Intraday dispatch optimization method for power grid considering photovoltaic output uncertainty under adverse weather

Abstract

The application belongs to the technical field of power systems, and more particularly to a power grid intra-day scheduling optimization method considering photovoltaic output uncertainty under severe weather, which uses an A2C algorithm framework and knowledge transfer technology to solve and optimize the power grid scheduling problem under photovoltaic output uncertainty under severe weather. First, the sequential decision-making process of power grid intra-day scheduling is determined, and a mathematical model is established, and then the optimization objectives and constraint conditions of power grid intra-day scheduling are determined; secondly, the CGAN scene generation method is used to generate training samples, and the A2C algorithm framework is used to optimize the scheduling intelligent agent that can adapt to various source and load scenes. Finally, transfer learning is used to improve the optimization of the specific intra-day scheduling intelligent agent, solving the problem that the accuracy of the prediction data of large-scale photovoltaic access to the power system greatly affects the stable operation of the power grid, and ensuring that the power grid operates in a relatively economical manner under the premise of safety.

Description

Technical Field

[0001] This invention belongs to the field of power system technology, and more specifically, relates to a method for optimizing intraday grid dispatch considering the uncertainty of photovoltaic output under severe weather conditions. This method uses the A2C algorithm framework and knowledge transfer technology to solve and optimize the grid dispatch problem with the uncertainty of photovoltaic output under severe weather conditions. Background Technology

[0002] Solar power output is significantly affected by seasons and weather, and there is a large discrepancy between the day-ahead forecast and the actual value. Therefore, in power systems with large-scale solar PV integration, the accuracy of solar PV forecast data has a significant impact on the stable operation of the power grid. Thus, considering various solar power output scenarios for the following day is crucial.

[0003] Existing research has widely used reinforcement learning algorithms to solve and optimize power grid scheduling problems in power systems. However, the output of photovoltaic (PV) units is greatly affected by weather and is difficult to predict accurately. A single deep reinforcement learning algorithm cannot fully adapt to the complex scheduling needs under such high uncertainty scenarios. Therefore, this paper proposes a daily power grid scheduling optimization method that considers the uncertainty of PV output under severe weather conditions. This aims to further improve the intelligence and adaptability of power grid scheduling, providing support for the operation of power grids with a high proportion of renewable energy integration. Summary of the Invention

[0004] This invention establishes a grid safety and economic dispatch model with photovoltaic new energy access based on intraday optimized dispatch mode, and proposes an intraday grid dispatch optimization method that considers the uncertainty of photovoltaic output under severe weather conditions, which enables the grid to operate in a more economical manner while ensuring safety.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A grid intraday dispatch optimization method considering the uncertainty of photovoltaic output under severe weather conditions includes the following steps:

[0007] Step 1: Determine the sequential decision-making process for intraday power grid dispatch and establish a mathematical model;

[0008] Step 2: Determine the daily scheduling optimization objectives and constraints of the power grid and establish its learning optimization model;

[0009] Step 3: Use the CGAN scene generation method to generate training samples, and use the A2C algorithm framework to optimize the scheduling agent that can adapt to various source load scenarios;

[0010] Step 4: Use transfer learning to obtain intraday scheduling agents for specific source load scenarios.

[0011] The sequential decision-making process for intraday power grid scheduling described in step 1 is specifically modeled as follows:

[0012] The sequential decision model first requires, at time t, to construct a state matrix from the observed power grid operation information, denoted as S. t The state matrix mainly includes the current time t and the observation information ob of the power grid operating state. t And the load forecast and the output forecast of new energy units for the next time t+1. t Therefore, the state matrix can be written in the following form:

[0013] S t =[t,ob t fore t ],t∈(0,1,2,...,T)

[0014] The time interval is 15 minutes, so T = 95. The agent receives a state matrix S. t After being input, the agent policy network π will output a unit action adjustment A based on the information in the current state matrix. t Assume the system contains N ther For each thermal power unit, the adjustment amount for the operation of the nth thermal power unit at time t is denoted as... Adjustment action matrix A for all units t It can be written in the form of a matrix as follows:

[0015]

[0016] The unit's adjustment needs to be mapped to the current unit's output value. The mapping formula is as follows:

[0017]

[0018] In the formula P t ther This represents the output value of all thermal power generating units in the system at time t. These represent the lower and upper limits of the allowable adjustable output of all thermal power units in the system at time t, respectively.

[0019] As mentioned above, the state matrix S at time t t The input is fed into strategy π, which determines the adjustment amount A. t Subsequently, the power grid adjusts and implements the corresponding thermal power units, and then proceeds to the next decision-making process, which can be described using mathematical notation as follows:

[0020] S t →π→A t →S t+1 →π→A t+1 …→ST

[0021] The constraints of the intraday power grid dispatch model described in step 2 include line power flow constraints, power generation and consumption balance, upper and lower limits of thermal power unit ramping power, and upper and lower limits of unit output power, as detailed below:

[0022] The power flow constraint of the line:

[0023]

[0024] in, Let be the line power matrix composed of the power of all lines at time t. The line limit power matrix is ​​composed of all line power limit values, l∈N line N line Numbering of all lines on all bus nodes.

[0025] The power generation and utilization balance:

[0026]

[0027] in N at time t PV The power matrix of the photovoltaic units output, Let P be the output power of the j-th photovoltaic unit at time t. t load It is the active power on the electricity consumption side.

[0028] The upper and lower limit power constraints of the unit at any time t:

[0029]

[0030] in This is the minimum / maximum output power matrix for thermal power units. This represents the minimum / maximum output power of the kth thermal power unit.

[0031] The climbing power constraint:

[0032]

[0033] in This is a matrix composed of the downward / upward ramping power of thermal power units. Let represent the downward / upward ramping power of the nth thermal power unit.

[0034] The power grid optimization objectives mentioned in step 2 are as follows:

[0035] The power grid's operational objectives primarily consider the operating costs of thermal power units and the grid integration rate of photovoltaic units. Among these, the operating costs of thermal power units... for:

[0036]

[0037] Where, α i ,β i ,γ i Let be the operating cost coefficient of group i when using the quadratic model. The photovoltaic grid integration rate can be expressed as:

[0038]

[0039] in This represents the short-term predicted maximum output of the j-th photovoltaic unit at time t. This represents the current output value of the j-th photovoltaic unit at time t.

[0040] Step 3 describes the CGAN scene generation method for generating training samples as follows:

[0041] CGAN is an extended version of Generative Adversarial Networks (GANs) that guides the data generation process by introducing conditional information. In power system dispatching scenarios, CGAN can generate realistic scenario data based on specific input conditions (such as meteorological data, load demand, and power generation plans) to simulate the output or source-load scenarios of new energy sources such as photovoltaics and wind power. CGAN consists of two main components: a generator and a discriminator. The generator generates virtual samples G(z|c) that meet the conditions based on the input conditional variable c and random noise z. The discriminator, by inputting real samples and generated samples, determines whether the input sample is real data and outputs the true probability. On the input side of CGAN, the random noise matrix d and the label matrix l are merged and used as the input data for the generator G(d|l), finally producing generated samples x′=G(d|l) that meet the label values. The discriminator D(x|l) not only needs to determine the magnitude p(x) of the sample conforming to the distribution of real samples, but also needs to determine whether the sample belongs to condition l. Therefore, during training, the sample distribution p output by the generator G(d|l) should be optimized as much as possible. g (x′) continuously matches the distribution p(x) of samples in historical data; while the discriminator D(x|l) should strive to improve its ability to distinguish between true and false samples. In new energy power systems, the output of photovoltaic and wind power is affected by various factors such as weather and time, and their volatility and uncertainty pose a great challenge to grid dispatch. Through CGAN scene generation technology, rich simulation data can be provided for grid dispatching agents, enhancing the robustness of dispatching models.

[0042] The photovoltaic output data is divided into multiple scene sets with different characteristics according to month or solar term, which serve as condition values. Assuming the dataset is divided into m scene sets, and considering the randomness of photovoltaic output, a noise value is matched to each moment of a sample. The noise sequence d is defined as follows: d = {d1, d2, d3, ..., d...} i In the formula: i — the total number of time points for a sample. Similarly, the condition values ​​are divided according to the number of scene sets, and the condition sequence l is defined as follows: l = {l1, l2, l3, ..., l...} m The two sequences are concatenated, and the input matrix is ​​first processed through a fully connected layer for feature extraction. Then, a nonlinear function is added between each fully connected layer. This nonlinearity allows the neural network to more accurately approximate the mapping relationship between the conditional label values, noise distribution, and the power output distribution of the new energy units. After layer-by-layer calculation, the input and output of each layer are obtained, ultimately yielding the output value of the fully connected layer.

[0043] This model generates realistic wind and solar power scenes by learning the temporal characteristics and error distribution of wind and solar power data. Specifically, the model first uses an LSTM network to capture the temporal features of the wind and solar power data, then combines kernel density estimation to fit the wind power prediction error and generate random noise that conforms to the error distribution. This noisy data is then used to train the CGAN model. Simultaneously, the day-ahead predicted values ​​of wind and solar power are used as conditional input to help the model learn the mapping relationship between the conditional noise distribution and the actual data. The entire design aims to more accurately simulate the randomness and uncertainty of wind and solar power, providing an effective solution for generating realistic wind and solar power scenes.

[0044] The scheduling agent is trained to adapt to various source-load scenarios. The CGAN scenario generation method is used to generate photovoltaic power output scenarios. Based on the characteristics of the photovoltaic data during the period, such as the maximum output, half-load probability and output period, three main source-load scenario samples of sunny, cloudy and rainy days are selected. These scenario data can be used to verify the grid scheduling strategy under different photovoltaic power output conditions.

[0045] The A2C algorithm optimization described in step 3 first obtains the state from the environment, then inputs it into the agent's network. The actor network outputs the mean and variance, constructs a corresponding Gaussian distribution, and samples actions. Finally, the actions are input into the environment for execution. After execution, the environment outputs the state information for the next time step. The agent's critic network calculates the corresponding value estimate based on the current state and the state at the next time step. It further calculates the mean squared error based on the real-time reward feedback from the environment and uses this error to update the agent's critic network. Finally, it updates the actor network based on the output action information.

[0046] In step 4, transfer learning is used to improve the optimization of the scheduling agent for a specific day:

[0047] Network-based transfer learning is used, which involves reusing a pre-trained network from the source domain, including its network structure and connection parameters, to transform it into parameters for initial training of a deep neural network in the target domain. The specific formula is shown below:

[0048]

[0049] In the formula π source ,π target The policy networks θ represent the source task and the target task, respectively. source The network parameters θ represent the source task policy network. t ′ arget The initial network parameters represent the target task's parameters. Therefore, the parameters of the source task's policy network are directly transferred to the target task's policy network, and then training continues. This significantly improves training efficiency and yields better results.

[0050] Unlike existing technologies, the above technical solution has the following beneficial effects:

[0051] To address the intraday optimal scheduling problem of power grids with photovoltaic (PV) renewable energy integration, this invention proposes an intraday intelligent scheduling optimization method based on the A2C algorithm framework and a knowledge transfer approach. CGAN scene generation technology is used to expand the output samples of PV units, providing a large set of source-load scenarios for the scheduling agent. Then, the knowledge transfer method is used to transfer scheduling knowledge adapted to agents with various source-load scenarios to agents targeting specific intraday scheduling. This optimization method enables the power grid to operate in the most economical way while ensuring safety. Attached Figure Description

[0052] Figure 1 A diagram illustrating the sequential decision-making process for intraday power grid dispatching;

[0053] Figure 2 Flowchart for A2C algorithm optimization;

[0054] Figure 3 A simplified diagram illustrating the CGAN scene generation method;

[0055] Figure 4 This is a diagram illustrating the knowledge transfer process for scheduling.

[0056] Figure 5 A comparison chart of off-load power and operating costs of thermal power units with and without transfer learning;

[0057] Figure 6 Optimize the total reward curve for learning with and without transfer learning. Detailed Implementation

[0058] To provide a detailed explanation of the technical content, structural features, objectives, and effects of the load under load technical solution, the following description, in conjunction with specific embodiments and accompanying drawings, will be provided.

[0059] This invention discloses a method for intraday grid dispatch optimization considering the uncertainty of photovoltaic (PV) output under severe weather conditions. Given that PV output is significantly affected by seasons and weather, and that its day-ahead forecasts deviate considerably from actual values, and that a single deep learning algorithm is insufficient to address grid dispatch issues arising from the integration of large-scale PV units, this invention employs a CGAN-based scenario generation method to construct a PV output sample set. Then, an A2C algorithm framework is used to optimize the dispatch agent. The proposed dispatch optimization method ensures the safe and economical operation of the power grid under various severe source-load scenarios.

[0060] This invention is specifically illustrated using a provincial power grid as an example. The invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0061] See also Figure 1 The diagram shows the sequential decision-making process for intraday power grid dispatching.

[0062] Step 1: The sequential decision-making process for intraday power grid dispatching is determined as follows:

[0063] First, state information is obtained from the environment and then input into the agent's network. The agent's actor network outputs the mean and variance, and then constructs the corresponding Gaussian distribution and samples the actions. Next, the actions are input into the environment for execution, and after execution, the environment outputs the state information for the next time step.

[0064] Power optimization scheduling divides the scheduling problem over a period of time into multiple scheduling problems at smaller time scales. The scheduling duration is 24 hours, and a scheduling plan is formulated every 15 minutes, for a total of 96 scheduling plans per day. First, at time t, the observed information is constructed into a state matrix, denoted as S. t This includes the current time t and observational information about the power grid's operating status ob. t And the load forecast and the output forecast of new energy units for the next time t+1. t Therefore, the state matrix can be written in the following form:

[0065] S t =[t,ob t fore t ],t∈(0,1,2,...,T)

[0066] Wherein, the observation matrix ob tIt consists of variables in the operation of the power grid, mainly including bus load, current generation information of each generating unit, and information of each node in the current system. When the agent receives a state matrix S at time t... t After being input, the agent policy network π will output a unit action adjustment A based on the information in the current state matrix. t Adjustment action matrix A for all units t It can be written in the form of a matrix as follows:

[0067]

[0068] The unit's adjustment needs to be mapped to the current unit's output value. The mapping formula is as follows:

[0069]

[0070] In the formula P t ther This represents the output value of all thermal power generating units in the system at time t, so in Let be the output value of the i-th thermal power unit at time t. These represent the lower and upper limits of the allowable adjustable output of all thermal power units in the system at time t, respectively. in Let be the adjustable lower or upper bound of the i-th thermal power unit at time t.

[0071] Step 2: Determine the daily power grid dispatch optimization objectives and constraints:

[0072] The most crucial aspect of ensuring the safe operation of a power system is guaranteeing grid safety during intraday power optimization dispatch. Dispatch safety primarily involves ensuring that the power transmitted through grid lines does not exceed their power limits, while also maintaining a balance between power generation and consumption on both the source and load sides. Safe grid operation also requires addressing the active power distribution problem among generating units, and further, satisfying certain constraints, such as line power flow constraints, upper and lower limits for thermal power unit ramping power, and upper and lower limits for unit output power. As shown in the following equation:

[0073] The power flow constraint of the line:

[0074]

[0075] in, Let be the line power matrix composed of the power of all lines at time t. The line limit power matrix is ​​composed of all line power limit values, l∈N line N line Numbering of all lines on all bus nodes.

[0076] For the safe operation of a power grid, frequency stability must be ensured. To ensure frequency stability, the active power on the generation side must be the same as the active power on the consumption side. Therefore, at any time t, the system should satisfy the following power balance equation:

[0077]

[0078] in N at time t PV The power matrix of the photovoltaic units output, Let be the output power of the j-th photovoltaic unit at time t.

[0079] The upper and lower power limits of the thermal power unit:

[0080]

[0081] in This is the minimum / maximum output power matrix for thermal power units. This represents the minimum / maximum output power of the kth thermal power unit.

[0082] The climbing constraint of the thermal power unit:

[0083]

[0084] in This is a matrix composed of the downward / upward ramping power of thermal power units. Let represent the downward / upward ramping power of the nth thermal power unit.

[0085] The power grid's operational objectives primarily consider the operating costs of thermal power units and the grid integration rate of photovoltaic units. The operating costs of the units... The calculation formula is:

[0086]

[0087] Where, α i ,β i ,γ i Let be the operating cost coefficient for group i when using a quadratic model. The formula for the photovoltaic grid integration rate is as follows:

[0088]

[0089] in This represents the short-term predicted maximum output of the j-th photovoltaic unit at time t. Let represent the current output value of the j-th photovoltaic unit at time t. If the short-term predicted output of the photovoltaic system is zero, its grid integration rate is set to 100%. In the actual optimization process, some constraints need to be introduced as penalty terms into the objective function; the specific objective function is given in the example.

[0090] Step 3: Use the CGAN scene generation method to generate training samples, and use the A2C algorithm framework to optimize the scheduling agent that can adapt to various source load scenarios.

[0091] CGAN is an extended version of Generative Adversarial Networks (GANs) that guides the data generation process by introducing conditional information. In power system dispatching scenarios, CGAN can generate realistic scenario data based on specific input conditions (such as meteorological data, load demand, and power generation plans) to simulate the output or source-load scenarios of new energy sources such as photovoltaics and wind power. CGAN consists of two main components: a generator and a discriminator. The generator generates virtual samples G(z|c) that meet the conditions based on the input conditional variable c and random noise z. The discriminator, by inputting real samples and generated samples, determines whether the input sample is real data and outputs the true probability. On the input side of CGAN, the random noise matrix d and the label matrix l are merged and used as the input data for the generator G(d|l), finally producing generated samples x′=G(d|l) that meet the label values. The discriminator D(x|l) not only needs to determine the magnitude p(x) of the sample conforming to the distribution of real samples, but also needs to determine whether the sample belongs to condition l. Therefore, during training, the sample distribution p output by the generator G(d|l) should be optimized as much as possible. g (x′) continuously matches the distribution p(x) of samples in historical data; while the discriminator D(x|l) should strive to improve its ability to distinguish between true and false samples. In new energy power systems, the output of photovoltaic and wind power is affected by various factors such as weather and time, and their volatility and uncertainty pose a great challenge to grid dispatch. Through CGAN scene generation technology, rich simulation data can be provided for grid dispatching agents, enhancing the robustness of dispatching models.

[0092] The photovoltaic output data is divided into multiple scene sets with different characteristics according to month or solar term, which serve as condition values. Assuming the dataset is divided into m scene sets, and considering the randomness of photovoltaic output, a noise value is matched to each moment of a sample. The noise sequence d is defined as follows: d = {d1, d2, d3, ..., d...} i In the formula: i — the total number of time points for a sample. Similarly, the condition values ​​are divided according to the number of scene sets, and the condition sequence l is defined as follows: l = {l1, l2, l3, ..., l...} mThe two sequences are concatenated, and the input matrix is ​​first processed through a fully connected layer for feature extraction. Then, a nonlinear function is added between each fully connected layer. This nonlinearity allows the neural network to more accurately approximate the mapping relationship between the conditional label values, noise distribution, and the power output distribution of the new energy units. After layer-by-layer calculation, the input and output of each layer are obtained, ultimately yielding the output value of the fully connected layer.

[0093] This model generates realistic wind and solar power scenes by learning the temporal characteristics and error distribution of wind and solar power data. Specifically, the model first uses an LSTM network to capture the temporal features of the wind and solar power data, then combines kernel density estimation to fit the wind power prediction error and generate random noise that conforms to the error distribution. This noisy data is then used to train the CGAN model. Simultaneously, the day-ahead predicted values ​​of wind and solar power are used as conditional input to help the model learn the mapping relationship between the conditional noise distribution and the actual data. The entire design aims to more accurately simulate the randomness and uncertainty of wind and solar power, providing an effective solution for generating realistic wind and solar power scenes.

[0094] The scheduling agent is trained to adapt to various source-load scenarios. The CGAN scenario generation method is used to generate photovoltaic power output scenarios. Based on the characteristics of the photovoltaic data during the period, such as the maximum output, half-load probability and output period, three main source-load scenario samples of sunny, cloudy and rainy days are selected. These scenario data can be used to verify the grid scheduling strategy under different photovoltaic power output conditions.

[0095] See also Figure 2 The flowchart shown below illustrates the optimization process of the A2C algorithm, which includes the following steps:

[0096] Step 3.1: First, a photovoltaic power output sample set under various weather scenarios is constructed using a scenario generation method based on conditional generative adversarial networks. Then, based on the characteristics of the photovoltaic data during that period, such as the maximum output, half-load probability, and output period, three main source-load scenario samples are selected: sunny day, cloudy day, and rainy day.

[0097] Step 3.2: Then, based on the generated sample set, the dominant actor-critic algorithm is used to optimize the scheduling agent that can adapt to multiple source-load scenarios. The intraday scheduling plan needs to be formulated every 15 minutes, therefore, timeliness is crucial. This section utilizes a data-driven deep reinforcement learning algorithm. First, the intraday scheduling agent is trained offline based on historical data and scenario generation technology. Then, the offline-learned scheduling agent receives real-time grid information and real-time forecast information of new energy sources and loads as input for online decision-making. The intraday scheduling agent adaptable to multiple source-load scenarios is trained using the A2C algorithm framework. First, the output of new energy sources and the load forecast curve are input as part of the state matrix S into the general agent. In addition, the state matrix S also includes bus load and other node information, as well as the current unit's output plan. After receiving the input, the agent outputs the adjustment action vector A for each unit and calculates the current unit's output value through these actions. The current unit output value is then imported into the grid environment. After power flow calculation and constraint checks, the system's safety, economic, cleanliness, and load curtailment penalties are calculated. These indicators are then weighted according to their importance to obtain the corresponding cost reward R. Finally, the grid simulation environment updates the current unit output plan in the current state with the final executed unit output plan.

[0098] To improve the learning efficiency of the agent, a parallel architecture using CPU multithreading is employed to accelerate computation. Specifically, the entire policy network is divided into a main network and multiple parallel, independent sub-threads. Each sub-thread loads parameters from the main network and interacts independently with its own environment to obtain independent sampling trajectories. Simultaneously, different source-load scenarios are set for each sub-grid environment within the power grid's operating environment: sunny day scenarios for sub-environments 1 to n, cloudy day scenarios for sub-environments (n+1 to 2n), and rainy day scenarios for sub-environments (2n+1 to 3n). By setting scenario parameters for different sub-environments, the policy network can explore and learn under varying climatic conditions, enabling the scheduling agent to learn decision-making methods for different situations and improving its ability to operate safely under various harsh power grid scenarios.

[0099] Step 4: Use transfer learning to improve the optimization of the scheduling agent for a specific day.

[0100] Step 4.1: Using network-based transfer learning, the pre-trained network in the source domain, including its network structure and connection parameters, is reused and transformed into deep neural network parameters for initial training in the target domain. The purpose of transfer learning is to utilize the source domain D... s In learning task T sChina's experience in safely and effectively adjusting unit output for various climatic load scenarios can help improve the target domain D. t In learning task T t The learning speed of the policy network and its ability to make safe scheduling decisions in harsh scenarios are discussed. The specific formula is shown below:

[0101]

[0102] In the formula π source ,π target The policy networks θ represent the source task and the target task, respectively. source The network parameters θ represent the source task policy network. t ′ arget The initial network parameters represent the target task. Therefore, the parameters of the source task's policy network are directly transferred to the target task's policy network, and then learning continues.

[0103] In power systems with large-scale photovoltaic (PV) integration, traditional deep learning algorithms struggle to fully meet the needs of grid dispatch optimization when dealing with the volatility and uncertainty of PV output. This invention proposes a grid intraday dispatch optimization method that considers the uncertainty of PV output under severe weather conditions.

[0104] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Unless otherwise specified, an element defined by the phrase "comprising..." or "including..." does not exclude the presence of additional elements in the process, method, article, or terminal device that includes said element. Additionally, in this document, "greater than," "less than," "exceeding," etc., are understood to exclude the stated number; "above," "below," "within," etc., are understood to include the stated number.

[0105] Although the above embodiments have been described, those skilled in the art, once they understand the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the above descriptions are merely embodiments of the present invention and do not limit the scope of patent protection of the present invention. Any equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A grid intraday dispatch optimization method considering the uncertainty of photovoltaic output under severe weather conditions, characterized in that, Includes the following steps: Step 1: Determine the sequential decision-making process for intraday power grid dispatch and establish a mathematical model; Step 2: Determine the daily scheduling optimization objectives and constraints of the power grid and establish its learning optimization model; Step 3: Use the CGAN scene generation method to generate training samples, and use the A2C algorithm framework to optimize the scheduling agent that can adapt to various source load scenarios; Step 4: Use transfer learning to obtain intraday scheduling agents for specific source-load scenarios; By setting scenario parameters for different sub-environments, the policy network can explore and learn under different climatic power grid operation scenarios, thereby enabling the scheduling agent to learn decision-making methods under different conditions and improve the agent's ability to operate safely under various harsh power grid scenarios.

2. The grid intraday dispatch optimization method considering the uncertainty of photovoltaic output under severe weather conditions as described in claim 1, characterized in that, The sequential decision-making process for intraday power grid scheduling in step 1 is specifically modeled as follows: The sequential decision model first in the... At any given moment, the observed power grid operation information needs to be constructed into a state matrix, and denoted as... The state matrix mainly includes the current time step number. Observational information on the operating status of the power grid and the next moment Forecast values ​​of load and output of new energy units The state matrix has the following form: The time interval is 15 minutes, so The agent receives a state matrix. After being input, the agent policy network It will output a unit action adjustment amount based on the information in the current state matrix. Assume the system contains Taiwan thermal power units, on the [date] At the nth moment, for the first The adjustment amount of the thermal power unit is recorded as follows: Adjustment action matrix for all units It takes the form of a matrix as follows: The unit's adjustment needs to be mapped to the current unit's output value. The mapping formula is as follows: In the formula represent The output values ​​of all thermal power generating units in the time system. , Represent The adjustable lower and upper limits of the output of all thermal power units in the time system; As mentioned above, Time-state matrix Input to Policy In the middle, the decision is made to adjust the amount. Subsequently, the power grid adjusts and implements the corresponding thermal power units, and then proceeds to the next decision-making process, which can be described using mathematical symbols: 。 3. The grid intraday dispatch optimization method considering the uncertainty of photovoltaic output under severe weather conditions as described in claim 2, characterized in that, The constraints of the intraday power grid dispatch model in step 2 include line power flow constraints, power generation and consumption balance, upper and lower limit power constraints of generating units, and upper and lower limit constraints of thermal power unit ramping power, as detailed below: The power flow constraint of the line: in, for The line power matrix, composed of the power of all lines at any given time. This is a line limit power matrix composed of all line power limits. , Numbering of all lines on all busbar nodes; The power generation and utilization balance: in for time The power matrix of the photovoltaic units output, for Time of the first The output power of the photovoltaic unit, It is the active power on the power consumption side; arbitrary Timing unit upper and lower limit power constraints: in This is the minimum / maximum output power matrix for thermal power units. For the first Minimum / maximum output power of the thermal power unit; the climbing power constraint: in This is a matrix composed of the downward / upward ramping power of thermal power units. For the first The downward / upward ramping power of the thermal power unit.

4. The grid intraday dispatch optimization method considering the uncertainty of photovoltaic output under severe weather conditions as described in claim 3, characterized in that, The power grid optimization objectives in step 2 are as follows: The power grid's operational objectives primarily consider the operating costs of thermal power units and the grid integration rate of photovoltaic units, with the operating costs of thermal power units being the most significant factor. for: in, When using a quadratic model The operating cost coefficient, and the photovoltaic grid integration rate are expressed as: in express Time of the first Taiwan's short-term forecast maximum power output express Time of the first The current output value of the photovoltaic unit.

5. The grid intraday dispatch optimization method considering the uncertainty of photovoltaic output under severe weather conditions as described in claim 1, characterized in that, The CGAN scene generation method in step 3 generates training samples as follows: CGAN consists of two main components: a generator and a discriminator. The generator generates virtual samples G(z|c) that meet the conditions based on the input conditional variable c and random noise z. The discriminator judges whether the input sample is real data by taking the input real sample and the generated sample, and outputs the true probability. The scheduling agent is trained to adapt to various source-load scenarios. The CGAN scenario generation method is used to generate photovoltaic power output scenarios. Based on the maximum output, half-load probability and output time characteristics of photovoltaic data during the period, three main source-load scenario samples of sunny day, cloudy day and rainy day are selected.

6. The grid intraday dispatch optimization method considering the uncertainty of photovoltaic output under severe weather conditions as described in claim 1, characterized in that, In step 3, the A2C algorithm optimization first obtains the state from the environment and then inputs it into the agent's network. The actor network outputs the mean and variance, and simultaneously constructs a corresponding Gaussian distribution and samples actions. Finally, the actions are input into the environment for execution. After execution, the environment outputs the state information for the next time step. The agent's critic network calculates the corresponding value estimate based on the current state and the state at the next time step. It further calculates the mean squared error based on the real-time reward feedback from the environment and uses it to update the agent's critic network. Finally, it updates the actor network based on the output action information.

7. The grid intraday dispatch optimization method considering the uncertainty of photovoltaic output under severe weather conditions as described in claim 1, characterized in that, In step 4, transfer learning is used to improve the optimization of the scheduling agent for a specific day: Network-based transfer learning is used, which involves reusing a pre-trained network in the source domain, including its network structure and connection parameters, and transforming it into parameters for initial training of a deep neural network in the target domain. The specific formula is shown below: In the formula These represent the policy networks for the source task and the target task, respectively. Network parameters representing the source task policy network, The initial network parameters represent the target task. Therefore, the parameters of the source task's policy network are directly transferred to the target task's policy network, and then learning continues.

Images