Park source network load storage cooperative scheduling method and system based on knowledge graph

Abstract

The invention discloses a park source network load storage cooperative scheduling method and system based on a knowledge graph, and the method comprises the steps: constructing the knowledge graph of a field based on a scheduling rule, and coding the knowledge graph into low-dimensional vector representation; fusing the real-time system data and the priori knowledge vector to generate enhanced state representation; constructing a reinforcement learning decision-making environment; a distributed Actor-Critic framework is adopted, and all agents share environment information but independently optimize a strategy; and realizing collaborative decision-making under global constraints through iterative updating of a strategy gradient algorithm. The system comprises a knowledge graph unit, a vector fusion unit, an environment definition unit and an optimization unit. According to the invention, a cooperative scheduling scheme satisfying the industrial park source-network-load-storage integrated scheduling scene can be provided. The method can be widely applied to the technical field of power system dispatching.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching technology, and in particular to a knowledge graph-based method and system for coordinated dispatching of power generation, grid, load and storage in a power plant. Background Technology

[0002] The structure of modern power systems is changing, with increasing integration and interaction between the generation side, grid, load side, and energy storage side. On the one hand, the integration of renewable energy sources such as photovoltaics and wind power, as well as various distributed power sources, has led to uncertainties and fluctuations in power output characteristics. On the other hand, demand-side resources such as electric vehicles and adjustable industrial loads provide new flexibility for system dispatch. This trend of "generation, grid, load, and storage integration" is particularly evident in scenarios such as industrial parks, which typically integrate self-built photovoltaic and combined heat and power units, energy storage equipment, and diverse production loads, placing higher demands on the safe, economical, and stable operation of electricity.

[0003] To achieve coordinated dispatching of power generation, grid, load, and storage in complex scenarios such as industrial parks, existing methods primarily employ centralized optimization approaches. These methods, traditional power system dispatching techniques, establish a unified mathematical programming model for the entire dispatching system, with a central controller collecting global information for centralized solution to obtain the theoretically optimal dispatching scheme. The effectiveness of this method heavily relies on accurate modeling and prediction of system topology, equipment parameters, and changes in power generation and load. In real-world environments, obtaining such precise information is costly and difficult to achieve, leading to insufficient robustness of the dispatching scheme. Summary of the Invention

[0004] In view of this, in order to solve the technical problem that existing collaborative scheduling methods cannot accurately model data, thus leading to insufficient robustness of the scheduling scheme, the present invention proposes a knowledge graph-based collaborative scheduling method for source-grid-load-storage systems in industrial parks. The method includes the following steps:

[0005] A domain knowledge graph is constructed based on scheduling rules and encoded into a low-dimensional vector representation.

[0006] By integrating real-time system data with prior knowledge vectors, an enhanced state representation with semantic enhancement is generated;

[0007] Construction of Reinforcement Learning Decision Environment: State Space: Define the system observation space based on enhanced state representation; Action Space: Set the executable decision variables and their feasible range; Reward Mechanism: Design a multi-scale reward function that takes into account both optimization objectives and constraints; Constraint Modeling: Ensure policy compliance through penalty terms or feasible region restrictions;

[0008] A distributed Actor-Critic framework is adopted, in which each agent shares environmental information but optimizes its strategy independently; collaborative decision-making under global constraints is achieved through iterative updates of the policy gradient algorithm.

[0009] In some embodiments, it also includes:

[0010] System security constraints are defined based on backup capacity constraints, network security constraints, and device health and safety constraints, and actions are corrected based on these system security constraints.

[0011] This invention also proposes a knowledge graph-based collaborative scheduling system for source-grid-load-storage in a park, which includes:

[0012] The knowledge graph unit constructs a domain-specific knowledge graph based on scheduling rules and encodes it into a low-dimensional vector representation.

[0013] The vector fusion unit fuses real-time system data with prior knowledge vectors to generate enhanced state representations with semantic enhancement.

[0014] The environment definition unit is used to construct the environment for performing reinforcement learning decision-making.

[0015] The optimization unit adopts a distributed Actor-Critic framework, where each agent shares environmental information but independently optimizes its strategy; it achieves collaborative decision-making under global constraints through iterative updates via a policy gradient algorithm.

[0016] Based on the above scheme, this invention provides a knowledge graph-based method and system for coordinated scheduling of source, network, load, and storage in a park. By constructing an expert knowledge graph and embedding it into the model, the decision-making process of the intelligent agent is based on existing rules and experience, thus making the decision-making logic traceable and shortening the training time required for the model to achieve effective performance. A safety correction module is set up to perform mandatory verification and correction on the action after the intelligent agent outputs the initial action and before it is sent to the environment for execution, ensuring that the final executed scheduling instruction is strictly within the preset safety boundary, thus overcoming the security risks of traditional methods. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps of a knowledge graph-based collaborative scheduling method for source-grid-load-storage in a park according to the present invention.

[0018] Figure 2 This is a structural block diagram of a knowledge graph-based collaborative scheduling system for source-grid-load-storage in a park, according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0021] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0022] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0023] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0024] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0025] Reference Figure 1 This is a flowchart illustrating an optional example of the knowledge graph-based collaborative scheduling method for source-grid-load-storage systems in a park proposed in this invention. This method can be applied to computer equipment, and the scheduling method proposed in this embodiment may include, but is not limited to, the following steps:

[0026] Step S1, Offline Preparation Stage: Collect the scheduling rules of the industrial park, construct a knowledge graph and generate knowledge vectors;

[0027] Step S2, Online Stage: Based on a gated neural network, knowledge vectors are fused with real-time data vectors to generate an enhanced state representation;

[0028] Step S3, Online Phase: Construct the state space based on the enhanced state representation, and define the action space, reward function, and constraints;

[0029] Step S4, Online Phase: Combining the state space, action space, reward function, and constraints, and with the operational goals of economy, stability, and low carbon emissions, a multi-agent Actor-Critic framework is used for decision optimization.

[0030] In some feasible embodiments, step S1 specifically includes:

[0031] This step is the offline preparation phase of the system. Its core objective is to transform the unstructured expert knowledge in the field of power system source-grid-load-storage scheduling into a structured mathematical form that can be understood and utilized by machine learning models.

[0032] S1.1 Knowledge Acquisition and Formalization: Systematically collect the scheduling rules of industrial parks, including: electricity pricing rules; production rules; economic scheduling strategies; and physical laws. Convert these rules into RDF triples.

[0033] S1.2 Knowledge Vectorization: The TransE model is used to train the knowledge graph described above. By minimizing the loss function L, a "low-dimensional knowledge vector" v for each entity and relation can be obtained. k This vector mathematically contains semantic information about the industrial park's scheduling knowledge system. The trained vector will be stored in a knowledge vector database.

[0034]

[0035] in, S is the set of positive samples (correct triples), and S' is the set of negative samples (generated by replacing the head entity, relation, or tail entity). This step yields the knowledge vector v. k,i .

[0036] In some feasible embodiments, step S2 specifically includes:

[0037] At each scheduling time t, this step generates an enhanced state representation h′ for each agent i that integrates real-time data and prior knowledge. i,t .

[0038] S2.1 Construction of real-time data vector: For each key node i in the power grid, obtain its normalized real-time measurement data and construct it as a data vector.

[0039] v d,i,t =[U i,t ,θ i,t ,P G,i,t Q G,i,t ,P L,i,t Q L,i,t SOC E,i,t ,...]

[0040] Among them, U i,t ,θ i,t These are the node voltage magnitude and phase angle, respectively; P G,i,t Q G,i,t It represents the active and reactive power generated by this node; P L,i,t Q L,i,t It refers to the active and reactive power of the load; SOC E,i,t It is in a state of energy storage and charging.

[0041] S2.2, Gating Fusion:

[0042] Based on the identifier of node i, retrieve its corresponding knowledge vector v from the knowledge vector base in step S1. k,i The fusion weights g are calculated using a gated neural network unit. i,t ∈(0,1):

[0043]

[0044] Where σ is the Sigmoid activation function; W g and b g These are the learnable weights and biases of the gating network; This indicates a vector concatenation operation.

[0045] Using gating weights g i,t Dynamically fuse data vectors and knowledge vectors to generate a fused feature vector:

[0046] v f,i,t =g i,t ·(W d v d,i,t )+(1-g i,t )·(W k,i )

[0047] Among them W d and W k Learnable linear transformation matrices are used to map data vectors and knowledge vectors to the same fused feature space dimension.

[0048] S2.3, Graph Attention Information Aggregation:

[0049] Fuse the features of all nodes {v f,1,t ,v f,2,t ,...,v f,N,t This serves as input to a Graph Attention Network (GAT). GAT calculates the attention coefficient e of neighboring node j to target node i. ij To aggregate information:

[0050]

[0051] Where a T and W v These are the learnable parameters of the GAT layer. The attention coefficients are normalized using the Softmax function to obtain the final attention weights α. ij,t :

[0052]

[0053] Where N i It is the set of neighboring nodes of node i.

[0054] Finally, the output of the GAT layer, i.e., the enhanced state representation h′, is obtained by weighted summation. i,t :

[0055]

[0056] In this embodiment, the gated fusion graph attention network designed in this invention can adaptively fuse real-time data with prior knowledge. This mechanism ensures that when the data is reliable, the decision-making leans towards data-driven optimization; when the data is abnormal, it relies more on stable expert rules, thereby improving the system's adaptability and the reliability of decisions under different operating conditions.

[0057] In some feasible embodiments, the state space in step S3 is: the state s of each agent i. i,t From its local real-time observation o i,t (v d,i,t (a subset of) and the enhanced state representation h generated in step two i,t Together they constitute:

[0058] In some feasible embodiments, the action space in step S3 is: the action a of agent i. i,t This includes devices that it can directly control. Specifically, it includes:

[0059] Control actions of self-built power source: Output P of controllable unit (cogeneration) CHP,t That is, the active power output of the cogeneration unit at time t; the start-stop state of the controllable unit u.CHP,t , is a binary variable that determines whether the unit is turned on (1) or off (0) at time t.

[0060] Control actions of the energy storage system (storage): charging power P ESS,ch,t That is, the charging power of the energy storage system at time t; the discharging power P ESS,dis,t , which is the discharge power of the energy storage system at time t.

[0061] Interaction with the main power grid: Purchased power P grid,buy,t That is, the power purchased from the external power grid at time t; the power sold, P. grid,sell,t That is, the power sold to the external power grid at time t.

[0062] Load-side (load) management actions: Interruptible load P IL That is, the amount of load interrupted at time t; the load reduction amount P that can be shifted. shift,cut,t That is, the amount of load reduced from the peak period at time t to prepare for relocation; the amount of load that can be relocated, P. shift,add,t That is, the load that is moved in at time t, which originally belonged to other time periods.

[0063] For a single agent i, its action space is a subset of the aforementioned actions, specifically depending on the unit that agent is responsible for controlling in the "source-grid-load-storage" system. The actions of all agents collectively constitute the system's joint action A = (a 1,t ,a 2,t ,...,a N,t ).

[0064] In some feasible embodiments, the reward function in step S3 specifically includes:

[0065] Reward function R t To guide the system towards economical, stable, and low-carbon operation, the negative of the system's objective function is used as the global reward signal. The objective function is defined as follows:

[0066] minC total =C grid +C dg +C ess +C dr

[0067] Where: C grid Represents the grid interaction cost; C dg C is the operating cost of a self-built power source; ess Indicates the operating cost of the energy storage system; C dr This refers to load-side management costs.

[0068] The final reward function is defined as:

[0069] R=C total +r push

[0070] Where, r push This is a penalty term; if the variable does not meet the constraints, this term will be negative.

[0071] Grid interaction cost C grid The calculation formula includes the time-of-use electricity purchase cost, the demand electricity fee, and minus the revenue from electricity sales, as follows:

[0072]

[0073] c buy (t) represents the electricity price at time t; P buy (t) represents the power purchased from the grid at time t; c sell The grid connection price at time (t)t; P sell (t) represents the power sold to the grid at time t; C demand Δt represents the maximum demand electricity cost within the scheduling cycle; Δt represents the duration of each time step.

[0074] Self-built power supply operating cost C dg The cost, including renewable energy and controllable units, is calculated using the following formula:

[0075]

[0076] C PV,om C WT,om P represents the unit power generation and operation and maintenance cost of photovoltaic and wind power; PV (t),P WT (t) represents the output of photovoltaic and wind power at time t; C CHP,fuel C CHP,om This represents the fuel and operation and maintenance costs of a combined heat and power (CHP) unit; P CHP (t) represents the output of the cogeneration unit at time t; C CHP,su The single start-up cost of a combined heat and power (CHP) unit; u CHP (t) represents a 0-1 variable, which is 1 when the unit starts at time t, and 0 otherwise; C env E represents the environmental cost per unit of carbon emissions. CHP (t) represents the carbon emissions of the cogeneration unit at time t.

[0077] Energy storage system operating cost C ess The main quantification is its cycle life loss, calculated using the following formula:

[0078]

[0079] c degP represents the equivalent depreciation cost per kilowatt-hour of electricity charged or discharged from the energy storage system; ch (t) represents the charging power of the stored energy at time t; P dis (t) represents the discharge power of the stored energy at time t.

[0080] Load-side management cost C dr This includes compensation and excitation for adjustable loads, calculated using the following formula:

[0081]

[0082] c il P represents the compensation cost per unit of load interruption; il (t) represents the load power interrupted at time t; c sh P represents the incentive cost per unit load transferred; sh,down (t) represents the load power that is reduced (waiting to be shifted) from the peak at time t.

[0083] In some feasible embodiments, the constraints in step S3 specifically include:

[0084] Power balance constraint: Ensure that the power supply and demand in the park remain balanced at any time t.

[0085] P PV (t)+P WT (t)+P CHP (t)+P dis (t)+P buy (t)=P load (t)+P ch (t)+P sell (t)

[0086] Among them, P PV (t) represents the actual output power of the photovoltaic at time t; P WT (t) represents the actual output power of the wind power at time t; P CHP (t) represents the output power of the cogeneration unit at time t; P dis (t) Discharge power of the energy storage system at time t; P buy (t) represents the power purchased from the grid at time t; P load (t) represents the total electrical load in the park at time t; P ch (t) represents the charging power of the energy storage system at time t; P sell (t) represents the power sold to the grid at time t.

[0087] Grid interaction constraints limit power exchange between the industrial park and the main power grid:

[0088] Power purchase constraint: 0≤P buy(t)≤u buy (t)P grid,max , where P grid,max Indicates the maximum transmission power of the connection line between the park and the power grid; u buy (t) is a binary variable. It is 1 when electricity is purchased from the grid at time t, and 0 otherwise.

[0089] Electricity sales power constraint: 0≤P sell (t)≤u sell (t)P grid,max , where u sell (t) A binary variable. It is 1 when electricity is sold to the grid at time t, and 0 otherwise.

[0090] Mutual exclusion constraint for power purchase and sale: u buy (t)+u sell (t)≤1, this formula ensures that the park cannot purchase and sell electricity at the same time at any given time.

[0091] Constraints of self-built power supply operation:

[0092] Renewable energy output constraints: The output of solar and wind power is affected by weather and cannot exceed their predicted maximum available power. The formula is as follows:

[0093] 0≤P PV (t)≤P PV,forecast (t)

[0094] 0≤P W T(t)≤P WT,forecast (t)

[0095] Among them, P PV,forecast (t) represents the predicted maximum output power of the photovoltaic system at time t; P WT,forecast This represents the predicted maximum output power of wind power at time t.

[0096] Controllable unit operating constraints: Combined heat and power (CHP) units have minimum / maximum output limits and ramp-up rate limits. The formulas are as follows:

[0097] u CHP (t)P CHP,min ≤P CHP (t)≤u CHP (t)P CHP,max

[0098] -R down ≤P CHP (t)-P CHP (t-1)≤R up

[0099] Where P CHP,min P CHP,maxThese represent the minimum and maximum output of the CHP unit, respectively; u CHP (t) is a binary variable. It is 1 when the CHP unit is running at time t, and 0 otherwise; R down R up These represent the maximum upward and downward ramp rates of the CHP unit, respectively.

[0100] Energy storage system operating constraints describe the physical characteristics of the energy storage system:

[0101] The dynamic update formula for the State of Charge (SOC) of an energy storage system is as follows:

[0102]

[0103] SOC(t) represents the state of charge of the stored energy at time t; E ESS,rated Indicates the rated capacity of the energy storage system; η ch η dis This indicates the charging and discharging efficiency of the energy storage system.

[0104] SOC boundary constraints: prevent overcharging and over-discharging, the formula is as follows.

[0105] SOC min ≤SOC(t)≤SOC max

[0106] SOC min SOC max These represent the minimum and maximum permissible values ​​for the state of charge of the energy storage, in order to protect battery life.

[0107] Charging power limit: 0≤P ch (t)≤u ch (t)P ch,max Charging power must not exceed the maximum charging power.

[0108] Discharge power limit: 0≤P dis (t)≤u dis (t)P dis,max The discharge power must not exceed the maximum discharge power.

[0109] Charge-discharge mutual exclusion constraint: u ch (t)+u dis (t)≤1, meaning that charging and discharging cannot occur simultaneously at any given time.

[0110] Load demand-side management constraints:

[0111] Total load composition: P load (t)=P load,base (t)-P il (t)-Psh,down (t)+P sh,up (t)

[0112] P load,base (t) represents the original predicted load at time t; P il (t) represents the interruptible load that is interrupted at time t; P sh,down (t) represents the amount of movable load that is reduced at time t; P sh,up (t) represents the amount of movable load that is moved in at time t.

[0113] Interruptible load constraints: The calculation formula is as follows. The first part limits the maximum interruptible load at time t to prevent system scheduling from exceeding the actual interruptible capacity; the second part limits the total interruption duration to prevent excessive impact or loss to users.

[0114] 0≤P il (t)≤P il,max (t)

[0115]

[0116] P il,max (t) represents the maximum interruptible load at time t; u il (t) is a binary variable representing whether a load interrupt was executed at time t; T il,max This indicates the longest possible interruption duration for interruptible loads within a single scheduling cycle.

[0117] Sliding load constraint: Ensures that the total amount of electricity cut during the entire scheduling cycle equals the total amount of electricity moved in, and that at any given time, the maximum amount of electricity that can be moved out or moved in does not exceed P. sh,max The load is (t). The calculation formula is as follows;

[0118]

[0119] 0≤P sh,up (t)≤P sh,max (t)

[0120] 0≤P sh,up (t)≤P sh,max (t)

[0121] In some feasible embodiments, it also includes:

[0122] SBF Safety Correction: Corrects actions based on system safety constraints.

[0123] Linearize or quadratize the system's core safety constraint C(x,a)≤0. Here, x represents the system state (e.g., voltage, phase angle), and a represents the joint action to be executed.

[0124] Backup capacity constraint: This is the most critical constraint for ensuring system security. It requires the system to reserve sufficient, rapidly deployable backup capacity at all times to cope with sudden load increases or power outages. Backup capacity is typically divided into upward and downward backup.

[0125] Increased reserve constraints: The system needs to be able to increase power generation or decrease power consumption in a short period of time to cope with sudden load increases or generator failures. The formula is as follows:

[0126] R req,up (t)-(R CHP,up (t)+R ESS,up (t)+R IL (t)+R Grid,up (t))≤0

[0127] R req,up (t)=max(P CHP,max ,P PV,forecast (t))

[0128] R CHP,up (t)=P CHP,max -P CHP (t)

[0129]

[0130] R IL (t)=P il,max (t)-P il (t)

[0131] R Grid,up (t)=P grid,max -P buy (t)

[0132] Among them, R req,up (t) represents the total increased reserve capacity required at time t; R CHP,up (t) represents the upward reserve that the CHP unit can provide, i.e., the difference between its rated maximum power and its current power; R ESS,up (t) represents the backup power provided by energy storage, which is limited by both its maximum discharge power and remaining power; R IL (t) represents the upward reserve that interruptible loads can provide; R Grid,up (t) The maximum up-regulation capacity provided by the external power grid at time t, which is the difference between the capacity of the external power grid and the current purchased power.

[0133] Lowering reserve constraints: The system needs to be able to reduce power generation or increase power consumption in a short period of time to cope with sudden load drops or unexpected increases in photovoltaic output. The calculation formula is as follows:

[0134] R req,down(t)-(R CHP,down (t)+R ESS,down (t)+R Grid,down (t))≤0

[0135] R CHP,down (t)=P CHP (t)-P CHP,min

[0136]

[0137] R Grid,down (t)=P grid,max -P sell (t)

[0138] Among them, R req,down (t) represents the total downsizing reserve capacity required at time t; R CHP,down (t) represents the downsizing reserve that the CHP unit can provide; R ESS,down (t) Energy storage can provide downsizing backup; R Grid,down (t) represents the down-limit reserve that the external power grid can provide.

[0139] Network security constraints ensure that the power flow within the park's power grid does not exceed limits, preventing damage to lines or transformers due to localized overloads and subsequent cascading failures. The formula is as follows:

[0140] |F l (t)∣-F l,max ≤0

[0141]

[0142] Among them, F l (t) represents the power flow on line l at time t; F l,max This represents the maximum permissible transmission power of line l. N bus Indicates the number of nodes in the system; PTDF l,n P represents the power transfer distribution factor, which is the contribution of unit power injected at node n to the power flow of line l. This is a pre-calculated constant. net,n (t) represents the net injected power of node n at time t, which is the sum of the power of all generating units at that node minus the sum of the power of all loads.

[0143] Equipment health and safety constraints: In addition to meeting real-time operation requirements, the health status of the equipment itself must also be considered to avoid severely shortening the equipment's lifespan due to overly aggressive scheduling strategies. Specifically, this includes:

[0144] Energy storage system SOC depth limit: To extend battery life and prevent it from being in a state of overcharging or over-discharging for extended periods. SOC min,safeand SOC max,safe It is a more conservative safety threshold than the conventional operating limits.

[0145] Controllable unit start-stop limit: Frequent start-stop cycles will severely shorten the lifespan of the generator set. CHP (t)-u CHP The expression (t-1) is only 1 when the unit changes from stopped (0) to started (1), capturing the start-up action. su,max This represents the maximum number of startups allowed within a scheduling period T.

[0146] Finally, given the initial joint action a output by the Actor network. raw,t The SBF module uses QP (Quadratic Programming) to solve for the core safety constraint C(x,a)≤0, obtaining the minimum corrected safety action: a safe,t .

[0147] In some feasible embodiments, a training step is also included:

[0148] This invention employs an Actor-Critic model of Centralized Training with Decentralized Execution (CTDE). Each agent relies only on its local observations and augmented state representations during execution. However, during the training phase, the experience data of all agents are aggregated to update the global Critic network, thereby guiding the optimization of each Actor network. The specific training process is as follows:

[0149] Initialization: Initialize the Actor network for each agent i in the system. and its corresponding target Actor network Initialize the Critic network Q(·|θ) shared by all agents. Q Initialize the learnable parameters θ of the gated fusion graph attention network. G Initialize an experience replay pool R with a capacity of M.

[0150] Environmental Interaction and Experience Collection: At each time step t in each training round: for each agent i, its local observation data is acquired and combined with knowledge vectors to generate an enhanced state representation through a gated fusion graph attention network. Together they constitute its state s i,t Each agent's Actor network is based on its state s i,t Output initial action The initial joint action of all agents a t =(a1,t ,...,a N,t The input is fed into the SBF safety correction module; the SBF module solves the optimization problem and outputs the corrected final action a′ that guarantees absolute safety. t Perform safety actions in the environment a′ t The global reward signal r was observed. t and the system's next state s′ t ; to complete the empirical tuple (s t ,a t ,r t ,s′ t ,a′ t Store it in the experience replay pool R.

[0151] Network model update: When the amount of data in the experience replay pool R exceeds a certain threshold, every few time steps, a batch of experience data is randomly sampled from R for network update.

[0152] Update the Critic network: The optimization objective of the Critic network is to minimize the temporal difference error. Its loss function is L(θ). Q ) is defined as:

[0153] L(θ Q ) = E (s,a,r,s′)～R [(yQ(s,|aθ Q )) 2 ]

[0154] The target value y is calculated by the target network to increase training stability.

[0155] y=r+γQ′(s′,a′|θ Q′ )

[0156] a′ is the action generated by the target Actor network π′ in the next state s′. By minimizing this loss function, the parameters θ of the Critic network are... Q And the gated fusion graph network parameters θ, which are part of its input. G Perform gradient updates.

[0157] Update the Actor networks for each agent: The update goal for each Actor network is to generate actions that will earn higher evaluation values ​​from the Critic network. Its policy gradient... Approximately:

[0158]

[0159] Update the target network: Perform soft updates on the parameters of all target networks to slowly track the parameters of the main network, using the following method:

[0160] θ′←τθ+(1-τ)θ′

[0161] Where τ is a very small update coefficient, θ represents the main network parameters, and θ′ represents the target network parameters.

[0162] like Figure 2 As shown, a knowledge graph-based collaborative scheduling system for source-grid-load-storage in a park includes:

[0163] Knowledge graph unit, used to execute step S1;

[0164] A vector fusion unit is used to perform step S2;

[0165] An environment definition unit is used to execute step S3;

[0166] The optimization unit is used to execute step S3.

[0167] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0168] A knowledge graph-based collaborative scheduling device for source-grid-load-storage systems in a park:

[0169] At least one processor;

[0170] At least one memory for storing at least one program;

[0171] When the at least one program is executed by the at least one processor, the at least one processor implements a knowledge graph-based collaborative scheduling method for source-grid-load-storage in a campus, as described above.

[0172] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0173] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement a knowledge graph-based collaborative scheduling method for source-grid-load-storage systems in a campus, as described above.

[0174] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0175] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A knowledge graph-based collaborative scheduling method for source-grid-load-storage systems in a park, characterized in that, Includes the following steps: A knowledge graph is constructed based on the scheduling rules, and knowledge vectors are generated. The knowledge vector is fused with the real-time data vector to generate an enhanced state representation; A state space is constructed based on the enhanced state representation, and an action space, reward function, and constraints are defined. By combining the state space, the action space, the reward function, and the constraints, a multi-agent Actor-Critic framework is used for decision optimization.

2. The knowledge graph-based collaborative scheduling method for source-grid-load-storage in a park according to claim 1, characterized in that, Also includes: Actions are corrected based on system safety constraints.

3. The knowledge graph-based collaborative scheduling method for source-grid-load-storage in a park according to claim 2, characterized in that, The step of constructing a knowledge graph and generating knowledge vectors based on scheduling rules specifically includes: Obtain the scheduling rules of the industrial park and convert the scheduling rules into triples to obtain a knowledge graph; The knowledge graph is vectorized to obtain knowledge vectors.

4. The knowledge graph-based collaborative scheduling method for source-grid-load-storage in a park according to claim 2, characterized in that, The step of fusing the knowledge vector with the real-time data vector to generate an enhanced state representation specifically includes: Acquire real-time measurement data and vectorize it to obtain real-time data vectors; The fusion weight is calculated based on the node identifier, and the real-time data vector and the knowledge vector are fused to obtain the fusion feature vector; The fused feature vectors of all nodes are used as input to the graph attention network to aggregate information and generate enhanced state representations.

5. The knowledge graph-based collaborative scheduling method for source-grid-load-storage in a park according to claim 4, characterized in that, The constraints include: Power balance constraints: P PV (t)+P WT (t)+P CHP (t)+P dis (t)+P buy (t)=P load (t)+P ch (t)+P sell (t); Among them, P PV (t) represents the actual output power of the photovoltaic at time t; P WT (t) represents the actual output power of the wind power at time t; P CHP (t) represents the output power of the cogeneration unit at time t; P dis (t) represents the discharge power of the energy storage system at time t; P buy (t) represents the power purchased from the grid at time t; P load (t) represents the total electrical load in the park at time t; P ch (t) represents the charging power of the energy storage system at time t; P sell (t) represents the power sold to the grid at time t; Power grid interaction constraints: 0≤P buy (t)≤u buy (t)P grid,max ; 0≤P sell (t)≤u sell (t)P grid,max ; u buy (t)+u sell (t)≤1; Among them, P grid,max Indicates the maximum transmission power of the connection line between the park and the power grid; u buy (t) represents the electricity purchased from the power grid; u sell (t) represents the electricity sold to the grid; Constraints of self-built power supply operation: Among them, P PV,forecast (t) represents the predicted maximum output power of the photovoltaic system at time t; P WT,forecast P represents the predicted maximum output power of wind power at time t; CHP,min P CHP,max These represent the minimum and maximum output of the CHP unit, respectively; u CHP (t) represents the operating status of the CHP unit at time t; R down R up These represent the maximum upward and downward ramp rates of the CHP unit, respectively. Energy storage system operating constraints: SOC min ≤SOC(t)≤SOC max ; 0≤P ch (t)≤u ch (t)P ch,max ; 0≤P dis (t)≤u dis (t)P dis,max ; u ch (t)+u dis (t)≤1; Where SOC(t) represents the state of charge of the stored energy at time t; E ESS,rated Indicates the rated capacity of the energy storage system; η ch η dis This indicates the charging and discharging efficiency of the energy storage system; Load demand-side management constraints: P load (t)=P load,base (t)-P il (t)-P sh,down (t)+P sh,up (t); 0≤P il (t)≤P il,max (t) 0≤P sh,up (t)≤P sh,max (t); 0≤P sh,up (t)≤P sh,max (t); Among them, P load,base (t) represents the original predicted load at time t; P il (t) represents the interruptible load that is interrupted at time t; P sh,down (t) represents the amount of movable load that is reduced at time t; P sh,up (t) represents the amount of movable load that is moved in at time t; P il,max (t) represents the maximum interruptible load at time t; u il (t) represents the execution status of the load interruption at time t; T il,max This indicates the longest possible interruption duration for interruptible loads within a single scheduling cycle.

6. The knowledge graph-based collaborative scheduling method for source-grid-load-storage in a park according to claim 5, characterized in that, The system security constraints include backup capacity constraints, network security constraints, and equipment security and health constraints.

7. A knowledge graph-based collaborative scheduling system for source-grid-load-storage in a park, characterized in that, include: The knowledge graph unit constructs a knowledge graph according to scheduling rules and generates knowledge vectors; The vector fusion unit fuses the knowledge vector with the real-time data vector and generates an enhanced state representation; The environment definition unit constructs a state space based on the enhanced state representation, and defines the action space, reward function, and constraints. The optimization unit, combining the state space, the action space, the reward function, and the constraints, employs a multi-agent Actor-Critic framework for decision optimization.

8. A knowledge graph-based collaborative scheduling device for source-grid-load-storage systems in a park, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the knowledge graph-based collaborative scheduling method for source-grid-load-storage in a park as described in any one of claims 1-6.

Images