A power grid random scenario generation method combining discrete probability and CGAN

Abstract

The application discloses a power grid random scenario generation method combining discrete probability and conditional generative adversarial network, first, a discrete joint probability model of the equivalent injection power time fluctuation and the prediction random error of each source and load node is constructed. Secondly, a conditional generative adversarial network (CGAN) model is constructed and trained, and the CGAN generator and discriminator are trained until convergence. Finally, based on the trained CGAN, the equivalent injection power prediction information of each source and load node in the research time window is taken as the conditional value of the CGAN input, random noise is generated based on the discrete joint probability distribution of each source and load node, the equivalent injection power of each source and load node in the research time window is generated through the CGAN, and the corresponding power grid random scenario of the research time window is constructed. The application proposes a power grid future random scenario generation method based on discrete probability model and conditional generative adversarial learning fusion, so as to enhance the pertinence and adaptability of scenario generation, and lay a foundation for formulating a robust scheduling strategy suitable for multiple scenarios.

Description

Technical Field

[0001] This invention belongs to the field of power system automation, and in particular relates to an intelligent construction method for power grid operation scenarios that integrates discrete probability and CGAN. Background Technology

[0002] The integration of high-penetration renewable energy sources into the grid will pose significant challenges to the power grid. Current grid operation, dispatch, and planning strategies need to be adjusted accordingly to adapt to the uncertainties of renewable energy. New energy power generation, represented by wind and solar power, is affected by the randomness and volatility of primary energy sources, resulting in random and fluctuating power output.

[0003] In existing research, scenario analysis can effectively characterize the uncertainty of renewable energy output through a set of discrete scenarios, thereby transforming the uncertainty optimization problem into a multi-scenario deterministic optimization problem or a stochastic optimization problem. However, due to the complexity and unknown nature of the uncertainty of renewable energy, especially when a high proportion of new energy is connected to the grid on a large scale, it leads to problems such as complex and diverse grid operation modes, difficulty in prediction, and difficulty in planning. Scenario construction also suffers from problems such as difficulty in covering historical scenarios and low scenario generation efficiency.

[0004] In recent years, with the rapid development of artificial intelligence technology and the continuous improvement of computer efficiency, it has become possible to analyze historical data through data-driven methods to obtain its inherent statistical regularities and to construct uncertainty models for renewable energy using black-box models. Artificial intelligence methods have advantages such as strong processing capabilities for high-dimensional nonlinear data, no need for complex physical models, and fast online computation speed, making them a research hotspot in the field of power grid security analysis. They have achieved good application results in areas such as power system mode analysis, transient stability, operating rules, network analysis and optimization, and dispatching operations. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for generating random power grid scenarios by fusing discrete probability and CGAN, thereby solving problems such as the complexity and diversity of power grid operation modes, the difficulty of planning and scheduling, and the low efficiency of scenario generation.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] In a first aspect, the present invention provides a method for generating stochastic power grid scenarios based on a discrete probability model of node equivalent injected power and a fusion of conditional generative adversarial learning, comprising the following steps:

[0008] Step 1: Obtain historical data of new energy sources and loads at source-load nodes, including historical datasets of predicted power and historical datasets of actual power, to form historical datasets of predicted power and actual power of equivalent injected power at nodes. Based on the historical data, statistically analyze the time-series volatility and prediction random error of equivalent injected power at each source-load node of the power grid, and construct a discrete joint probability model of the time-series volatility and prediction random error of equivalent injected power at each source-load node.

[0009] Step 2: Construct a Conditional Generative Adversarial Network (CGAN) model, including a generator and a discriminator neural network architecture. Use the historical predicted power of the equivalent injected power of each source load node as the conditional value. Randomly sample noise based on the discrete joint probability distribution of the equivalent injected power. Train the CGAN generator and discriminator until convergence to obtain the trained CGAN generator network model and discriminator network model.

[0010] Step 3: Obtain the power prediction data of new energy sources and loads at the source-load nodes corresponding to the research time window, and form a prediction dataset of the equivalent injected power of the nodes; based on the trained CGAN, use the predicted information of the equivalent injected power of each source-load node corresponding to the research time window as the conditional value of the CGAN input, generate random noise based on the discrete joint probability distribution of each source-load node, generate the equivalent injected power of each source-load node in the research time window through CGAN, and construct the random power grid scenario corresponding to the research time window.

[0011] Further, in step 1, the predicted power historical dataset and the actual power historical dataset of the new energy and load at the source-load nodes are obtained to form the predicted power historical dataset and the actual power historical dataset of the node equivalent injected power. Based on the temporal volatility and prediction random error of the node equivalent injected power, a discrete probability joint distribution function of the node equivalent injected power is constructed, including:

[0012] Historical datasets of predicted power and actual power of renewable energy and load at source and load nodes are obtained from the renewable energy prediction and load prediction database of the power grid EMS energy management system, forming historical datasets of predicted power and actual power of node equivalent injected power.

[0013]

[0014]

[0015] Among them, P a，i (t) represents the equivalent actual power of node i at time t. Let be the wind power output of node i at time t. Let the photovoltaic output of node i at time t be... P represents the load power of node i at time t. f，i(t) represents the equivalent predicted power of node i at time t. Let be the predicted wind power at node i at time t. Let be the predicted photovoltaic power of node i at time t. The predicted load power of node i at time t;

[0016] The temporal volatility and prediction random error of the equivalent injected power of each source load node are statistically analyzed, and a temporal volatility model and a discrete probability model of the prediction error of the equivalent injected power of each source load node are constructed.

[0017] The discrete probability joint distribution function is obtained by combining the time-series fluctuation model of node equivalent injection power based on historical data statistical analysis and the discrete probability model of prediction error.

[0018] Furthermore, in step 1, the time-series fluctuation model of the node equivalent injection power based on historical data statistical analysis and the discrete probability model of the prediction error are combined to obtain the discrete probability joint distribution function, including:

[0019] The historical temporal volatility and prediction random error of the equivalent injected power of each source load node are statistically analyzed, and the equivalent injected power t of each source load node is constructed. i Time (t) i The time-series fluctuation model ∈[t1, t2]) and the discrete probability model for predicting random errors;

[0020] Combined with t i By calculating the discrete probability joint distribution of the injected power of each node at time t, considering the historical temporal fluctuations and the discrete probability of the predicted random error, we obtain the discrete joint probability distribution of the injected power of the node at that time.

[0021] Furthermore, combined with t i The discrete probability model of the historical temporal fluctuation and prediction random error of the equivalent injected power of each source load node at time t is used to calculate the discrete joint probability distribution of the equivalent injected power of each node at that time, including:

[0022] The temporal fluctuation characteristics of the equivalent injected power at node i at time t can be expressed as:

[0023] ΔP i (t)=P a，i (t)-P a，i (t-1) (3)

[0024] The prediction error of the equivalent injected power at node i at this moment can be expressed as:

[0025] ε i (t)=P a，i (t)-p f，i (t) (4)

[0026] Based on formulas (3) and (4), the steps for constructing the discrete probability model of a node are as follows:

[0027] As can be seen from formula (3), the volatility is the equivalent injected power at time t+1 minus the equivalent injected power at time t; take the discretization formula factor C, divide the volatility value of the node equivalent injected power into its range of variation, and count the frequency of the volatility value in each discretization interval as the estimated value of the probability of the volatility value falling into the discretization interval, and calculate the expected value of all volatility values ​​falling into each interval.

[0028] Use (2×N) w The matrix G w G is a discrete probability distribution characterizing the temporal volatility of the equivalent injected power at nodes in historical data. w The first line represents the fluctuation range, and the second line represents the probability of falling within that fluctuation range, G. w It can be represented as:

[0029]

[0030] In the formula N w Let be the number of states of the discrete probability distribution of time-series volatility, satisfying:

[0031]

[0032] In the formula δ w，max The maximum value of the equivalent power fluctuation injected into the nodes in the historical data; Indicates to The value is rounded down;

[0033] The equivalent actual power of node i at k time points in the statistical historical data is given by δ. w (t) represents the fluctuation value in hour t, G w The elements in can be calculated using the following formula:

[0034]

[0035]

[0036] Note: k fluctuation values ​​can be calculated from k time periods; where f i (x) is the indicator function, defined as follows:

[0037]

[0038] According to the above formulas (6)(7)(8)(9), the (k-1) fluctuation values ​​of k time period nodes i can be converted into N... wThe discrete probability distribution of each state value and its corresponding probability, i.e., formula (5), retains the probabilistic characteristics of temporal fluctuations while merging states.

[0039] The prediction error probability is divided based on the Δt between the predicted time and the current time. The error of the corresponding time period is calculated according to formula (4). The error is divided into regions within its range of variation. The frequency of error values ​​in each discretized interval is statistically analyzed as the estimated value of the probability of the error value falling into the discretized interval. The expected value of all error values ​​falling into each interval is calculated to obtain its discrete probability distribution.

[0040] The equivalent injection power of node i in k time periods of statistical historical data is obtained, and k error values ​​are converted into N. w Discrete probability distribution of a state value and its corresponding probability;

[0041] For discrete random variables, the joint probability function is P(X=x&Y=y), that is:

[0042] P(Y=y|X=x)P(X=x)=P(X=x|Y=y)P(Y=y) (10)

[0043] Joint distribution function:

[0044]

[0045] Using formulas (10) and (11), the joint distribution function of volatility and random error can be obtained.

[0046] By using the obtained discrete joint distribution function, the fluctuation range of the equivalent injected power of each source load node at each time step can be determined. Combined with the corresponding node power prediction information, the possible random change trend of each source load node based on historical actual data can be obtained.

[0047] Further, in step 2, a conditional generative adversarial network (CGAN) model is constructed, including a generator and a discriminator neural network architecture. The historical predicted power of the equivalent injected power of each source load node is used as a conditional value. Noise is generated by random sampling based on the discrete joint probability distribution of the equivalent injected power. The CGAN generator and discriminator are then trained until convergence, resulting in the trained CGAN generator and discriminator network models, including:

[0048] A conditional generative adversarial network (CGAN) model is constructed. The generator's input and hidden layers each contain fully connected layers, batch normalization (BN) layers, and activation functions, with 256 and 512 neurons respectively. The output layer contains a fully connected layer and a Tanh function. The discriminator's input layer contains a fully connected layer, BN layers, and activation functions, with 256 neurons. The output layer contains a fully connected layer and a Sigmond function.

[0049] Based on the historical dataset of predicted power equivalent injected power of the power grid source and load nodes, the time-series horizontal concatenation of the predicted power equivalent injected power of each source and load node is used as the conditional value C of the CGAN input. Random numbers are generated based on the discrete joint probability distribution obtained for each node, and the corresponding time-series horizontal concatenation forms the noise Z input sequence. The above conditional value C and noise Z are vertically concatenated as the input of the generator. The historical scene node equivalent injected power is used as the real sample X input. The above real sample X and noise Z are vertically concatenated as the input of the generator to train the conditional generative adversarial network model CGAN.

[0050] Gradient penalty sampling points are selected, the loss functions of the generator and discriminator are calculated, and the Adam optimization algorithm is used to update the network weights of the generator and discriminator respectively.

[0051] If training is not completed, return to the next round of training. Training ends when the loss functions of both the generator and the discriminator converge.

[0052] Furthermore, training ends when the loss functions of both the generator and the discriminator converge.

[0053] Furthermore, the temporal horizontal concatenation of the equivalent injected power prediction information of each source load node is used as the conditional value C for the CGAN input. Random numbers are generated based on the discrete joint probability distribution obtained for each node, and the corresponding temporal horizontal concatenation forms the noise input sequence Z. The conditional value C and the noise Z are vertically concatenated as the generator input. The equivalent injected power of historical scene nodes is used as the real sample X input. The real sample X and the noise Z are vertically concatenated as the generator input. The Conditional Generative Adversarial Network (CGAN) model is trained, including:

[0054] For each moment, the predicted equivalent injected power of each source-load node is matched with a noise. Within the set research time window [t1, t2], there are m moments, Z... i，j and C i，j Let Z represent the random sampled noise input and the conditional input at time j of node i, respectively. Then the noise input sequence set of node i is Z. i ={Z i，1 Z i，2 , ..., Z i，m-1 Z i，m The conditional input sequence set is C. i ={C i，1 C i，2 C i，m-1 C i，m The noise input length is the same as the condition input length. The noise input z and the condition input c are vertically concatenated into a matrix, so that the noise and the prediction information form a vertical correspondence at each time step.

[0055] The true sample input sequence set for historical data node i is X i ={X i，1 X i，2 , ..., X i，m-1 X i，m The input length of the historical data real sample is the same as the condition input length. The historical data real sample X input and the condition input c are vertically concatenated into a matrix, so that the historical data real sample and the prediction information form a vertical correspondence at each time step.

[0056] The sample input sequence set is X′ i ={X′ i，1 , X′ i，2 , ..., X′ i，m-1 , X′ i，m The length of the generated sample input is the same as the length of the conditional input. The generated sample X′ input and the conditional input c are vertically concatenated into a matrix, so that the generated sample and the prediction information form a vertical correspondence at each time step.

[0057] The data after vertically concatenating the noise Z and the predicted value C is input into the generator G, which outputs the generated sample X′. The data after vertically concatenating the predicted value C and the real sample X is input into the discriminator D, and the data after vertically concatenating the predicted value C and the generated sample X′ is also input into the discriminator D. The discriminator D outputs the discrimination value between the real sample and the generated sample.

[0058] Furthermore, the method also includes:

[0059] The generated scenario sets were evaluated from three aspects: volatility, time correlation, and effectiveness, and the selected scenario sets were then selected.

[0060] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0061] 1. This invention discloses an intelligent construction method for power grid operation scenarios that takes into account multiple uncertainties, belonging to the field of power system automation. Addressing the problems of complex and diverse power grid operation modes and difficulties in planning caused by the large-scale integration of high-proportion renewable energy sources into the grid, this invention proposes a method for generating random power grid scenarios based on a fusion of discrete probability models and conditional generative adversarial learning, to enhance the relevance of sample generation and scenario adaptability.

[0062] 2. The specific function of the generator in this invention is to input historical data and output corresponding scene extension scenes, which largely solves the problem of insufficient scenes. Attached Figure Description

[0063] Figure 1 This is a flowchart of this application. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0065] Example 1:

[0066] To address the challenges posed by the large-scale integration of high-proportion renewable energy sources into the power grid, leading to complex and diverse grid operation modes and difficulties in planning, this embodiment proposes a method for generating stochastic grid scenarios based on a fusion of a discrete probability model of power injection nodes and conditional generative adversarial learning. The method includes the following steps:

[0067] Step 1: Obtain historical data of power injected into power nodes at grid source and load nodes from the grid EMS energy management system, perform statistical analysis on the fluctuation characteristics and prediction random errors of the node power injected, and calculate the time-series fluctuation model and the discrete probability model of the prediction error of the power injected into grid source and load nodes based on the statistical analysis of historical data.

[0068] Step 1.1: For the time period information [t1, t2] corresponding to the power grid scenario to be generated, obtain the source-load power prediction data of the power grid bus node from the new energy prediction and load prediction software of the power grid EMS energy management system, form the node injection power prediction information, and input it into the discrete probability model mentioned in Step 1 to calculate t1 and t2 respectively. i Time (t) i ∈[t1, t2]), the time-series discrete probability model of node injection power fluctuation and the random discrete probability model of prediction error.

[0069] Step 1.2: Combine t i By using the historical time-series fluctuation model of the injected power at a given time point and the discrete probability model of the prediction error, the joint discrete probability distribution function of the injected power at that time point is calculated, and the discrete probability distribution of the random prediction value is obtained.

[0070] Step 2: Construct a Conditional Generative Adversarial Network (CGAN) model. Divide the power source-load prediction dataset of the power grid bus node mentioned in Step 1.1 into a training set and a test set. Use the node injection power prediction information in the [t1, t2] time period in the training set as the conditional value C for the CGAN input. Based on the discrete probability joint distribution function obtained in Step 1.2, generate random numbers and form a noise input sequence Z. Vertically concatenate the above conditional value C and noise Z as the input of the generator. Use the historical data of the power grid source-load injection node as the real sample X for the CGAN input. Vertically concatenate the above conditional value C and real sample X as the input of the discriminator.

[0071] Gradient penalty sampling points are selected, and the loss functions of the generator and discriminator are calculated. The Adam optimization algorithm is then used to update the network weights of the generator and discriminator, respectively. Training ends when the loss function graphs of the generator and discriminator converge. If training is not yet complete, the process returns to the next round of training.

[0072] Step 3: After training, extract the generator model from CGAN, select the power prediction information of the nodes in the corresponding time period of the test set as the condition value C of the CGAN input, and concatenate it longitudinally with the discrete probability distribution noise Z that satisfies the random prediction value. Input the concatenated data into the trained generator G to generate a scene dataset based on the prediction value.

[0073] Specifically, the technical solution adopted in this invention is as follows:

[0074] First, a discrete random probability model for new energy nodes and load nodes is constructed based on historical data statistical analysis. The time-series fluctuation discrete random probability model can be expressed as:

[0075]

[0076] The discrete random probability model of prediction error can be expressed as:

[0077]

[0078] in This represents the predicted power value of power injection node i in time period t. This represents the actual power injected into node i during time period t.

[0079] Based on formulas (1) and (2), the steps for constructing the discrete random probability model for new energy nodes are as follows:

[0080] As shown in formula (1), the volatility is the injected power at time t+1 minus the injected power at time t. Taking the discretization formula factor C, the volatility value of the node injected power is divided into partitions within its range of variation. The frequency of the volatility value in each discretization interval is statistically analyzed as the estimated value of the probability that the volatility value falls into the discretization interval. The expected value of all volatility values ​​falling into each interval is then calculated.

[0081] Use 1 (2×N) w The matrix G w G represents the discrete probability distribution of the bus load volatility over the study period. w The first line represents the fluctuation value, i.e., the state value, and the second line represents the probability corresponding to the fluctuation value, G. w It can be represented as:

[0082]

[0083] In the formula Nw Let be the number of states of a discrete probability distribution, satisfying:

[0084]

[0085] In the formula δ w，max To study the maximum value of fluctuations within a period; Indicates to The value is rounded down, and C is the discretization formula factor;

[0086] The equivalent actual power of node i at k time points in the statistical historical data is given by δ. w (t) represents the fluctuation value in hour t, G w The elements in can be calculated using the following formula:

[0087]

[0088]

[0089] Note: k time points can yield k-1 fluctuation values. In the formula, f... i (x) is the indicator function, defined as follows:

[0090]

[0091] Based on the above formulas (4)(5)(6)(7), the k-1 fluctuation values ​​of the bus load at time node k can be converted into N... w The discrete probability distribution of each state value and its corresponding probability, i.e., formula (3), retains the probabilistic characteristics of fluctuation while merging the states.

[0092] Based on the Δt between the predicted time and the current time, the probability of prediction error is divided. The error of the corresponding time period is calculated according to formula (2). The error is divided into regions within its range of variation. The frequency of error values ​​in each discretized interval is statistically analyzed as the estimated value of the probability of the error value falling into the discretized interval. The expected value of all error values ​​falling into each interval is calculated to obtain its discrete probability distribution.

[0093] The equivalent injection power of node i in k time periods of statistical historical data is obtained, and k error values ​​are converted into N. w Discrete probability distribution of each state value and its corresponding probability.

[0094] For discrete random variables, the joint probability function is P(X=x&Y=y), that is:

[0095] P(Y=y|X=x)P(X=x)=P(X=x|Y=y)P(Y=y) (8)

[0096] Observing the formula, we can see that it is a probability function of X and Y occurring simultaneously. Therefore, its (joint) distribution function is simply the sum of these probability functions. Similarly, the joint distribution function is:

[0097]

[0098] Using formulas (8) and (9), the joint distribution function of volatility and random error can be obtained.

[0099] The obtained joint distribution function can be used to determine the range of variation for each time point based on historical data. By feeding the historical data into the aforementioned probability model, a random prediction value that accounts for volatility and error and varies around the historical data can be obtained.

[0100] Secondly, the power prediction information of the nodes in the [t_1, t_2] time period in the training set is used as the conditional value C of the CGAN input. Random numbers are generated based on the discrete probability joint distribution function obtained in 1.2 to form the noise Z input sequence. The conditional value C and the noise Z are vertically concatenated as the input of the generator. The historical data of the power grid source-load injection nodes are used as the real sample X of the CGAN input. The conditional value C and the real sample X are vertically concatenated as the input of the discriminator.

[0101] Assuming there are m time points within a given prediction time interval, Z i，j and C i，j Let Z represent the random sampled noise input and the conditional input at time j of node i, respectively. Then the noise input sequence set of node i is Z. i ={Z i，1 Z i，2 , ..., Z i，m-1 Z i，m The conditional input sequence set is C. i ={C i，1 C i，2 C i，m-1 C i，m For each time step, a noise input is matched to the predicted value. The noise input length is the same as the conditional input length. The noise input z and the conditional input c are concatenated vertically into a matrix, so that the noise and the predicted information form a vertical correspondence at each time step. Similarly, the input data processing for the historical cross-sectional data real sample X and the predicted data conditional value C is the same as the input data processing for the predicted data conditional value C and the generated sample X′.

[0102] The data obtained by vertically concatenating the noise Z and the predicted value C is input into the generator G, which outputs the generated sample X′. The data obtained by vertically concatenating the predicted value C and the real sample X′ is input into the discriminator D, and the data obtained by vertically concatenating the predicted value C and the generated sample X′ is also input into the discriminator D. The discriminator D outputs the discrimination value between the real sample and the generated sample.

[0103] Third, select gradient penalty sampling points, calculate the loss functions of the generator and discriminator, and use the Adam optimization algorithm to update the network weights of the generator and discriminator respectively. If training is not yet complete, return to the next round of training.

[0104] Finally, training ends when the loss functions of both the discriminator and the generator converge. The generator model is extracted from the Conditional Generative Adversarial Network (CGAN), and the predicted data of the power grid source-load injection nodes in the test set are selected as condition C and longitudinally concatenated with random noise Z. The concatenated data is then input into the trained generator G to generate a scene dataset based on the predicted values.

[0105] In summary, the proposed method first involves statistically analyzing the power injection volatility and prediction random error of power injection nodes based on historical data, thereby constructing a discrete stochastic probability model for the power injection nodes. Then, a Conditional Generative Adversarial Network (CGAN) model is built using PyTorch. Random sampling is performed based on the discrete stochastic probability model of the injection nodes, and the sampling results are used as noise input to the CGAN, while historical actual operation scenario data are used as the true input values. Finally, the generated scenario set is evaluated from three aspects: volatility, temporal correlation, and effectiveness. Simulation results verify the effectiveness of the proposed intelligent construction method for power grid operation scenarios.

[0106] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0107] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0108] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0109] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0110] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating stochastic power grid scenarios based on a discrete probability model of equivalent injected power at source-load nodes and conditional generative adversarial learning, characterized in that, Includes the following steps: Historical data on renewable energy and load at source-load nodes are obtained, including historical datasets of predicted power and historical datasets of actual power. Historical datasets of predicted power and actual power are formed for the equivalent injected power of nodes. Based on the historical data, the temporal volatility and prediction random error of the equivalent injected power of each source-load node in the power grid are statistically analyzed, and a discrete joint probability model of the temporal volatility and prediction random error of the equivalent injected power of each source-load node is constructed. A conditional generative adversarial network (CGAN) model is constructed, including a generator and a discriminator neural network architecture. The historical predicted power of the equivalent injection power of each source load node is used as the condition value. Noise is generated by random sampling based on the discrete joint probability distribution of the equivalent injection power. The CGAN generator and discriminator are trained until convergence, and the trained CGAN generator network model and discriminator network model are obtained. The power prediction data of new energy sources and loads at the source-load nodes corresponding to the research time window are obtained to form a prediction dataset of the equivalent injected power of the nodes. Based on the trained CGAN, the prediction information of the equivalent injected power of each source-load node corresponding to the research time window is used as the condition value of the CGAN input. Random noise is generated based on the discrete joint probability distribution of each source-load node. The equivalent injected power of each source-load node in the research time window is generated by CGAN, and the random scenario of the power grid corresponding to the research time window is constructed.

2. The method for generating random power grid scenarios according to claim 1, characterized in that, Obtain historical datasets of predicted and actual power for new energy sources and loads at source-load nodes, forming historical datasets of predicted and actual power for node equivalent injected power. Based on the temporal volatility and prediction random error of node equivalent injected power, construct a joint discrete probability distribution function for node equivalent injected power, including: The predicted power historical datasets and actual power historical datasets of new energy and load at source and load nodes are obtained from the new energy prediction and load prediction database of the power grid EMS energy management system, forming the predicted power historical datasets and actual power historical datasets of the node equivalent injected power. Among them, P a，i (t) represents the equivalent actual power of node i at time t. Let be the wind power output of node i at time t. Let the photovoltaic output of node i at time t be... P represents the load power of node i at time t. f，i (t) represents the equivalent predicted power of node i at time t. Let be the predicted wind power at node i at time t. Let be the predicted photovoltaic power of node i at time t. The predicted load power of node i at time t; The temporal volatility and prediction random error of the equivalent injected power of each source load node are statistically analyzed, and a temporal volatility model and a discrete probability model of the prediction error of the equivalent injected power of each source load node are constructed. The discrete probability joint distribution function is obtained by combining the time-series fluctuation model of node equivalent injection power based on historical data statistical analysis and the discrete probability model of prediction error.

3. The method for generating random power grid scenarios according to claim 2, characterized in that, By combining the time-series fluctuation model of node equivalent injection power based on historical data statistical analysis with the discrete probability model of prediction error, a joint discrete probability distribution function is obtained, including: The historical temporal volatility and prediction random error of the equivalent injected power of each source load node are statistically analyzed, and the equivalent injected power t of each source load node is constructed. i Time (t) i The time-series fluctuation model ∈[t1, t2]) and the discrete probability model for predicting random errors; Combined with t i By calculating the discrete probability joint distribution of the injected power of each node at time t, considering the historical temporal fluctuations and the discrete probability of the predicted random error, we obtain the discrete joint probability distribution of the injected power of the node at that time.

4. The method for generating random power grid scenarios according to claim 3, characterized in that, Combined with t i The discrete probability model of the historical temporal fluctuation and prediction random error of the equivalent injected power of each source load node at time t is used to calculate the discrete joint probability distribution of the equivalent injected power of each node at that time, including: The temporal fluctuation characteristics of the equivalent injected power at node i at time t can be expressed as: ΔP i (t)=P a，i (t)-P a，i (t-1) (3) The prediction error of the equivalent injected power at node i at this moment can be expressed as: ε i (t)=P a，i (t)-p f，i (t) (4) Based on formulas (3) and (4), the steps for constructing the discrete probability model of a node are as follows: As can be seen from formula (3), the volatility is the equivalent injected power at time t+1 minus the equivalent injected power at time t; take the discretization formula factor C, divide the volatility value of the node equivalent injected power into its range of variation, and count the frequency of the volatility value in each discretization interval as the estimated value of the probability of the volatility value falling into the discretization interval, and calculate the expected value of all volatility values ​​falling into each interval. Use (2×N) w The matrix G G G is a discrete probability distribution characterizing the temporal volatility of the equivalent injected power at nodes in historical data. w The first line represents the fluctuation range, and the second line represents the probability of falling within that fluctuation range, G. w It can be represented as: In the formula N w Let be the number of states of the discrete probability distribution of time-series volatility, satisfying: In the formula δ w，max The maximum value of the equivalent power fluctuation injected into the nodes in the historical data; Indicates to The value is rounded down; The equivalent actual power of node i at k time points in the statistical historical data is given by δ. w (t) represents the fluctuation value in hour t, G w The elements in can be calculated using the following formula: Note: k fluctuation values ​​can be calculated from k time periods; where f i (x) is the indicator function, defined as follows: According to the above formulas (6)(7)(8)(9), the (k-1) fluctuation values ​​of k time period nodes i can be converted into N... w The discrete probability distribution of each state value and its corresponding probability, i.e., formula (5), retains the probabilistic characteristics of temporal fluctuations while merging states. The prediction error probability is divided based on the Δt between the predicted time and the current time. The error of the corresponding time period is calculated according to formula (4). The error is divided into regions within its range of variation. The frequency of error values ​​in each discretized interval is statistically analyzed as the estimated value of the probability of the error value falling into the discretized interval. The expected value of all error values ​​falling into each interval is calculated to obtain its discrete probability distribution. The equivalent injection power of node i in k time periods of historical data is statistically analyzed, and k error values ​​are converted into N values. w Discrete probability distribution of a state value and its corresponding probability; For discrete random variables, the joint probability function is P(X=x&Y=y), that is: P(Y=y|X=x)P(X=x)=P(X=x|Y=y)P(Y=y) (10) Joint distribution function: Using formulas (10) and (11), the joint distribution function of volatility and random error can be obtained. By using the obtained discrete joint distribution function, the fluctuation range of the equivalent injected power of each source load node at each time step can be determined. Combined with the corresponding node power prediction information, the possible random change trend of each source load node based on historical actual data can be obtained.

5. The method for generating random power grid scenarios according to claim 1, characterized in that, A conditional generative adversarial network (CGAN) model is constructed, including a generator and a discriminator neural network architecture. The historical predicted power of the equivalent injected power of each source load node is used as a conditional value. Noise is generated by random sampling based on the discrete joint probability distribution of the equivalent injected power. The CGAN generator and discriminator are trained until convergence, resulting in the trained CGAN generator and discriminator network models, including: A conditional generative adversarial network (CGAN) model was constructed. The generator's input layer and hidden layer both contain fully connected layers, batch normalization (BN) layers, and activation functions, with 256 and 512 neurons respectively. The output layer contains a fully connected layer and a Tanh function. The discriminator's input layer contains a fully connected layer, BN layers, and activation functions, with 256 neurons. The output layer contains a fully connected layer and a Sigmond function. Based on the historical dataset of predicted power equivalent injected power of the power grid source and load nodes, the time-series horizontal concatenation of the predicted power equivalent injected power of each source and load node is used as the conditional value C of the CGAN input. Random numbers are generated based on the discrete joint probability distribution obtained for each node, and the corresponding time-series horizontal concatenation forms the noise Z input sequence. The above conditional value C and noise Z are vertically concatenated as the input of the generator. The historical scene node equivalent injected power is used as the real sample X input. The above real sample X and noise Z are vertically concatenated as the input of the generator to train the conditional generative adversarial network model CGAN. Gradient penalty sampling points are selected, the loss functions of the generator and discriminator are calculated, and the Adam optimization algorithm is used to update the network weights of the generator and discriminator respectively. If training is not completed, return to the next round of training. Training ends when the loss functions of both the generator and the discriminator converge.

6. The method for generating random power grid scenarios according to claim 5, characterized in that, Training ends when the loss functions of both the generator and the discriminator converge.

7. The method for generating random power grid scenarios according to claim 5, characterized in that, The equivalent injected power prediction information of each source load node is temporally concatenated laterally as the conditional value C input to the CGAN. Random numbers are generated based on the discrete joint probability distribution obtained for each node, and the corresponding temporally concatenated sequence forms the noise input sequence Z. The conditional value C and the noise Z are vertically concatenated as the generator input. The equivalent injected power of historical scene nodes is used as the real sample X input. The real sample X and the noise Z are vertically concatenated as the generator input. The Conditional Generative Adversarial Network (CGAN) model is trained, including: For each moment, the predicted equivalent injected power of each source-load node is matched with a noise. Within the set research time window [t1, t2], there are m moments, Z... i，j and C i，j Let Z represent the random sampled noise input and the conditional input at time j of node i, respectively. Then the noise input sequence set of node i is Z. i ={Z i，1 Z i，2 , ..., Z i，m-1 Z i，m The conditional input sequence set is C. i ={C i，1 C i，2 C i，m-1 C i，m The noise input length is the same as the condition input length. The noise input z and the condition input c are vertically concatenated into a matrix, so that the noise and the prediction information form a vertical correspondence at each time step. The true sample input sequence set for historical data node i is X i ={X i，1 X i，2 , ..., X i，m-1 X i，m The input length of the historical data real sample is the same as the condition input length. The historical data real sample X input and the condition input c are vertically concatenated into a matrix, so that the historical data real sample and the prediction information form a vertical correspondence at each time step. The sample input sequence set is X′ i ={X′ i，1 , X′ i，2 , ..., X′ i，m-1 , X′ i，m The length of the generated sample input is the same as the length of the conditional input. The generated sample X′ input and the conditional input c are vertically concatenated into a matrix, so that the generated sample and the prediction information form a vertical correspondence at each time step. The data obtained by vertically concatenating the noise Z and the predicted value C is input into the generator G, which outputs the generated sample X′. The data obtained by vertically concatenating the predicted value C and the real sample X′ is input into the discriminator D. The data obtained by vertically concatenating the predicted value C and the generated sample X′ is also input into the discriminator D. The discriminator D outputs the discrimination value between the real sample and the generated sample.

Images