Power distribution network load transfer method and system based on data knowledge driving

Abstract

The invention belongs to the technical field of power distribution networks, and particularly relates to a power distribution network load transfer method and system based on data knowledge driving. The method comprises the following steps: constructing a mathematical model of a power distribution network load transfer problem; constructing a multi-agent space according to the mathematical model; combining the multi-agent space with the mathematical model to obtain a multi-agent initial model; obtaining a teaching experience library, and training the multi-agent initial model in combination with a learning algorithm to obtain a pre-trained multi-agent model; and generating a load transfer strategy by using the pre-trained multi-agent model. The load transfer decision can be generated according to the observation information through the multi-agent system, generation of the load transfer decision is assisted, and safe operation of the power system is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution network, and in particular, relates to a method and system for load transfer in a distribution network driven by data knowledge. Background Art

[0002] As the uncertainty of power systems becomes increasingly complex, the types of faults in distribution networks are becoming more and more numerous. Their wiring branches and equipment are very complex, and the transition resistance is greatly affected by the environment during a fault. It is crucial to accurately locate the fault location and transfer the load in the shortest possible time.

[0003] Under the background of rapid construction and development of new power systems, a large number of new energy sources and emerging loads are connected. The power system under the background of high proportion of renewable energy has undergone profound changes in many aspects such as stability mechanism, dispatching operation and planning development. The power generation principle, control method, grid-connected equipment and external characteristics of non-hydro renewable energy units represented by wind power and photovoltaics are quite different from those of conventional units. After the access of distributed power sources and new loads with strong randomness and volatility, the source-load uncertainty characteristics will bring more severe challenges to the safe and reliable operation of the power system, and the risks of photovoltaic grid disconnection and load interruption will increase sharply. In solving the problem of load transfer, current technology usually adopts traditional mathematical programming solution methods, but this method has a slow solution rate and low efficiency in generating load transfer strategies, which cannot meet the real-time requirements of load transfer tasks. In recent years, with the rapid development of artificial intelligence technology, data-driven methods have provided a new way for the research on distribution network fault handling and intelligent regulation. Summary of the invention

[0004] In view of the above problems, the present disclosure provides a distribution network load transfer method based on data knowledge drive, the method comprising:

[0005] Construct a mathematical model for the load transfer problem in distribution networks;

[0006] Construct a multi-agent space based on mathematical models;

[0007] Combine the multi-agent space with the mathematical model to obtain the multi-agent initial model;

[0008] Obtain the teaching experience library, combine it with the learning algorithm, train the multi-agent initial model, and obtain a pre-trained multi-agent model;

[0009] Generate load transfer strategies using pre-trained multi-agent models.

[0010] Furthermore, a mathematical model of the load transfer problem in the distribution network is constructed, including:

[0011] Obtain the weight coefficient, the restored load capacity within the entire restoration process time, and the constraints for safe operation of the distribution network;

[0012] Determine the objective function based on the weight coefficient and the load capacity to be restored within the entire restoration process time;

[0013] According to the objective function and the constraints of safe operation of distribution network, a mathematical model of load transfer problem in distribution network is constructed.

[0014] Furthermore, the constraints for safe operation of the distribution network include: power flow constraints, radial topology constraints, node voltage constraints, branch current constraints and node output constraints of the three-phase balanced distribution network.

[0015] Furthermore, the multi-agent space includes: a state observation space, a global state space, an action space and an immediate reward function.

[0016] Furthermore, the acquisition of the instant reward function includes:

[0017] Determine the sum of the load powers restored at each time step and the penalty function;

[0018] The sum of the load power restored at each time step and the penalty function are subtracted to obtain the immediate reward function.

[0019] Furthermore, a teaching experience library is obtained, including:

[0020] Obtain actual operation data of the power grid in the target area;

[0021] Integrate the actual operation data of the power grid in the target area into a priori data set;

[0022] According to the principle of maximum action information entropy, the recovery trajectory of partial load transfer is selected from the prior data set to build a teaching experience library.

[0023] Furthermore, the multi-agent initial model is trained to obtain a pre-trained multi-agent model, including:

[0024] Obtain fusion agent interaction experience and obtain training data from the teaching experience library;

[0025] Set up the algorithm framework, combine and integrate the agent interaction experience and training data, train the multi-agent initial model, and obtain a pre-trained multi-agent model.

[0026] Furthermore, before obtaining the fusion agent interaction experience, it also includes:

[0027] Build the replay buffer framework.

[0028] The present disclosure also proposes a distribution network load transfer system based on data knowledge drive, the system comprising:

[0029] The first building module is used to build a mathematical model for the load transfer problem of the distribution network;

[0030] The second building module is used to construct a multi-agent space based on a mathematical model;

[0031] A combination module is used to combine the multi-agent space with the mathematical model to obtain the multi-agent initial model;

[0032] The training module is used to obtain the teaching experience library and train the multi-agent initial model in combination with the learning algorithm to obtain a pre-trained multi-agent model;

[0033] A generation module is used to generate load transfer strategies using a pre-trained multi-agent model.

[0034] Furthermore, the first building block is used to build a mathematical model for the load transfer problem of the distribution network, including:

[0035] The first building block is used to obtain the weight coefficient, the restored load capacity within the entire restoration process time, and the constraints for the safe operation of the distribution network;

[0036] Determine the objective function based on the weight coefficient and the load capacity to be restored within the entire restoration process time;

[0037] According to the objective function and the constraints of safe operation of distribution network, a mathematical model of load transfer problem in distribution network is constructed.

[0038] Furthermore, the constraints for safe operation of the distribution network include: power flow constraints, radial topology constraints, node voltage constraints, branch current constraints and node output constraints of the three-phase balanced distribution network.

[0039] Furthermore, the multi-agent space includes: a state observation space, a global state space, an action space and an immediate reward function.

[0040] Furthermore, the acquisition of the instant reward function includes:

[0041] Determine the sum of the load powers restored at each time step and the penalty function;

[0042] The sum of the load power restored at each time step and the penalty function are subtracted to obtain the immediate reward function.

[0043] Furthermore, the training module is used to obtain the teaching experience library, including:

[0044] The training module is used to obtain the actual operation data of the power grid in the target area;

[0045] Integrate the actual operation data of the power grid in the target area into a priori data set;

[0046] According to the principle of maximum action information entropy, the recovery trajectory of partial load transfer is selected from the prior data set to build a teaching experience library.

[0047] Furthermore, a training module is used to train the multi-agent initial model to obtain a pre-trained multi-agent model, including:

[0048] The training module is used to obtain the interaction experience of the fusion agent and obtain training data from the teaching experience library;

[0049] Set up the algorithm framework, combine and integrate the agent interaction experience and training data, train the multi-agent initial model, and obtain a pre-trained multi-agent model.

[0050] Furthermore, before the training module is used to obtain the fusion agent interaction experience, it also includes:

[0051] Build the replay buffer framework.

[0052] The present disclosure has the following beneficial effects:

[0053] (1) This paper establishes a data-driven mathematical model and a multi-agent reinforcement learning framework based on the Qmix algorithm, and proposes a distribution network load transfer method based on data knowledge drive. Compared with the traditional mathematical programming solution method, it has a faster solution rate in solving the load transfer problem and can efficiently generate load transfer strategies, thereby meeting the real-time requirements of the load transfer task.

[0054] (2) This disclosure introduces historical expert operation process experience and uses prior knowledge to guide the training of the multi-agent system model, which greatly accelerates the model training process and reduces the model training cost. This makes the load transfer strategy generated by the model more in line with the expert operation preferences, improves the reliability of the load transfer strategy generated by the model, and provides effective decision-making support for the corresponding power business scenarios.

[0055] Other features and advantages of the present disclosure will be described in the following description, and partly become apparent from the description, or be understood by implementing the present disclosure. The purpose and other advantages of the present disclosure can be realized and obtained by the structures indicated in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 A diagram showing a method for load transfer in a distribution network driven by data knowledge in an embodiment of the present disclosure is shown;

[0058] Figure 2 A diagram of a load transfer system of a distribution network driven by data knowledge in an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0059] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as being limited to the examples set forth herein; on the contrary, these embodiments are provided so that the present disclosure will be more comprehensive and complete, and the concepts of the example embodiments are fully conveyed to those skilled in the art. The described features, structures, or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced while omitting one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known technical solutions are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0060] In addition, the accompanying drawings are only schematic illustrations of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or implemented in one or more hardware units or integrated circuits, or implemented in different networks and / or processor devices and / or microcontroller devices.

[0061] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the steps. For example, some steps may be decomposed, while some steps may be combined or partially combined, so the actual execution order may change according to the actual situation.

[0062] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein, for example.

[0063] In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or sub-modules is not necessarily limited to those steps or sub-modules explicitly listed, but may include other steps or sub-modules not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0064] like Figure 1 As shown, the present disclosure proposes a distribution network load transfer method based on data knowledge driving, the method comprising:

[0065] Construct a mathematical model for the load transfer problem in distribution networks;

[0066] Construct a multi-agent space based on mathematical models;

[0067] Combine the multi-agent space with the mathematical model to obtain the multi-agent initial model;

[0068] Obtain the teaching experience library, combine it with the learning algorithm, train the multi-agent initial model, and obtain a pre-trained multi-agent model;

[0069] Generate load transfer strategies using pre-trained multi-agent models.

[0070] Specifically, the detailed technical solution of this method is as follows:

[0071] Step 1: Establish a mathematical model for the decision-making problem of load transfer in the distribution network; consider the safe operation constraints in the distribution network, such as the operation restrictions of the node lines, to construct the constraint items. Consider the possible load recovery impact in the load transfer process and the need to drive small-capacity loads first during the black start process. In order to ensure that these loads are restored first and that the loads are restored in a focused and orderly manner in the subsequent process, a weight coefficient is set for each individual load. By setting a reasonable weight coefficient, the product of the load capacity restored within the entire recovery process time and the weight coefficient is used as the objective function to guide the load transfer process. After the objective function is determined, the mathematical model of the load transfer problem in the distribution network is further constructed.

[0072] Step 1.1: Consider the safe operation constraints in the distribution network, such as the operation restrictions of node lines, and construct the constraint items under the mathematical model of the load transfer task based on the relevant constraints:

[0073]

[0074] Formulas (1) to (4) are the power flow constraints of the three-phase balanced distribution network. i,t and q i,t is the active power and reactive power of node i at time t; v i,t and θi,t is the voltage amplitude and phase of node i at time t; v j,t is the voltage amplitude of node j at time t; θ ij,t The voltage phase difference between node i and node j at time t; and is the reactive power demand and active power demand of the load connected to node i at time t; is the load recovery status of node i at time t; and are respectively the active power and reactive power generated by the generator on node i at time t; is the actual active output of PV connection node i at time t; G ij and B ij are the conductance and susceptance of branch ij; D is the collection of distribution network lines.

[0075] Strictly speaking, formula (1) and formula (2) are not used to calculate parameters, but to construct the equation constraint conditions of the load transfer mathematical model. Formula (1) establishes the equation constraint that the net active power consumption of the node is equal to the current node voltage and the adjacent node voltage and the line admittance and phase difference relationship. Formula (2) establishes the equation constraint that the net active power consumption of the node is equal to the active power consumption of the node minus the active power generated by the node. The specific constraints established by the two formulas are different and have different meanings. Formula (3) and formula (4) are similar, but the object is reactive power.

[0076] Step 1.2: Construct radial topology constraints:

[0077]

[0078] In the formula, is the line switch connection state between lines ij at time t; n s,t is the number of microgrids in the distribution network at time t; n b is the total number of nodes in the distribution network.

[0079] Step 1.3: Construct node voltage and branch current constraints:

[0080] V min ≤v i,t ≤V max (6)

[0081] I ij,min ≤i ij,t ≤I ij,max (7)

[0082] In the formula, v i,t is the voltage amplitude of node i at time t; i ij,t is the current of branch ij at time t.

[0083] Step 1.4: Construct node output constraints:

[0084]

[0085] In the formula, and is the actual active output and maximum allowable active output of PV connection node i at time t; The maximum generation capacity limit issued for node i.

[0086] Step 1.5: Considering the load priority, the product of the load capacity restored within the entire restoration process time and the weight coefficient is used as the objective function. Specifically, the total length of the restoration steps of the distribution network load transfer strategy is set to T, and the objective function in the scenario where the system node set is V is:

[0087]

[0088] In the formula, c i is the weight coefficient of node i load; is the load recovery state of node i at time t. If it is restored, otherwise, load is the active power demand of the load connected to node i at time t.

[0089] Step 2: Based on the established mathematical model, configure an agent for each movable line connection switch, further construct the agent set of multi-agent reinforcement learning, the environment state space, the agent observation space and the agent reward mechanism, and transform the distribution network fault recovery problem into a Markov decision process.

[0090] Step 2.1: Configure an agent for each movable line connection switch, which can obtain information in the corresponding state space and output action decisions within the action space.

[0091] Step 2.2: Construct the state observation space of each agent, including node voltage, line current, fault line ID, switch state, node load recovery state, etc., and further construct the state observation space of the entire multi-agent system. For example, the state observation information of a line connection switch agent is the system operation parameters that have a strong correlation with this line connection switch, such as the voltage of nearby nodes and the recovery state of node loads, the current of nearby lines and switch states, the ID of the fault line, etc. These local state observation information constitute the state observation space of a certain agent.

[0092] Step 2.3: Construct the global state space of the multi-agent system, which includes global observation information such as the voltage of each node, the current of each line branch, the ID of the fault line, the closed state of the switch, the recovery state of the target load, etc. Unlike the local state observation space of a single agent, the global state space of the multi-agent system can fully reflect the global information, including the voltage and load recovery state of all nodes in the environment, the current of each line branch and the closed state of the switch, the ID of the fault line, and other global information.

[0093] Step 2.4: Construct the action space of the agent. The agent can determine the closed state of the switch based on observations, so its action space is a discrete quantity of 0 or 1.

[0094] Step 2.5: The task goal of the entire multi-agent system is to restore power to the loads on the network nodes in an orderly manner according to the importance level after the failure occurs as much as possible. The relationship between the agents is completely cooperative, so the immediate reward function r t It is set to the sum of the load power restored at each time step under the premise of considering the load priority, and a penalty mechanism is added to make the power grid operate beyond the constraint limit.

[0095]

[0096] In the formula, the immediate reward r t is the sum of the restored loads at time t considering the load priority; V is the set of system nodes; c i is the weight factor; is the recovery state of the connected node i at time t. If it is recovered, otherwise, load is the active power demand of the load connected to node i at time t; φ is the penalty function of the system exceeding the limit. When the system operation status exceeds the limit, the penalty function φ is obtained by multiplying each constraint by the penalty coefficient and performing a weighted summation.

[0097] Step 3: Based on the actual operation data of the regional power grid, random fault scenarios covering various fault types and fault locations are generated, and the load transfer decision process under the corresponding faults is generated. Furthermore, according to the principle of maximum action information entropy, a certain number of load transfer decision processes are selected to generate a teaching experience library.

[0098] Step 3.1: Based on the actual operation data of the regional power grid, generate random fault scenarios covering various fault types and fault locations and generate load transfer decision process data under corresponding faults, which are further integrated into a priori data sets.

[0099] Step 3.2: From the prior data set, select the recovery trajectory of partial load transfer according to the principle of maximum action information entropy to build a teaching experience library. The form of mathematical expression is as follows:

[0100]

[0101] Λ * =arg max H(a|Λ) (13)

[0102] Where H(a|Λ) is the action information entropy; a is the action trajectory; Λ is the teaching dataset constructed by sampling a certain amount of trajectories from the prior dataset; p j is the frequency of the jth action in the action set appearing in all action trajectories of the teaching dataset Λ.

[0103] Step 4: Build a multi-agent reinforcement learning framework, use the multi-agent reinforcement learning algorithm Qmix algorithm to train multi-agents, and accelerate the training of the model by introducing the teaching experience library and buffer reuse structure. After the training is completed, the agent is put into use.

[0104] Step 4.1: Build a replay buffer framework; use the replay buffer of historical tasks and current task experience to mix the interaction trajectory of historical tasks into the interaction trajectory of the current task, and accelerate the learning of the current task by reusing historical experience.

[0105] Step 4.2: Construct an experience database that integrates the interaction experience of intelligent agents and a certain proportion of teaching experience to provide training data for the multi-agent system during the multi-agent training process.

[0106] Step 4.3: Based on the Markov decision process established in the previous step, set up the Qmix algorithm framework to implement reinforcement learning training for multiple agents.

[0107] The loss function L(θ) of the Qmix algorithm is:

[0108]

[0109] In the formula, is the current reward function generated based on the target network; Q tot (τ, a, s; θ) is the reward function generated by the current training network; b is the data capacity of the set training set.

[0110]

[0111] In the formula, r is the immediate reward of the agent, The reward function for the next stage generated by the target network, γ is the reward decay coefficient.

[0112] The gradient descent of L(θ) is used to realize the training of multiple agents. At the same time, the experience database constructed earlier that integrates the agent interaction experience and teaching experience is used to accelerate the training convergence process of the agent.

[0113] Step 4.4: After training, the multi-agent system can generate load transfer decisions based on observation information and assist in the generation of load transfer strategies. At this time, the trained multi-agent system can be put into use.

[0114] The specific steps of step 4 are shown in the flow chart of Table 1.

[0115] Table 1

[0116]

[0117] like Figure 2 As shown, the present disclosure also proposes a distribution network load transfer system driven by data knowledge, the system comprising:

[0118] The first building module is used to build a mathematical model for the load transfer problem of the distribution network;

[0119] The second building module is used to construct a multi-agent space based on a mathematical model;

[0120] A combination module is used to combine the multi-agent space with the mathematical model to obtain the multi-agent initial model;

[0121] The training module is used to obtain the teaching experience library and train the multi-agent initial model in combination with the learning algorithm to obtain a pre-trained multi-agent model;

[0122] A generation module is used to generate load transfer strategies using a pre-trained multi-agent model.

[0123] Those skilled in the art should understand that although the present disclosure is described in detail with reference to the aforementioned embodiments, it is still possible to modify the technical solutions recorded in the aforementioned embodiments, or to make equivalent replacements for some of the technical features therein; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for load transfer in a distribution network driven by data knowledge, characterized in that: The method comprises: Construct a mathematical model for the load transfer problem in distribution networks; Construct a multi-agent space based on mathematical models; Combine the multi-agent space with the mathematical model to obtain the multi-agent initial model; Obtain the teaching experience library, combine it with the learning algorithm, train the multi-agent initial model, and obtain a pre-trained multi-agent model; Generate load transfer strategies using pre-trained multi-agent models.

2. The method for load transfer in distribution network based on data knowledge drive according to claim 1 is characterized in that: Construct a mathematical model for the load transfer problem in the distribution network, including: Obtain the weight coefficient, the restored load capacity within the entire restoration process time, and the constraints for safe operation of the distribution network; Determine the objective function based on the weight coefficient and the load capacity to be restored within the entire restoration process time; According to the objective function and the constraints of safe operation of distribution network, a mathematical model of load transfer problem in distribution network is constructed.

3. The method for load transfer in distribution network based on data knowledge drive according to claim 2 is characterized in that: The constraints for the safe operation of the distribution network include: power flow constraints, radial topology constraints, node voltage constraints, branch current constraints and node output constraints of the three-phase balanced distribution network.

4. The method for load transfer in distribution network based on data knowledge drive according to claim 1, characterized in that: The multi-agent space includes: a state observation space, a global state space, an action space and an immediate reward function.

5. The method for load transfer in distribution network based on data knowledge drive according to claim 4 is characterized in that: The acquisition of the instant reward function includes: Determine the sum of the load powers restored at each time step and the penalty function; The sum of the load power restored at each time step and the penalty function are subtracted to obtain the immediate reward function.

6. The method for load transfer in distribution network based on data knowledge drive according to claim 1, characterized in that: Get a library of teaching experience, including: Obtain actual operation data of the power grid in the target area; Integrate the actual operation data of the power grid in the target area into a priori data set; According to the principle of maximum action information entropy, the recovery trajectory of partial load transfer is selected from the prior data set to build a teaching experience library.

7. The method for load transfer in distribution network based on data knowledge drive according to claim 1, characterized in that: Train the multi-agent initial model to obtain a pre-trained multi-agent model, including: Obtain fusion agent interaction experience and obtain training data from the teaching experience library; Set up the algorithm framework, combine and integrate the agent interaction experience and training data, train the multi-agent initial model, and obtain a pre-trained multi-agent model.

8. The method for load transfer in distribution network based on data knowledge drive according to claim 7 is characterized in that: Before obtaining the fusion agent interaction experience, it also includes: Build the replay buffer framework.

9. A distribution network load transfer system driven by data knowledge, characterized in that: The system comprises: The first building module is used to build a mathematical model for the load transfer problem of the distribution network; The second building module is used to construct a multi-agent space based on a mathematical model; A combination module is used to combine the multi-agent space with the mathematical model to obtain the multi-agent initial model; The training module is used to obtain the teaching experience library and train the multi-agent initial model in combination with the learning algorithm to obtain a pre-trained multi-agent model; A generation module is used to generate load transfer strategies using a pre-trained multi-agent model.

10. The data knowledge driven distribution network load transfer system according to claim 9, characterized in that: The first building block is used to construct a mathematical model for the load transfer problem of the distribution network, including: The first building block is used to obtain the weight coefficient, the restored load capacity within the entire restoration process time, and the constraints for the safe operation of the distribution network; Determine the objective function based on the weight coefficient and the load capacity to be restored within the entire restoration process time; According to the objective function and the constraints of safe operation of distribution network, a mathematical model of load transfer problem in distribution network is constructed.

11. The data knowledge driven distribution network load transfer system according to claim 10, characterized in that: The constraints for the safe operation of the distribution network include: power flow constraints, radial topology constraints, node voltage constraints, branch current constraints and node output constraints of the three-phase balanced distribution network.

12. The data knowledge driven distribution network load transfer system according to claim 9, characterized in that: The multi-agent space includes: a state observation space, a global state space, an action space and an immediate reward function.

13. The data knowledge driven distribution network load transfer system according to claim 12, characterized in that: The acquisition of the instant reward function includes: Determine the sum of the load powers restored at each time step and the penalty function; The sum of the load power restored at each time step and the penalty function are subtracted to obtain the immediate reward function.

14. The data knowledge driven distribution network load transfer system according to claim 12, characterized in that: Training module, used to obtain teaching experience library, including: The training module is used to obtain the actual operation data of the power grid in the target area; Integrate the actual operation data of the power grid in the target area into a priori data set; According to the principle of maximum action information entropy, the recovery trajectory of partial load transfer is selected from the prior data set to build a teaching experience library.

15. The data knowledge driven distribution network load transfer system according to claim 12, characterized in that: The training module is used to train the multi-agent initial model to obtain a pre-trained multi-agent model, including: The training module is used to obtain the interaction experience of the fusion agent and obtain training data from the teaching experience library; Set up the algorithm framework, combine and integrate the agent interaction experience and training data, train the multi-agent initial model, and obtain a pre-trained multi-agent model.

16. The data knowledge driven distribution network load transfer system according to claim 15, characterized in that: Before the training module is used to obtain the interaction experience of the fusion agent, it also includes: Build the replay buffer framework.

Images