Power system dynamic economic dispatch system and method based on generative adversarial imitation learning

Abstract

The present invention discloses a power system dynamic economic dispatch system and method based on generative adversarial imitation learning. First, the generator network module observes the power system state and uses a reinforcement learning proximal strategy optimization algorithm to generate a dispatch strategy; secondly, the perfect dispatch module generates a perfect dispatch strategy; thirdly, the discriminator network module compares the dispatch strategy generated by the generator network module with the perfect dispatch strategy, and obtains feedback information to train the generator network parameters and the discriminator network parameters; finally, the generator network module obtains the final dispatch strategy based on the generator network parameters trained by the discriminator network module and the power system state. The present invention combines generative adversarial networks to avoid the subjectivity introduced by artificially defining reward functions in deep reinforcement learning, realizes end-to-end learning from strategy to strategy, improves the convergence problem of the algorithm, reduces the difficulty of modeling, and enhances the ability of the algorithm to deal with high-dimensional complex problems.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and specifically to a system and method for dynamic economic dispatching of power systems based on generative adversarial imitation learning. Background Art

[0002] Building a new power system with a high proportion of renewable energy is the only way to achieve the "dual carbon" goal. However, renewable energy power generation is intermittent, volatile and random, and the uncertainty introduced by large-scale renewable energy grid connection is bound to bring challenges to the dispatching and operation of the power system.

[0003] Dynamic economic dispatch arranges the output plan of thermal power units based on new energy and load forecast data to minimize the system power generation cost within the dispatch cycle. Traditional methods model dynamic economic dispatch as an optimization problem with uncertainty, which can be solved using stochastic constraint programming and robust optimization methods. However, stochastic constraint programming requires artificial representation of the probability distribution model of random variables, and its optimization results depend on the degree of consistency between the assumed probability distribution and the actual random variables. The robust optimization algorithm sacrifices economy to meet the safety of the power system in the worst scenario, making the dispatch plan too conservative. Reinforcement learning has the characteristics of adapting to environmental uncertainty, seeking long-term control target optimization, and high real-time decision-making efficiency. Reinforcement learning is a necessary means to improve the effectiveness and flexibility of dynamic economic dispatch of power systems. However, reinforcement learning faces too large action exploration space due to the lack of prior knowledge guidance in the early stage of intelligent agent training, resulting in too long strategy convergence time.

[0004] Inspired by the idea of ​​imitation learning, dynamic economic dispatch is regarded as an imitation learning problem, and the dispatch center is allowed to imitate the perfect dispatch strategy to arrange the unit output plan. Dynamic economic dispatch based on generative adversarial imitation learning (GAIL) is proposed, which allows the intelligent agent to obtain expert prior knowledge, imitate the perfect dispatch strategy, and optimize its own strategy. Under the premise that the goals of imitation learning and perfect dispatch are consistent, the discriminator is used to avoid artificially defining the reward function, and the interaction between the generator and the power system environment is used to enable the intelligent agent to learn a strategy that is better than expert experience. Summary of the invention

[0005] The purpose of the present invention is to provide a system and method for dynamic economic dispatch of power systems based on generative adversarial imitation learning, construct a generator network, and generate a dispatch strategy by observing the system state. Inspired by the objective function of the proximal strategy optimization algorithm of reinforcement learning, the loss function of the generator network is innovatively constructed, and the network parameters are updated in reverse to optimize the dispatch strategy. At the same time, the ideal dispatch plan is calculated and the learning of the generator network is guided. Finally, a discriminator network is constructed to identify the generation strategy and the perfect dispatch strategy, output the identification result, and assist the generator network in updating.

[0006] To achieve this purpose, the present invention designs a power system dynamic economic dispatch system based on generative adversarial imitation learning, including a generator network module, a perfect dispatch module and a discriminator network module; the generator network module is used to observe the state of the power system, generate a dispatch strategy through the generator network, and adopt a reinforcement learning proximal strategy optimization algorithm to obtain the determined output of the unit; the perfect dispatch module is used to generate a perfect dispatch strategy; the discriminator network module uses a discriminator network to compare the dispatch strategy generated by the generator network module with the perfect dispatch strategy generated by the perfect dispatch module, and obtain feedback information for training the generator network parameters of the generator network module and the discriminator network parameters of the discriminator network module.

[0007] A method for dynamic economic dispatch of power system based on generative adversarial imitation learning, comprising the following steps:

[0008] Step 1: The generator network module observes the power system state, generates a dispatch strategy through the generator network, and uses the reinforcement learning proximal strategy optimization algorithm to obtain the determined output of the unit;

[0009] Step 2, the perfect scheduling module generates a perfect scheduling strategy;

[0010] Step 3: The discriminator network module uses the discriminator network to compare the scheduling strategy generated by the generator network module with the perfect scheduling strategy generated by the perfect scheduling module to obtain feedback information for training the generator network parameters of the generator network module and the discriminator network parameters of the discriminator network module;

[0011] Step 4: The generator network module obtains the final dispatch strategy based on the trained generator network parameters obtained in step 3 and the power system status.

[0012] The beneficial effects of the present invention are:

[0013] 1. In order to improve the shortcomings of a single reinforcement learning algorithm that has no prior knowledge to guide and requires artificially defined reward functions, we deepened the research on the dynamic economic dispatch model of the power system based on deep reinforcement learning, combined with the perfect dispatch theory, introduced the idea of ​​imitation learning, and proposed a dynamic economic dispatch model of the power system based on generative adversarial imitation learning. Under the guidance of the perfect dispatch strategy, the exploration of meaningless action space can be reduced and the learning process of the intelligent agent can be accelerated. Combined with the generative adversarial network, the subjectivity introduced by artificially defined reward functions in deep reinforcement learning is avoided, and end-to-end learning from strategy to strategy is realized, which can significantly improve the convergence problem of the algorithm.

[0014] 2. The dynamic economic scheduling model based on generative adversarial imitation learning improves the convergence and stability issues of deep reinforcement learning, and the algorithm effect has been improved to a certain extent. Under the guidance of the perfect scheduling plan, the model reduces meaningless action exploration and shortens the convergence time. At the same time, it eliminates the subjectivity of artificially defined reward functions, reduces the difficulty of modeling, reduces the workload of parameter adjustment, and enhances the ability of the algorithm to deal with high-dimensional complex problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a system structure diagram of the present invention;

[0016] Figure 2 is a flow chart of the method of the present invention;

[0017] Figure 3 This is a schematic diagram of the generator network structure;

[0018] Figure 4 Schematic diagram of the discriminator network structure;

[0019] Among them, 1-generator network module, 2-perfect scheduling module and 3-discriminator network module. DETAILED DESCRIPTION

[0020] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0021] A dynamic economic dispatch system for power systems based on generative adversarial imitation learning, such as Figure 1 As shown, it includes a generator network module 1, a perfect scheduling module 2 and a discriminator network module 3; the generator network module 1 is used to observe the state of the power system, generate a scheduling strategy through the generator network, and adopt a reinforcement learning proximal strategy optimization algorithm to obtain the determined output of the unit; the perfect scheduling module 2 is used to generate a perfect scheduling strategy; the discriminator network module 3 uses a discriminator network to compare the scheduling strategy generated by the generator network module 1 with the perfect scheduling strategy generated by the perfect scheduling module 2, and obtain feedback information for training the generator network parameters of the generator network module 1 and the discriminator network parameters of the discriminator network module 3.

[0022] A power system dynamic economic dispatch system based on generative adversarial imitation learning, further comprising a generator output optimization module 4 and a network model offline training module 5;

[0023] In the above technical scheme, the generator output optimization module 4 generates an optimized scheduling control strategy by randomly sampling the normal distribution of the unit output to avoid the learning of the generator from falling into the local optimal point; the network model offline training module 5 adopts the generator network module 1 to generate a scheduling plan according to the offline training scenario, and executes the scheduling plan in the power system environment, and the power system environment is transferred to a new state; after a decision cycle, the state and action of each decision step are collected into the generator memory to generate a generator memory state-action sequence; for the same scheduling scenario, the perfect scheduling plan is obtained according to the perfect scheduling module 2, the system state and the perfect scheduling plan are collected into the expert memory to generate an expert memory state-action sequence; the generator memory state-action sequence and the expert memory state-action sequence are input into the discriminator network module 3, and the discriminator network of the discriminator network module 3 outputs a discrimination result, and the generator network in the generator network module 1 updates the generator network parameters according to the discrimination result to generate a better scheduling plan.

[0024] In the above technical solution, the generator network in the generator network module 1 includes an 8-layer network structure, such as Figure 3 As shown, they are, in order, the generator network input layer, the generator network normalization layer, the generator network first hidden layer, the generator network first activation function layer, the generator network second hidden layer, the generator network second activation function layer, the generator network output layer, and the generator network third activation function layer;

[0025] The input data of the generator network input layer is the power system observation state s t ; The normalization layer is used to offset and scale different input features in the input data of the input layer; the output data of the output layer of the generator network is n groups of output layer neurons, and the output layer neurons contain mean and standard deviation. The n groups of units are determined through the n groups of output layer neurons.

[0026] In the above technical solution, the power system observation state s t It includes the load forecast value at time t+1, the new energy output forecast value at time t+1, the actual output of the thermal power unit at time t and the actual output of the new energy at time t, where the new energy includes wind power and photovoltaic power;

[0027] The input variable s t It is expressed as follows:

[0028] s t ={t,P′ load,t+1 ,P′ wind,t+1 ,P′ pv,t+1 ,P g,t ,P wind,t ,Ppv,t}

[0029] Among them, P′ load,t+1 represents the load forecast value at time t+1, P′ wind,t+1 represents the predicted wind power output at time t+1, P′ pv,t+1 represents the predicted photovoltaic output value at time t+1, P g,t represents the actual output of the thermal power unit at time t; P wind,t represents the actual wind power output value at time t, P pv,t is the actual photovoltaic output value at time t;

[0030] The output variable of the generator network output layer is the output value a of the thermal power unit. tg :

[0031] a tg = {P G1 ,…,P Gk ,…,P GK}

[0032] Among them, P Gk is the output value of the kth thermal power unit; P GK is the output value of the Kth thermal power unit; K is the total number of thermal power units in the system.

[0033] In the above technical solution, the specific implementation method of the generator output optimization module 4 to generate the optimized scheduling control strategy is:

[0034] The output of the thermal power unit P output by the generator network module 1 Gk The mean value of Thermal power unit output P Gk The standard deviation of The mean and standard deviation form a probability distribution of the output of thermal power units such as a normal distribution, and its probability density function is as follows:

[0035]

[0036] Standard Deviation in Probability Distributions It reflects the exploration ability of the action. A larger standard deviation in the early stage of training can increase the exploration space of the action. As the training process gradually stabilizes, the standard deviation will stabilize at a smaller value in the later stage of training.

[0037] The mean and standard deviation of the thermal power unit output output by the generator network module 1 must conform to the probability density function. If not, the generator network module 1 will iteratively train to obtain a new thermal power unit output until the mean and standard deviation of the new thermal power unit output satisfy the normal distribution.

[0038] In the above technical solution, the specific method of the perfect scheduling module 2 to generate the perfect scheduling strategy is:

[0039] After the real-time market ends every day, the perfect scheduling module 2 uses the known information at each moment of the day, given the system operating status and operating conditions, with the goal of minimizing the system power generation cost, and uses the optimal unit combination and optimal power flow SCOPF that take into account the system safety operation constraints to calculate the perfect economic scheduling plan for 24 moments of the day to generate a perfect scheduling strategy; the known information includes actual power load, actual new energy output and actual topology structure.

[0040] In the above technical solution, the discriminator network in the discriminator network module 3 includes a 7-layer network structure, such as Figure 4 As shown, they are, in order, the discriminator network input layer, the discriminator network normalization layer, the discriminator network first hidden layer, the discriminator network first activation function layer, the discriminator network second hidden layer, the discriminator network second activation function layer, and the discriminator network output layer;

[0041] The input data of the discriminator network input layer includes state-action pairs (s t ,a tg ) and state-action pairs (s t ,a te ), the state-action pair (s t ,a tg ) is the output of the generator network output layer a tg and the power system observation state s t The state-action pair (s t ,a te ) is the scheduling strategy a generated by the perfect scheduling module 2 te and the power system observation state s t The output data of the output layer of the discriminator network is the probability value of the perfect scheduling solution.

[0042] The discriminator network adds the discriminator network normalization layer after the discriminator network input layer to normalize the state-action pair; adds the discriminator network second activation function layer before the discriminator network output layer to map the output to the interval [0,1] as the corresponding probability value that the input strategy is the perfect scheduling strategy;

[0043] The discriminator network D ω The objective function is:

[0044]

[0045] Among them, (s t ,a tg) is the output of the generator network output layer a tg and the power system observation state s t The state-action pair (s t ,a te ) is the scheduling strategy a generated by the perfect scheduling module 2 te and the power system observation state s t The state-action pair composed of G The generator strategy formed for the generator network; τ E The perfect scheduling strategy generated by the perfect scheduling module 2.

[0046] In the above technical solution, the specific implementation method of the reinforcement learning proximal strategy optimization algorithm in the generator network module 1 is:

[0047] The reinforcement learning proximal policy optimization algorithm defines the generator G θ The objective function is trained to generate a scheduling strategy. The expression of the objective function is:

[0048]

[0049] clip(r t (θ),1-ε,1+ε)D ω (s t ,a tg ))]

[0050]

[0051] Among them, (s t ,a tg ) is the output of the generator network output layer a tg and the power system observation state s t The state-action pair composed of G is the generator strategy formed by the generator network, θ is the learning rate of the generator network, w is the learning rate of the discriminator network, D ω (s t ,a tg ) is the output data of the discriminator output layer, ε is a hyperparameter, generally set to 0.2, π θ (a tg |s t ) is to follow the current strategy in state s t When selecting action a tg The probability of To follow the old strategy in state s t When selecting action a tg The probability of t (θ) is the value in state st When following the current strategy and the old strategy, action a is selected tg The ratio of the probability of ω (s t ,a tg ) reflects the gap between the current scheduling strategy and the perfect scheduling solution. The strategy can be updated under the guidance of the expert strategy to avoid the subjectivity of artificially setting the reward function and achieve end-to-end learning from strategy to strategy. Clip is the clipping function.

[0052] A method for dynamic economic dispatch of power systems based on generative adversarial imitation learning, such as Figure 2 As shown, the following steps are included:

[0053] Step 1, the generator network module 1 observes the power system state, generates a dispatch strategy through the generator network, and uses the reinforcement learning proximal strategy optimization algorithm to obtain the determined output of the unit;

[0054] Step 2, the perfect scheduling module 2 generates a perfect scheduling strategy;

[0055] Step 3, the discriminator network module 3 uses the discriminator network to compare the scheduling strategy generated by the generator network module 1 with the perfect scheduling strategy generated by the perfect scheduling module 2, and obtains feedback information for training the generator network parameters of the generator network module 1 and the discriminator network parameters of the discriminator network module 3;

[0056] Step 4: The generator network module 1 obtains the final dispatching strategy based on the trained generator network parameters obtained in step 3 and the power system status.

[0057] The contents not described in detail in this specification belong to the prior art known to those skilled in the art. It should be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0058] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0059] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0060] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims of the invention.

Claims

1. A power system dynamic economic dispatch system based on generative adversarial imitation learning, characterized by: It comprises a generator network module (1), a perfect dispatch module (2) and a discriminator network module (3); the generator network module (1) is used to observe the state of the power system, generate a dispatch strategy through the generator network and adopt a reinforcement learning proximal strategy optimization algorithm to obtain the determined output of the unit; the perfect dispatch module (2) is used to generate the perfect dispatch strategy; the discriminator network module (3) uses a discriminator network to compare the dispatch strategy generated by the generator network module (1) with the perfect dispatch strategy generated by the perfect dispatch module (2), and obtain feedback information for training the generator network parameters of the generator network module (1) and the discriminator network parameters of the discriminator network module (3); It also includes a generator output optimization module (4) and a network model offline training module (5); The generator output optimization module (4) generates an optimized dispatch control strategy by randomly sampling the normal distribution of the unit output, so as to avoid the generator learning falling into the local optimum point; the network model offline training module (5) uses the generator network module (1) to generate a dispatch plan according to the offline training scenario, and executes the dispatch plan in the power system environment, and the power system environment is transferred to a new state; after a decision cycle, the state and action of each decision step are collected into the generator memory to generate a generator memory state action sequence; for the same dispatch scenario, the perfect dispatch plan is obtained according to the perfect dispatch module (2), the system state and the perfect dispatch plan are collected into the expert memory to generate an expert memory state action sequence; the generator memory state action sequence and the expert memory state action sequence are input into the discriminator network module (3), and the discriminator network of the discriminator network module (3) outputs a discrimination result, and the generator network in the generator network module (1) updates the generator network parameters according to the discrimination result to generate a better dispatch plan; The specific implementation method of the generator output optimization module (4) generating the optimized scheduling control strategy is: The output P of the thermal power unit output by the generator network module (1) Gk The mean value of Thermal power unit output P Gk The standard deviation of The mean and standard deviation form a probability distribution of the output of thermal power units such as a normal distribution, and its probability density function is as follows: The mean and standard deviation of the output of the thermal power unit outputted by the generator network module (1) must conform to the probability density function. If they do not conform to the probability density function, the generator network module (1) will iteratively train to obtain a new output of the thermal power unit until the mean and standard deviation of the output of the new thermal power unit satisfy the normal distribution. The specific implementation method of the reinforcement learning proximal strategy optimization algorithm in the generator network module (1) is: The reinforcement learning proximal policy optimization algorithm defines the generator G θ The objective function is trained to generate a scheduling strategy. The expression of the objective function is: Among them, (s t ,a tg ) is the output of the generator network output layer a tg and the power system observation state s t The state-action pair composed of G is the generator strategy formed by the generator network, θ is the learning rate of the generator network, w is the learning rate of the discriminator network, D ω (s t ,a tg ) is the output data of the discriminator output layer, ε is a hyperparameter, π θ (a tg |s t ) is to follow the current strategy in state s t When selecting action a tg The probability of To follow the old strategy in state s t When selecting action a tg The probability of t (θ) is the state of t When following the current strategy and the old strategy, action a is selected tg The ratio of the probabilities of , clip is the clipping function.

2. The power system dynamic economic dispatch system based on generative adversarial imitation learning according to claim 1, characterized in that: The generator network in the generator network module (1) comprises an 8-layer network structure, which is, in order, a generator network input layer, a generator network normalization layer, a generator network first hidden layer, a generator network first activation function layer, a generator network second hidden layer, a generator network second activation function layer, a generator network output layer, and a generator network third activation function layer; The input data of the generator network input layer is the power system observation state s t ; The normalization layer is used to offset and scale different input features in the input data of the input layer; the output data of the output layer of the generator network is n groups of output layer neurons, and the output layer neurons contain mean and standard deviation. The n groups of units are determined through the n groups of output layer neurons.

3. The power system dynamic economic dispatch system based on generative adversarial imitation learning according to claim 2, characterized in that: The power system observation state s t It includes the load forecast value at time t+1, the new energy output forecast value at time t+1, the actual output of the thermal power unit at time t and the actual output of the new energy at time t, where the new energy includes wind power and photovoltaic power; The power system observation state s t It is expressed as follows: s t ={t,P′ load,t+1 ,P′ wind,t+1 ,P′ pv,t+1 ,P g,t ,P wind,t ,P pv,t } Among them, P′ load,t+1 represents the load forecast value at time t+1, P′ wind,t+1 represents the predicted wind power output at time t+1, P′ pv,t+1 represents the predicted photovoltaic output value at time t+1, P g,t represents the actual output of the thermal power unit at time t; P wind,t represents the actual wind power output value at time t, P pv,t is the actual photovoltaic output value at time t; The output variable of the generator network output layer is the output value a of the thermal power unit. tg : a tg ={P G1 ,…,P Gk ,…,P GK } Among them, P Gk is the output value of the kth thermal power unit; P GK is the output value of the Kth thermal power unit; K is the total number of thermal power units in the system.

4. The power system dynamic economic dispatch system based on generative adversarial imitation learning according to claim 1, characterized in that: The specific method of the perfect scheduling module (2) generating the perfect scheduling strategy is: After the real-time market ends each day, the perfect dispatch module (2) uses the known information at each time of the day, given the system operating state and operating conditions, and with the goal of minimizing the system power generation cost, uses the optimal unit combination and optimal power flow SCOPF that take into account the system safety operation constraints to calculate the perfect economic dispatch plan at 24 times of the day to generate a perfect dispatch strategy; the known information includes actual power load, actual new energy output and actual topological structure.

5. The power system dynamic economic dispatch system based on generative adversarial imitation learning according to claim 4, characterized in that: The discriminator network in the discriminator network module (3) comprises a 7-layer network structure, which is, in order, a discriminator network input layer, a discriminator network normalization layer, a discriminator network first hidden layer, a discriminator network first activation function layer, a discriminator network second hidden layer, a discriminator network second activation function layer, and a discriminator network output layer; The input data of the discriminator network input layer includes state-action pairs (s t ,a tg ) and state-action pairs (s t ,a te ), the state-action pair (s t ,a tg ) is the output of the generator network output layer a tg and the power system observation state s t The state-action pair (s t ,a te ) is the scheduling strategy a generated by the perfect scheduling module (2) te and the power system observation state s t The output data of the output layer of the discriminator network is the probability value of the perfect scheduling solution; The discriminator network adds the discriminator network normalization layer after the discriminator network input layer to normalize the state-action pairs; Adding the second activation function layer of the discriminator network before the output layer of the discriminator network, mapping the output to the interval [0,1] as the corresponding probability value that the input strategy is the perfect scheduling strategy; The discriminator network D ω The objective function is: Among them, (s t ,a tg ) is the output of the generator network output layer a tg and the power system observation state s t The state-action pair (s t ,a te ) is the scheduling strategy a generated by the perfect scheduling module (2) te and the power system observation state s t The state-action pair composed of G The generator strategy formed for the generator network; τ E The perfect scheduling strategy generated by the perfect scheduling module (2).

6. A method for dynamic economic dispatch of power system based on generative adversarial imitation learning using the system of claim 1, characterized in that: It includes the following steps: Step 1, the generator network module (1) observes the power system state, generates a dispatch strategy through the generator network using a reinforcement learning proximal strategy optimization algorithm, and obtains the determined output of the unit; Step 2, the perfect scheduling module (2) generates a perfect scheduling strategy; Step 3, the discriminator network module (3) uses a discriminator network to compare the scheduling strategy generated by the generator network module (1) with the perfect scheduling strategy generated by the perfect scheduling module (2), and obtains feedback information for training the generator network parameters of the generator network module (1) and the discriminator network parameters of the discriminator network module (3); Step 4, the generator network module (1) obtains the final dispatch strategy based on the trained generator network parameters obtained in step 3 and the power system status.

Images