatch center and power trading center platforms without modification, and directly connects with edge-side intelligent agents to provide them with global demand data, including global power grid security constraint quantitative data (voltage constraint threshold, power constraint threshold), regional load forecast quantitative data (forecast accuracy ≥95%), and power market response quantitative data (real-time trading price, carbon quota trading price).

[0076] 5. The collaborative operation process of the architecture

[0077] The collaborative operation process of the architecture in this embodiment is as follows:

[0078] 1) The edge agent collects real-time operating parameters and environmental parameters of the distributed energy units through the high-precision edge perception module, performs protocol parsing and format conversion through the multi-protocol conversion module, and transmits them to the edge agent; at the same time, the edge agent completes autonomous decision-making and closed-loop control of the distributed energy units based on the lightweight swarm intelligence algorithm and combined with the output prediction value.

[0079] 2) The information of each side agent in the aggregation area of ​​the edge agent is used to solve the multi-objective optimization model through a multi-objective collaborative swarm intelligence algorithm to achieve inter-regional mutual assistance and intra-regional optimization;

[0080] 3) The edge agent dynamically allocates computing power based on the global demand data from the cloud side and the grid-constrained adaptive enhancement optimization algorithm to achieve inter-group mutual assistance and intra-group optimization.

[0081] 4) The cloud provides global demand data to edge agents, enabling global demand guidance without requiring them to participate in local collaborative decision-making, thus reducing the computing power pressure on the cloud.

[0082] Through the above-mentioned collaborative operation, the architecture of this embodiment realizes the collaborative optimization and control of massive, decentralized, and heterogeneous distributed energy resources, increases the renewable energy absorption rate by more than 15%, reduces carbon emissions by more than 10%, improves the economic efficiency of operation by more than 8%, significantly enhances the stability of power grid operation, and enables multi-market linkage arbitrage, stimulates the vitality of the electricity market, and fully meets the needs of large-scale development of distributed energy resources and energy transformation.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A four-layer distributed energy collaborative optimization architecture based on swarm intelligence (cloud-edge-device-edge), characterized in that: It comprises a top-down, collaborative four-layer structure: cloud-side, edge-side, terminal-side, and along-side. This four-layer structure forms a hierarchical intelligent operation mode: autonomous decision-making by individuals along-side, regional collaborative optimization by the terminal-side, mutual assistance and regulation by the edge-side, and global demand guidance by the cloud-side. Intelligent agents integrating customized swarm intelligence algorithms are constructed at the along-side, terminal-side, and edge-side levels. The along-side achieves individual perception and closed-loop control of distributed energy resources, the terminal-side completes multi-objective optimization of regions, and the edge-side achieves inter-swarm mutual assistance and dynamic computing power allocation. The cloud-side is built upon a basic platform, reusing its original data output interface to provide global demand data for the edge-side intelligent agents. The global demand data includes quantitative data on global power grid security constraints, quantitative data on regional load forecasts, and quantitative data on electricity market response.

2. The cloud-edge-device-edge four-layer distributed energy collaborative optimization architecture based on swarm intelligence as described in claim 1, characterized in that: The side-by-side, end-by-end, and edge-by-edge are each independently constructed as intelligent agents integrating swarm intelligence algorithms. Each intelligent agent is configured with appropriate hardware modules and algorithm models according to the differences in hierarchical functions, and has the ability to make autonomous decisions, interact with information, and execute instructions. The swarm intelligence algorithm is an intelligent optimization algorithm adapted to distributed multi-agent collaborative scenarios. The swarm intelligence algorithm supports the hierarchical collaborative control of intelligent agents at each level through self-organization, self-learning, and autonomous negotiation.

3. The cloud-edge-device-edge four-layer distributed energy collaborative optimization architecture based on swarm intelligence as described in claim 2, characterized in that: The edge-aware control intelligent agent is bound one-to-one with the distributed energy individual. As the exclusive intelligent agent between the distributed energy physical device and the edge-side intelligent agent, it is responsible for completing the real-time status acquisition, autonomous closed-loop control and cross-level information interaction of the distributed energy individual. The endpoint is a distributed energy zone aggregation intelligent agent, which aggregates distributed energy individuals and their corresponding side-side intelligent agents to form a distributed energy zone by dividing them according to voltage level or geographical area; the core function of the endpoint distributed energy zone aggregation intelligent agent is to complete the information interaction, autonomous negotiation and autonomous decision-making of multiple side-side intelligent agents within the zone, and realize the mutual assistance between distributed energy zones and the multi-objective optimization and control within the zone. The edge side is a distributed energy cluster aggregation intelligent agent, which aggregates distributed energy areas and their corresponding edge-side intelligent agents to form a distributed energy cluster by dividing them according to voltage level or geographical region. The edge-side distributed energy cluster aggregation intelligent agent connects to the cloud side to obtain and parse global demand data, autonomously generates cluster-level response plans and output optimization plans, completes autonomous information interaction and decision command implementation between edge-side intelligent agents in the cluster, and realizes mutual assistance between distributed energy clusters and secondary optimization and control within the cluster under the premise of meeting the global demand of the cloud side and the safety constraints of the power grid.

4. The cloud-edge-device-edge four-layer distributed energy collaborative optimization architecture based on swarm intelligence as described in claim 3, characterized in that: The edge-sensing control agent along the side is a lightweight, adaptable agent deployed at the local control end of the distributed energy unit. It integrates a high-precision edge sensing module, a multi-protocol conversion module, and a high-precision acquisition and control module. The high-precision edge sensing module collects the electrical operating parameters and environmental sensing parameters of the distributed energy unit. The electrical operating parameters include voltage, current, active power, reactive power, and energy storage state of charge; the environmental sensing parameters include photovoltaic irradiance, ambient temperature, wind speed, and solar irradiance. The multi-protocol conversion module integrates an industrial-standard hardware communication interface and has a built-in customized and optimized protocol parsing engine for automatic... The system identifies the common industrial communication protocols of various distributed energy converters, extracts and converts the core fields of communication data packets, and enables parallel parsing and priority sorting of multiple protocols to improve the efficiency and accuracy of protocol parsing. The high-precision acquisition and control module is the execution unit of the edge-sensing control agent, used to realize the real-time acquisition of individual operating parameters and status parameters of distributed energy and the execution of autonomous closed-loop control. The cross-level information interaction between the edge-sensing control agent on the edge and the distributed energy area aggregation agent on the end side specifically includes standardized operating parameters acquired on the edge, equipment status feedback data uploaded to the end side, and control commands obtained from the end side.

5. A four-layer distributed energy collaborative optimization method based on claim 4, characterized in that: The edge-sensing control agent along the side is equipped with a lightweight swarm intelligence algorithm, which is a collaborative optimization algorithm based on the fusion of simplified particle swarm algorithm and neural network. The lightweight adaptation method of the lightweight swarm intelligence algorithm is as follows: nodes with iteration errors less than a preset threshold are used as redundant iteration nodes and pruned, retaining only the iteration nodes of local optimal solutions; at the same time, principal component analysis is performed on the multidimensional parameters collected by the high-precision edge perception module to compress the data, select the core feature dimensions, and remove redundant feature dimensions. The neural network adopts a feedforward neural network structure, with the number of neurons in its input layer being consistent with the number of core feature dimensions obtained after principal component analysis, and the number of neurons in its output layer being 1. The neural network is used to make short-term predictions of the output trend of individual distributed energy sources on a time scale of 5-15 minutes, and to obtain the predicted output value. The predicted output value of the neural network is used as the initial input parameter of the lightweight swarm intelligence algorithm; and after each iteration of the lightweight swarm intelligence algorithm completes a local optimum, the algorithm corrects the iteration parameters by adjusting the deviation between the predicted output value and the actual collected value; the deviation is calculated as follows: ； Where: ε t P represents the relative deviation of the output at time t. pred,t The neural network predicts the output force; P act,t For actual data collection output; P rated The rated output of the equipment is used; the specific correction rule is as follows: when the deviation is >10%, the algorithm iteration step size is increased linearly according to the deviation ratio; when the deviation is ≤10%, the algorithm learning rate is fine-tuned to 0.01-0.

05.

6. The cloud-edge-device-edge four-layer distributed energy collaborative optimization method according to claim 5, characterized in that: The autonomous decision-making process of the distributed energy zone aggregation agent on the edge side is based on a multi-objective optimization model. The constraints of the multi-objective optimization model include upper and lower limits of distributed energy output, grid voltage / power constraints, and energy storage SOC constraints. The objective functions include four objectives: maximizing the absorption of new energy, minimizing carbon emissions, optimizing operational economy, and maximizing energy utilization. Each objective function is configured with dynamic weight coefficients calculated based on the analytic hierarchy process (AHP). These dynamic weight coefficients are dynamically adjusted according to the proportion of new energy output in the distributed energy zone, real-time electricity market trading prices, and carbon quota trading prices. The adjustment formula for the dynamic weighting coefficient is as follows: ； in, The dynamic weighting coefficients of the single objective function are as follows: x corresponds to the four objectives mentioned above; α is the weighting coefficient corresponding to the proportion of new energy output, corresponding to the objective of maximizing new energy consumption; β is the weighting coefficient corresponding to the electricity market transaction price, corresponding to the objective of optimal operating economy; γ is the weighting coefficient corresponding to the carbon quota transaction price, corresponding to the objective of minimizing carbon emissions; δ is the weighting coefficient corresponding to the comprehensive energy utilization coefficient, corresponding to the objective of maximizing energy utilization rate; and satisfies α+β+γ+δ=1; Pn' is the linear normalized value of the proportion of regional new energy output, Pe' is the linear normalized value of the real-time electricity market transaction price, Pc' is the linear normalized value of the carbon quota transaction price, and Pu' is the linear normalized value of the regional energy utilization rate; the values ​​of Pn', Pe', Pc', and Pu' are all in the range of 0-1; the linear normalized value is calculated as follows: linear normalized value = (actual value of the parameter - minimum value of the parameter) / (maximum value of the parameter - minimum value of the parameter); the maximum and minimum values ​​of each parameter are taken as the historical statistical extreme values ​​of the past 30 natural days.

7. The cloud-edge-device-edge four-layer distributed energy collaborative optimization method according to claim 6, characterized in that: The analytic hierarchy process (AHP) index system includes four primary indicators: new energy output volatility, market electricity price fluctuation rate, carbon quota trading return rate, and comprehensive energy utilization rate. The weight values ​​of the four primary indicators are calculated based on the AHP, and these weight values ​​are directly converted into specific values ​​of α, β, γ, and δ. The objective function of the multi-objective optimization model is solved iteratively by a multi-objective collaborative swarm intelligence algorithm mounted on the edge agent. The algorithm iteration terminates when the objective function converges to a preset accuracy range with the total active power deviation of the distributed energy zone as the calculation dimension, or when the number of algorithm iterations reaches a preset maximum number of iterations. The total active power deviation is calculated as follows: Total active power deviation = |Total actual active power output of the zone - Total planned active power output of the zone| / Total scheduled active power output of the zone. The preset maximum number of iterations is set according to the scale of the distributed energy zone; the more distributed energy units there are, the larger the number of iterations.

8. The cloud-edge-device-edge four-layer distributed energy collaborative optimization method according to claim 7, characterized in that: The multi-objective collaborative swarm intelligence algorithm carried by the distributed energy zone aggregation intelligent agent on the end side is an operator layer fusion algorithm based on non-dominated sorting genetic algorithm and ant colony algorithm, and integrates a convolutional neural network feature extraction model for data feature extraction. The feature extraction model is used to extract features from the interaction data of multiple side agents within the region collected by the edge agent; the core feature dimensions extracted include output deviation, adjustment response time, and resource utilization, and each feature is a quantitative feature. Adjustment response time = Decision command issuance time - Equipment response completion time; Resource utilization rate = Actual energy utilization / Total available energy × 100%; The quantized feature dimension results extracted by the feature extraction model are used as the initial input parameters of the fusion algorithm; The fusion of the non-dominated sorting genetic algorithm and the ant colony algorithm at the operator layer specifically involves fusing the selection and crossover operators of the non-dominated sorting genetic algorithm with the path search and pheromone update operators of the ant colony algorithm. The fusion rule is as follows: based on the three core dimensions of quantified features extracted by the feature extraction model, an initial population is generated through the selection and crossover operators. The path search operator performs local optimization based on the initial population, and the pheromone update operator updates the pheromone according to the local optimization results. The population iteration and path optimization are achieved through iterative loops, and the optimal solution, i.e., the multi-objective optimization and control instruction within the region, is output as the decision instruction for the edge agent.

9. The cloud-edge-device-edge four-layer distributed energy collaborative optimization method according to claim 8, characterized in that: The edge-side distributed energy cluster aggregation intelligent agent is equipped with a load sensing module and a computing power allocation module; the load sensing module provides real-time quantitative data input to the computing power allocation module, the computing power allocation module generates computing power allocation scheduling data, and the edge-side intelligent agent dynamically adjusts and optimizes the computing strategy through the computing power allocation scheduling data; the load sensing module and the computing power allocation module are the core functional modules for the edge-side intelligent agent to realize secondary optimization and control within the cluster. The load sensing module is used to collect two types of quantitative data in real time: the overall operating load parameters of the distributed energy group and the quantitative indicators of computing power demand for optimization tasks of each end-side intelligent agent within the group. The overall operating load parameters include the total active power, total reactive power, and group load rate, where the group load rate = total active power / rated total active power of the group × 100%. The quantitative indicators of computing power demand include the amount of task data, the number of algorithm iterations, and the computing power occupancy rate. The computing power allocation module dynamically allocates computing power to each edge agent within the group based on a two-factor weighting algorithm of load ratio and task complexity; the formula for calculating the computing power allocation ratio is as follows: Where η is the computing power allocation ratio of a single end-side intelligent agent, and its value ranges from 0 to 1; The proportion of the distributed energy zone load corresponding to this edge agent to the total group load is calculated as follows: =Total active power of the district / Total active power of the group × 100%; When the total active power of the group is 0 The number of endpoint agents within the group; The proportion of the optimization task complexity of this edge agent to the total task complexity of the group; This is the load proportion weighting coefficient. Let be the task complexity weight coefficient, and satisfy . The The value is dynamically adjusted based on the power grid operating conditions; the power grid operating conditions are quantitatively defined according to the group-level load rate and adjusted accordingly for different operating conditions. The value of makes the allocation of computing power more in line with the needs of power grid operation; The task complexity is quantified using a weighted summation method. The calculation formula is: Task complexity = 0.4 × Task data volume ratio + 0.3 × Algorithm iteration count ratio + 0.3 × Computing power utilization ratio. Each ratio is the ratio of the indicator corresponding to a single edge agent to the sum of the indicators corresponding to all edge agents in the group, and all are normalized, with a value range of 0-1. The computing power allocation module also performs a secondary dynamic adjustment of the computing power allocation ratio of each end-side agent in the group based on the grid constraint verification results of the edge-side agents: for end-side agents whose optimization results have over-constraint items, the corresponding ratio is increased based on the original computing power allocation ratio; for end-side agents whose optimization results have no over-constraint items, the corresponding ratio is decreased based on the original computing power allocation ratio; after adjustment, the sum of the computing power allocation ratios of all end-side agents in the group is still 1; the over-constraint items are items whose edge-side optimization results exceed the cloud-side global grid security constraint quantification threshold, and the global grid security constraint quantification threshold is clearly defined by the cloud-side global demand data.

10. The cloud-edge-device-edge four-layer distributed energy collaborative optimization method according to claim 9, characterized in that: The swarm intelligence algorithm carried by the distributed energy swarm aggregation intelligent agent on the side is a grid constraint-adaptive enhancement optimization algorithm based on ant colony algorithm fusion feature extraction model and deep reinforcement learning model. The input parameters of the algorithm are the group-level operating load parameters collected by the load sensing module and the computing power allocation ratio results of the computing power allocation module. The number of algorithm iterations is dynamically and adaptively adjusted according to the computing power allocation ratio. The higher the computing power allocation ratio, the larger the number of iterations. The number of iterations on the edge side is matched with the number of iterations on the end side within the group. The feature extraction model is constructed using a random forest model, which is used to extract features from inter-group interaction data collected by edge agents and global power grid constraint data sent from the cloud; the core of the extraction is 1214. The quantitative characteristics include inter-group output deviation, grid constraint margin, and load forecast deviation; The deep reinforcement learning model uses the core quantitative features extracted by the feature extraction model as the model input and maximizes the inter-group collaborative decision matching degree as the model reward function. After offline training and online optimization, it outputs the inter-group collaborative decision weight coefficient, i.e., the group decision factor. The offline training is based on the historical data of inter-group interactions and power grid constraints over the past 30 natural days, and the online optimization dynamically adjusts the model parameters based on real-time operation data. The inter-group collaborative decision matching degree = inter-group decision instruction fit degree × inter-group resource complementarity degree, with a value range of 0-1. The group decision factor is the priority weight of inter-group collaborative decision. The group decision factor is directly integrated into the pheromone update rule of the ant colony algorithm; the pheromone update formula of the ant colony algorithm is: τ ij (t+1)=(1-ρ)×τ ij (t)+ρ×Q ij ×λ; where τ ij (t+1) represents the pheromone concentration of path ij at time t+1, τ ij (t) represents the initial pheromone concentration at time t along path ij, ρ is the pheromone evaporation coefficient, and Q... ij Q is the fitness value of path ij, which is linearly positively correlated with the inter-group collaborative decision-making matching degree. The higher the matching degree, the better Q. ij The larger the value, the more λ becomes the group decision factor, used to adjust the priority of collaborative decision-making among groups.

Images