A regional power grid dispatch optimization system, method and storage medium

By building a scheduling agent in the regional power grid, using deep neural networks and sample generation modules, the problems of real-time scheduling strategy calculation efficiency and rational decision-making under source load uncertainty are solved, and more efficient grid scheduling decisions and power generation plans are achieved. response.

CN118646012BActive Publication Date: 2024-11-05STATE GRID ANHUI ELECTRIC POWER CO LTD ELECTRIC POWER SCI RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411097826.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-11-05
Estimated Expiration
2044-08-12

AI Technical Summary

Technical Problem

The prior art is difficult to achieve the computing efficiency and rationality of decision-making of real-time scheduling strategies of regional power grids in the case of uncertainty of source loads, especially when taking into account the response to power generation plan and real-time effort decision-making.

Method used

A regional power grid scheduling optimization system is adopted, including basic action generation module, modulation action generation module and sample generation and amplification module. The scheduling agent is constructed through deep neural networks, and the real-time operating status of the power grid system is used to make decisions, combining historical data and predictions. Information generation and scheduling actions improve the robustness and efficiency of decision-making.

Benefits of technology

Improve the real-time scheduling and calculation efficiency and rational decision-making in the regional power grid under the uncertainty of source and load, and can better complete the response of online decision-making and long-term power generation plans, improve the rapid perception ability of high-latitude systems and Computational efficiency of scheduling strategies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118646012B_ABST
    Figure CN118646012B_ABST
Patent Text Reader

Abstract

A regional power grid dispatch optimization system, method and storage medium belong to the technical field of electric power system, and are used for active real-time power dispatch auxiliary decision-making, solving the problem of how to improve the calculation efficiency of the regional power grid real-time dispatch strategy under the source-load uncertainty scenario, thereby improving the rationality of decision-making; the system includes a basic action generation module, a modulation action generation module and a sample generation and amplification module for collaborative operation; the method includes: establishing a power grid dispatch model including power grid topology information, unit physical parameters, dispatching targets, power generation plans and dispatching data, new energy output and load forecast information; defining a dispatching strategy that at least includes the unit output in the next dispatching period; using auxiliary decision-making to solve the dispatching strategy, thereby improving the calculation efficiency, realizing the effectiveness of the regional power grid real-time dispatching strategy under the source-load uncertainty scenario and the reasonable response to the power generation plan on a longer time scale, and improving the rationality of decision-making.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of electric power systems and relates to a regional power grid dispatch optimization system, method and storage medium. Background Art

[0002] In order to better achieve the goal of energy conservation and emission reduction, the installed capacity of highly random renewable energy represented by wind power and photovoltaics has continued to increase in recent years. "Green", "low-carbon" and "sustainable" have become a major feature of power system operation. While renewable energy improves the cleanliness of power generation, it also introduces new risks and challenges to the safe and stable operation of the power system. Its more frequent short-term power fluctuations often bring additional grid power imbalances. High uncertainty often causes the power generation plans of traditional hydropower and thermal power units to deviate from the actual electricity demand, and greatly increases the secondary frequency regulation pressure of AGC units. Active real-time dispatch can correct the deviation between the power generation plan and the forecast results based on ultra-short-term forecasts, realize the advance dispatch of unit output, alleviate the impact of new energy uncertainty on the safe and stable operation of the power grid, and save power generation costs.

[0003] The existing data-driven autonomous learning optimization algorithm represented by deep reinforcement learning is an important means of solving scheduling strategies. This type of algorithm has excellent system global state perception and time-sequential decision-making capabilities, and does not rely on accurate mathematical models. It has good robustness when dealing with decision-making problems in uncertain environments. In addition, since it directly establishes a state-to-strategy mapping by deep neural networks and uses forward propagation to directly calculate the power generation instructions of the computer group when making decisions, it also has great advantages in the speed of solving strategies and is more suitable for high-dimensional online decision-making with high real-time requirements. The final output of the unit is usually determined by the day-ahead plan, intraday plan, real-time scheduling and frequency control. The existing solution method based on deep reinforcement learning usually decouples real-time scheduling from power generation plan when calculating the target setting point of the computer group. This method often ignores the reasonable response to the unit plan. Therefore, it is of great theoretical and application value to design an improved deep reinforcement learning algorithm that takes into account the reasonable response of the power generation plan and effectively completes the online decision-making of real-time output. Summary of the invention

[0004] The technical solution of the present invention is used to solve the problem of how to improve the calculation efficiency of the real-time dispatch strategy of the regional power grid under the source-load uncertainty scenario, thereby improving the rationality of decision-making.

[0005] The present invention solves the above technical problems through the following technical solutions:

[0006] A regional power grid dispatch optimization system includes: a basic action generation module, a modulation action generation module, and a sample generation and amplification module; the basic action generation module and the modulation action generation module are used to form a dispatching intelligent agent, and the sample generation and amplification module is used to assist in the training of the dispatching intelligent agent; the modules operate in coordination; the dispatching intelligent agent is trained by interacting with a power grid dispatching model;

[0007] The basic action generation module is constructed based on deep deterministic policy gradient, using deep neural network as the dispatch decision carrier, and the real-time operation state matrix of the power grid system is transformed into Input into the policy network to get the basic action , basic actions The next scheduling period generated by the basic action generation module Normalized reference value of active output of thermal power units;

[0008] The modulation action generation module uses a deep neural network as a modulation action generation carrier, and Input into the modulation network to obtain the modulation action for the next scheduling period , The role is based on the state component Generate basic actions in the basic action module Make corrections.

[0009] Furthermore, the basic action The calculation formula is as follows:

[0010] (1)

[0011] (2)

[0012] (3)

[0013] (4)

[0014] (5)

[0015] (6)

[0016] Among them, the status It consists of four parts, where For the moment To time A matrix of unit power generation plans in multiple windows, represents the time interval between each scheduling decision time, in formula (3) for time The real-time output of the unit is: Indicates time To time The load forecast data within the time window, in formula (5) Indicates time To time New energy output forecast data within the time window; Each element in corresponds to Taiwan's thermal power units in the future The planned power generation output for each time section; Each element in corresponds to the active output of each unit at that moment; Each element in the Load forecast data for each time section; Each element in the The predicted value of new energy output at each time section; in formula (6) Indicates The basic action vector is solved for input. The basic action vector contains elements, corresponding to There are thermal power units that need to be dispatched.

[0017] Furthermore, the modulation action The calculation formula is as follows:

[0018] (7)

[0019] in, Represents the output value calculated by the output layer of the neural network in the modulation action generation module.

[0020] Furthermore, the sample generation and amplification module is constructed as follows:

[0021] The definition is generated from the historical operation data. The training samples are the original samples , including features With label ; Define the first training samples, called augmented samples , including features With label ; All original samples constitute the original sample set , the amplified samples are stored in the amplified sample pool ; Based on the historical operation data of the power grid, Typical source and load scenario collection , and each original sample All belong to a specific scene in the scene set;

[0022] Assume that the power grid system operation database stores The historical operation data of the time section, then for the original sample set Middle Samples For example, the characteristics Characterizing the scheduling time section in historical operation data The system operation status information, that is, ,in ;Label Each dimension Characterization Units in time section The calculation formula for the scheduling task participation tendency is as follows:

[0023] (8)

[0024] (9)

[0025] (10)

[0026] In formula (8), Representative crew In the period The maximum power change within the range represents the unit's climbing ability. represents the rated output of the unit; in formula (9), For this unit The difference between the climbing mileage of the power generation plan and the scheduling point during the period; In formula (10), For this unit Deviation between actual and planned power generation during the time period; and Take a constant;

[0027] For a typical source-load scenario set Middle Scenes For example, suppose the scene contains The original samples are recorded as , define the scene label value as , the calculation formula is as follows:

[0028] (11)

[0029] For a state obtained during the interaction between the dispatching agent and the power grid dispatching model, its effect on the scene set is quantified according to the similarity distance such as Euclidean distance. The membership degree of each specific scene in the amplification sample is For example, suppose the sample has a scene With the highest membership, take the sample label value .

[0030] Furthermore, the method for the collaborative operation between the modules is as follows:

[0031] The scheduling agent uses the basic actions generated by the basic action generation module Modulation action generated by the modulation action generation module , combined with the state matrix middle The dispatching action output by the dispatching agent is calculated based on the power generation plan information of the units at each future time section given by :

[0032] (12)

[0033] (13)

[0034] In formula (12), is a constant quantity, and Respectively Time crew The upper and lower limits of active power that can be achieved, By Status Sub-items in Given; in formula (13) Each element is the scheduling action output by the scheduling agent, that is, Unit value of active output of thermal power units;

[0035] The samples generated by the sample generation and amplification module are used to train the parameters of the modulation network in the modulation action generation module in the form of supervised learning;

[0036] The loss function of the modulation network is calculated as follows:

[0037] (14)

[0038] In formula (14), From the original sample set Amplification sample pool The number of samples randomly selected by the CPC is The characteristics of the samples are , tagged as ,Will Input into the modulation network, the output is recorded as .

[0039] An optimization method applied to the above regional power grid dispatch optimization system comprises the following steps:

[0040] (1) Obtain topological and operational data including grid structure, unit physical parameters, power generation plan and dispatching data, renewable energy output and load forecast, and build a grid dispatching model to determine the dispatching decision interval , determine the number of thermal power units that need to be dispatched ;

[0041] (2) Constructing the original sample set based on historical operation data ;

[0042] (3) Constructing a set of typical source-load scenarios based on the binary k-nearest neighbor method ;

[0043] (4) Constructing the economic cost of unit output , Branch current over-limit cost Cost of system power imbalance , and construct the system cost function ;

[0044] (5) Construct optimization objective function;

[0045] (6) The dispatching agent interacts with the power grid dispatching model to collect training samples, train the deep neural network in each module, and save the parameters after training.

[0046] Furthermore, the economic cost of the unit output is , Branch current over-limit cost Cost of system power imbalance , and construct the system cost function The method is as follows:

[0047] (15)

[0048] (16)

[0049] (17)

[0050] (18)

[0051] In formula (15) , , For the The unit economic cost coefficient of each unit; in formula (16) is the number of branches in the power grid, for Moment The over-limit power of each branch, is the maximum power limit of the branch, is a constant, which takes the value of 1 when the system power exceeds the limit, otherwise it takes the value of 0; in formula (17) Indicates The power generation capacity of each new energy station, For the The actual power consumption of each load node, is the number of new energy stations, is the number of load nodes; in formula (18) , , They are the weight coefficients of economic cost, branch power flow over-limit cost, and system power imbalance cost respectively.

[0052] Furthermore, the construction optimization objective function is:

[0053] (19)

[0054] In formula (19), Represents the system cost function Take the minimum value.

[0055] Furthermore, the method of collecting training samples by interacting with the power grid dispatching model through the dispatching agent and training the deep neural network in each module is as follows:

[0056] 1) Set algorithm hyperparameters, including the parameter update step of the actor network, critic network, and modulation network , , , soft update rate , set up the experience replay sample pool for DDPG training Capacity , set up the augmented sample pool for modulation network training Capacity , set the number of samples used for training each time to , set the interaction time step between the dispatch agent and the power grid dispatch model , the maximum interaction time step is ;

[0057] 2) Initialize deep neural network parameters , , , and the network parameters , Assign values ​​to , ,make , ;

[0058] 3) Obtaining status from the power grid dispatch model ,make ;

[0059] 4) Enter the actor network and get basic actions ;Will Enter the modulation network and get the modulation action ; and generate scheduling actions scheduling actions ;

[0060] 5) Remember ,Will Input into the power grid dispatch model to obtain the cost And the next moment of power grid operation status ,remember ;

[0061] 6) Store experience replay sample pool ;

[0062] 7) Replay sample pool from experience Extract Samples , ;

[0063] 8) , Input into the critic network and get output ,Will Input into the target actor network and get output ,Will , Input into the target critic network and get the output , calculate the critic network loss according to formula (20), and update the step size based on Update network parameters using gradient descent ;

[0064] (20)

[0065] 9) Input into the actor network and get output , then , Input to the critic network and get output , calculate the actor network loss according to formula (21) , and based on the update step size Update network parameters using gradient descent ;

[0066] (twenty one)

[0067] 10) Generate amplified samples based on typical source-load scenarios, i.e., use the features Get the label value of the amplified sample , and store the amplified samples in the amplified sample pool ;

[0068] 11) From the amplified sample pool With the original sample set Combined sampling by random proportion Samples , ;

[0069] 12) Input modulation network, get output , calculate the modulation network loss , and is the parameter update step size, and the network parameters are updated using the gradient descent method;

[0070] 13) Use soft update to update the target actor network parameters and target critic network parameters as follows:

[0071] (twenty two)

[0072] (twenty three)

[0073] 14) Interaction time step ,like , then return to step 3), if , go to the next step.

[0074] A storage medium stores a computer program, which executes the steps of the above optimization method when executed by a processor.

[0075] The advantages of the present invention are:

[0076] The present invention enhances the learning ability of the dispatching agent on the empirical knowledge in the historical operation data, so that the real-time dispatching strategy generated by the dispatching agent can better complete the online decision-making of the real-time dispatching of the regional power grid under the source-load uncertainty scenario and the reasonable response to the power generation plan with a longer time scale. The present invention uses a deep neural network to realize the rapid perception of the power grid system state and the online generation of the dispatching strategy, which improves the rapid perception ability of the real-time operation data of the high-latitude system and the calculation efficiency of the dispatching strategy compared with the traditional method; on the premise of making a moderate response to the power generation plan with a longer time scale, the present invention flexibly and efficiently completes the calculation of the real-time dispatching strategy of the regional power grid under the source-load uncertainty scenario, thereby improving the rationality of the decision. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a structural diagram of the regional power grid dispatch optimization system of the present invention;

[0078] Figure 2 is a flow chart of the regional power grid dispatch optimization method of the present invention;

[0079] Figure 3 (a) is a heat map of the participation of each unit in the real-time scheduling task in a certain test scenario using the existing deep reinforcement learning-based optimization algorithm;

[0080] FIG3( b ) is a heat map of the participation of each unit in the real-time scheduling task in a certain test scenario according to the regional power grid scheduling optimization method of the present invention. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0082] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments:

[0083] Embodiment 1

[0084] like Figure 1 As shown, a regional power grid dispatch optimization system of the present invention includes: a dispatching agent including a basic action generation module, a modulation action generation module, and a sample generation and amplification module for auxiliary dispatching agent training, and the modules operate in coordination with each other; the dispatching agent is trained by interacting with the power grid dispatching model.

[0085] 1. Basic action generation module

[0086] Defining basic actions The next scheduling period generated by the basic action generation module The basic action generation module is built based on the Deep Deterministic Policy Gradient (DDPG), which uses a deep neural network as the dispatch decision carrier. Input into the policy network to get the basic action , the specific calculation formula is as follows:

[0087] (1)

[0088] (2)

[0089] (3)

[0090] (4)

[0091] (5)

[0092] (6)

[0093] Among them, the status It consists of four parts, where For the moment To time A matrix of unit power generation plans in multiple windows, represents the time interval between each scheduling decision time, in formula (3) for time The real-time output of the unit is: Indicates time To time The load forecast data within the time window, in formula (5) Indicates time To time New energy output forecast data within the time window; Each element in corresponds to Taiwan's thermal power units in the future The planned power generation output for each time section; Each element in corresponds to the active output of each unit at that moment; Each element in the Load forecast data for each time section; Each element in the The predicted value of new energy output at each time section; in formula (6) Indicates The basic action vector is solved for input. The basic action vector contains elements, corresponding to There are thermal power units that need to be dispatched.

[0094] 2. Modulation action generation module

[0095] The modulation action generation module uses a deep neural network as the modulation action generation carrier, which can be called a modulation network. The modulation action generation module converts the real-time operation state matrix of the power grid system into Input into the modulation network, and obtain the modulation action of the next scheduling period as shown in formula (7): . The physical meaning of is to describe the participation tendency of each unit in the real-time scheduling task, and its role is based on the state component Generate basic actions in the basic action module Corrections are made so that the final dispatching instructions of the dispatching agent do not deviate too much from the daily power generation plan, and sufficient dispatching margin is reserved to deal with the randomness of renewable energy; modulation action It is expressed as follows:

[0096] (7)

[0097] in, Represents the output value calculated by the output layer of the neural network in the modulation action generation module.

[0098] 3. Sample generation and amplification module

[0099] The definition is generated from the historical operation data. The training samples are the original samples , including features With label ; Define the first training samples, called augmented samples , including features With label , which is generated in the interaction process between the intelligent agent and the power grid dispatch model; all original samples constitute the original sample set , the amplified samples are stored in the amplified sample pool The definition contains Typical source and load scenario collection Based on the historical operation data of the power grid, each original sample All belong to a specific scene in the scene set.

[0100] Assume that the power grid system operation database stores The historical operation data of the time section, then for the original sample set Middle Samples For example, the characteristics Characterizing the scheduling time section in historical operation data The system operation status information, that is, ,in ;Label Each dimension Characterization Units in time section The calculation formula for the scheduling task participation tendency is as follows:

[0101] (8)

[0102] (9)

[0103] (10)

[0104] In formula (8) Representative crew In the period The maximum power change within the range represents the unit's climbing ability. represents the rated output of the unit; in formula (9), For this unit The difference between the climbing mileage of the power generation plan and the scheduling point during the period; In formula (10), For this unit Deviation between actual and planned power generation during the time period; and Take a constant.

[0105] For scene sets Middle Scenes For example, suppose the scene contains The original samples are temporarily recorded as , define the scene label value as , the calculation formula is as follows:

[0106] (11)

[0107] Furthermore, for a certain state obtained during the interaction between the dispatching agent and the power grid dispatching model, its effect on the scene set can be quantified according to similarity distances such as Euclidean distance. The membership degree of each specific scene in the amplification sample is For example, suppose the sample has a scene With the highest membership, take the sample label value In particular, if we remember any The characteristics are , tagged as In the present invention, the process of generating enhanced samples in the above manner is referred to as generating amplified samples based on typical source-load scenarios.

[0108] The basic action generation module, the modulation action generation module and the sample generation and amplification module operate in coordination as follows:

[0109] Defining a dispatch action is the next scheduling period output by the scheduling agent The dispatching agent uses the basic action generation module to generate the basic action. Modulation action generated by the modulation action generation module , combined with the state matrix middle The dispatching action output by the dispatching agent is calculated based on the power generation plan information of the units at each future time section given by the item :

[0110] (12)

[0111] (13)

[0112] In formula (12), is a constant quantity, and Respectively Time crew The upper and lower limits of active power that can be achieved, By Status Sub-items in Given; in formula (13) Each element is the scheduling action output by the scheduling agent, that is, The per unit active output value of each thermal power unit.

[0113] The samples generated by the sample generation and amplification module are used to train the parameters of the modulation network in the modulation action generation module in the form of supervised learning. Amplification sample pool CCP random sampling Samples, record The characteristics of the samples are , tagged as ,Will Input into the modulation network, the output is recorded as , based on the above The samples are averaged in the mean square error (MSE) mode, and the modulation network loss function is calculated by formula (14): as follows:

[0114] (14)

[0115] Embodiment 2

[0116] like Figure 2 As shown, a regional power grid dispatch optimization method of the present invention comprises the following steps:

[0117] Step 1: Obtain topological and operational data including grid structure, unit physical parameters, power generation plan and dispatching data, new energy output and load forecast, and build a grid dispatching model to determine the dispatching decision interval. , determine the number of thermal power units that need to be dispatched ;

[0118] Step 2: Build the original sample set based on historical operation data ;

[0119] Step 3: Construct a typical source-load scenario set based on the bi-k-means method ;

[0120] Step 4: Establish the economic value of unit output , Branch current over-limit cost Cost of system power imbalance , and construct the system cost function ;

[0121] Economic cost of unit output , Branch current over-limit cost Cost of system power imbalance , the calculation formula is as follows:

[0122] (15)

[0123] (16)

[0124] (17)

[0125] In formula (15) , , For the The unit economic cost coefficient of each unit; in formula (16) is the number of branches in the power grid, for Moment The over-limit power of each branch, is the maximum power limit of the branch, is a constant, which takes the value of 1 when the system power exceeds the limit, otherwise it takes the value of 0; in formula (17) Indicates The power generation capacity of each new energy station, For the The actual power consumption of each load node, is the number of new energy stations, is the number of load nodes;

[0126] System cost function for , , The weighted sum of is expressed as:

[0127] (18)

[0128] In formula (18), , , are the weight coefficients of economic cost, branch power flow over-limit cost, and system power imbalance cost respectively;

[0129] Step 5: Construct the optimization objective function as follows:

[0130] (19)

[0131] In formula (19), Represents the system cost function Take the minimum value.

[0132] Step 6: Collect training samples through the interaction between the dispatching agent and the power grid dispatching model, and train the deep neural network in each module. The specific method is as follows:

[0133] Step 6.1: Set the algorithm hyperparameters, including the parameter update step size of the actor network, critic network, and modulation network. , , , soft update rate , set up the experience replay sample pool for DDPG training Capacity , set up the augmented sample pool for modulation network training Capacity , set the number of samples used for training each time to , set the interaction time step between the dispatch agent and the power grid dispatch model , the maximum interaction time step is ;

[0134] Step 6.2: Initialize deep neural network parameters , , , and the network parameters , Assign values ​​to , ,make , ;

[0135] Step 6.3: Get the status from the power grid dispatch model ,make ;

[0136] Step 6.4: Enter the actor network and get basic actions ;Will Enter the modulation network and get the modulation action ; Generate scheduling actions according to formulas (12) and (13) ;

[0137] Step 6.5, record ,Will Input into the power grid dispatch model to obtain the cost And the next moment of power grid operation status ,remember ;

[0138] Step 6.6: Store experience replay sample pool ;

[0139] Step 6.7: Replay sample pool from experience Extract Samples , ;

[0140] Step 6.8: , Input into the critic network and get output ,Will Input into the target actor network and get output ,Will , Input into the target critic network and get the output , calculate the critic network loss according to formula (20), and update the step size based on Update network parameters using gradient descent ;

[0141] (20)

[0142] Step 6.9: Input into the actor network and get output , then , Input to the critic network and get output , calculate the actor network loss according to formula (21) , and based on the update step size Update network parameters using gradient descent ;

[0143] (twenty one)

[0144] Step 6.10: Generate amplified samples based on typical source-load scenarios, i.e., use the feature Get the label value of the amplified sample , and store the amplified samples in the amplified sample pool ;

[0145] Step 6.11: From the amplification sample pool With the original sample set Combined sampling by random proportion Samples , ;

[0146] Step 6.12: Input modulation network, get output , calculate the modulation network loss , and is the parameter update step size, and the network parameters are updated using the gradient descent method;

[0147] Step 6.13: As shown in equations (22) and (23), use soft-update to update the target actor network parameters and the target critic network parameters respectively.

[0148] (twenty two)

[0149] (twenty three)

[0150] Step 6.14: Interaction time step ,like , then return to step 6.3. If , the training is completed and the deep neural network parameters are saved.

[0151] Experimental Testing

[0152] FIG3 (a) is a heat map of the participation of each unit in the real-time scheduling task under a certain test scenario of the existing optimization algorithm based on deep reinforcement learning; FIG3 (b) is a heat map of the participation of each unit in the real-time scheduling task under a certain test scenario of the regional power grid scheduling optimization method of the present invention; the lighter the color in the figure, the greater the deviation between the real-time scheduling power set point and the power generation plan, and the darker the color, the closer the real-time scheduling power set point is to the power generation plan.

[0153] As shown in Figure 3 (a), the existing optimization algorithm based on deep reinforcement learning only considers the real-time scheduling stage for decision-making, and the strategy it generates usually causes the real-time power to deviate too much from the expected scheduling plan; as shown in Figure 3 (b), the method proposed in the present invention makes real-time scheduling decisions, and the strategy it generates can take into account the reasonable response to the power generation plan, achieving the goal of not deviating too much from the expected plan while coping with uncertainty, thereby greatly reducing the frequency of plan revisions.

[0154] Embodiment 3

[0155] A storage medium stores a computer program, which executes the steps of the regional power grid dispatch optimization method in embodiment 2 when the computer program is executed by a processor.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A regional power grid dispatch optimization system, characterized in that: include: Basic action generation module, modulation action generation module, sample generation and amplification module; the basic action generation module and modulation action generation module are used to form a dispatching agent, and the sample generation and amplification module is used to assist in the training of the dispatching agent; the modules operate in coordination; the dispatching agent is trained by interacting with the power grid dispatching model; The basic action generation module is constructed based on deep deterministic policy gradient, using deep neural network as the dispatch decision carrier, and the real-time operation state matrix of the power grid system is transformed into Input into the policy network to get the basic action , basic actions The next scheduling period generated by the basic action generation module Normalized reference value of active output of thermal power units; The modulation action generation module uses a deep neural network as a modulation action generation carrier, and Input into the modulation network to obtain the modulation action for the next scheduling period , The role is based on the state component Generate basic actions in the basic action module make corrections; The basic actions described The calculation formula is as follows: (1) (2) (3) (4) (5) (6) Among them, the status It consists of four parts, where For the moment To time A matrix of unit power generation plans in multiple windows, represents the time interval between each scheduling decision time, in formula (3) for time The real-time output of the unit is: Indicates time To time The load forecast data within the time window, in formula (5) Indicates time To time New energy output forecast data within the time window; Each element in corresponds to Taiwan's thermal power units in the future The planned power generation output for each time section; Each element in corresponds to the active output of each unit at that moment; Each element in the Load forecast data for each time section; Each element in the The predicted value of new energy output at each time section; in formula (6) Indicates The basic action vector is solved for input. The basic action vector contains elements, corresponding to The thermal power units that need to be dispatched; The modulation action The calculation formula is as follows: (7) in, represents the output value calculated by the output layer of the neural network in the modulation action generation module; The method for the collaborative operation between the modules is as follows: The scheduling agent uses the basic actions generated by the basic action generation module Modulation action generated by the modulation action generation module , combined with the state matrix middle The dispatching action output by the dispatching agent is calculated based on the power generation plan information of the units at each future time section given by the item : (12) (13) In formula (12), is a constant quantity, and Respectively Time crew The upper and lower limits of active power that can be achieved, By Status Sub-items in Given; in formula (13) Each element is the scheduling action output by the scheduling agent, that is, The per unit active output value of each thermal power unit; The samples generated by the sample generation and amplification module are used to train the parameters of the modulation network in the modulation action generation module in the form of supervised learning; The loss function of the modulation network is calculated as follows: (14) In formula (14), From the original sample set Amplification sample pool The number of samples randomly selected by the CPC is The characteristics of the samples are , tagged as ,Will Input into the modulation network, the output is recorded as .

2. The regional power grid dispatch optimization system according to claim 1, characterized in that: The construction method of the sample generation and amplification module is as follows: The definition is generated from the historical operation data. The training samples are the original samples , including features With label ; Define the first training samples, called augmented samples , including features With label ; All original samples constitute the original sample set , the amplified samples are stored in the amplified sample pool ; Based on the historical operation data of the power grid, Typical source and load scenario collection , and each original sample All belong to a specific scene in the scene set; Assume that the power grid system operation database stores The historical operation data of the time section, then for the original sample set Middle Samples For example, the characteristics Characterize the scheduling time section in historical operation data The system operation status information, that is, ,in ;Label Each dimension Characterization Units in time section The calculation formula for the scheduling task participation tendency is as follows: (8) (9) (10) In formula (8), Representative crew In the period The maximum power change within the range represents the unit's climbing ability. represents the rated output of the unit; in formula (9), For this unit The difference between the climbing mileage of the power generation plan and the scheduling point during the period; In formula (10), For this unit Deviation between actual and planned power generation during the time period; and Take a constant; For a typical source-load scenario set Middle Scenes For example, suppose the scene contains The original samples are recorded as , define the scene label value as , the calculation formula is as follows: (11) For a state obtained during the interaction between the dispatching agent and the power grid dispatching model, its effect on the scene set is quantified according to the similarity distance such as Euclidean distance. The membership degree of each specific scene in the amplification sample is For example, suppose the sample has a scene With the highest membership, take the sample label value .

3. An optimization method applied to the regional power grid dispatch optimization system according to any one of claims 1 to 2, characterized in that: The following steps are involved: (1) Obtain topological and operational data including grid structure, unit physical parameters, power generation plan and dispatching data, renewable energy output and load forecast, and build a grid dispatching model to determine the dispatching decision interval , determine the number of thermal power units that need to be dispatched ; (2) Constructing the original sample set based on historical operation data ; (3) Constructing a set of typical source-load scenarios based on the binary k-nearest neighbor method ; (4) Constructing the economic cost of unit output , Branch current over-limit cost Cost of system power imbalance , and construct the system cost function ; (5) Construct optimization objective function; (6) Collect training samples through the interaction between the dispatching agent and the power grid dispatching model, train the deep neural network in each module, and save the parameters after the training is completed; The economic cost of the unit output , branch current over-limit cost Cost of system power imbalance , and construct the system cost function The method is as follows: (15) (16) (17) (18) In formula (15) , , For the The unit economic cost coefficient of each unit; in formula (16) is the number of branches in the power grid, for Moment The over-limit power of each branch, is the maximum power limit of the branch, is a constant, which takes the value 1 when the system power exceeds the limit, otherwise it takes the value 0; in formula (17) Indicates The power generation capacity of each new energy station, For the The actual power consumption of each load node, is the number of new energy stations, is the number of load nodes; in formula (18) , , are the weight coefficients of economic cost, branch power flow over-limit cost, and system power imbalance cost respectively; The construction optimization objective function is: (19) In formula (19), Represents the system cost function Take the minimum value; The method of collecting training samples by interacting with the power grid dispatching model through the dispatching agent and training the deep neural network in each module is as follows: 1) Set algorithm hyperparameters, including the parameter update step of the actor network, critic network, and modulation network , , , soft update rate , set up the experience replay sample pool for DDPG training Capacity , set up the augmented sample pool for modulation network training Capacity , set the number of samples used for training each time to , set the interaction time step between the dispatch agent and the power grid dispatch model , the maximum interaction time step is ; 2) Initialize deep neural network parameters , , , and the network parameters , Assign values ​​to , ,make , ; 3) Obtaining status from the power grid dispatch model ,make ; 4) Enter the actor network and get basic actions ;Will Enter the modulation network and get the modulation action ; and generate scheduling actions scheduling actions ; 5) Remember ,Will Input into the power grid dispatch model to obtain the cost And the next moment of power grid operation status ,remember ; 6) Store experience replay sample pool ; 7) Replay sample pool from experience Extract Samples , ; 8) , Input into the critic network and get output ,Will Input into the target actor network and get output ,Will , Input into the target critic network and get the output , calculate the critic network loss according to formula (20), and update the step size based on Update network parameters using gradient descent ; (20) 9) Input into the actor network and get output , then , Input to the critic network and get output , calculate the actor network loss according to formula (21) , and based on the update step size Update network parameters using gradient descent ; (21) 10) Generate amplified samples based on typical source-load scenarios, i.e., use the features Get the label value of the amplified sample , and store the amplified samples in the amplified sample pool ; 11) From the amplified sample pool With the original sample set Combined sampling by random proportion Samples , ; 12) Input modulation network, get output , calculate the modulation network loss , and is the parameter update step size, and the network parameters are updated using the gradient descent method; 13) Use soft update to update the target actor network parameters and target critic network parameters as follows: (22) (23) 14) Interaction time step ,like , then return to step 3), if , go to the next step.

4. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the optimization method according to claim 3 are executed.