Zone area two-stage distribution robust line loss optimization method considering wind and light uncertainty

Abstract

The invention provides a transformer area two-stage distribution robust line loss optimization method considering wind and light uncertainty, and belongs to the technical field of new energy. Firstly, a day-ahead-real-time two-stage line loss optimization method is constructed, and in the day-ahead stage, a K-means clustering algorithm and a conditional generative adversarial network are combined to generate a multi-scene sample set capable of reflecting the uncertainty of wind and light output; in a real-time stage, based on a day-ahead optimization result and updated output information, a multi-scene distribution robust optimization model is utilized to realize joint correction of line loss and economic dispatching; in the optimization solution process, a forward-backward substitution method is adopted to embed the load flow calculation model so as to accurately reflect the line transmission loss of the transformer area power distribution network; secondly, a distributed robust optimization method based on scene generation is constructed, historical wind and light data are utilized, a diversified scene set is generated through K-means clustering and a conditional generative adversarial network, and distributed robust optimization planning and correction are carried out in different scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy, and particularly relates to a two-stage distribution robust line loss optimization method for a transformer area considering wind and light uncertainty. BACKGROUND

[0002] At present, the penetration rate of distributed wind power, photovoltaic and other renewable energy in the transformer area power distribution network is continuously improved. The randomness and volatility of wind and light output lead to the uncertainty of power flow distribution and the complexity of line loss mechanism in the transformer area power distribution network. At present, there mainly exist line loss optimization methods based on deterministic optimal power flow and line loss optimization methods based on probabilistic power flow and fuzzy calculation. However, the traditional two kinds of line loss calculation and optimization methods all face severe challenges, which are specifically as follows:

[0003] 1. The line loss optimization method based on deterministic optimal power flow. The technical core of the method is that the wind and light output and load demand in a period of time (such as day-ahead) in the future are assumed to be determined and known (usually point prediction value), and a deterministic optimal power flow model is established on this basis. Linear programming or interior point method and other optimization algorithms are used to solve the model, and the optimal scheduling plan of distributed power sources, energy storage and other resources is obtained. The fundamental defect of the method is the "deterministic assumption". It completely ignores the inherent uncertainty of wind and light output, and the optimization result is based on a single, possibly inaccurate prediction scenario. Once the actual wind and light output deviates from the predicted value, the optimization scheduling scheme made according to the method will no longer be the optimal solution, and may even lead to the problem that the system line loss does not decrease but increases, the node voltage exceeds the limit and other safety problems. The optimization result has poor robustness, weak anti-interference ability, and it is difficult to maintain the expected economic and safety benefits in the actual fluctuating environment.

[0004] 2、Based on the probabilistic power flow and fuzzy calculation line loss optimization method, the uncertainty of wind and light output is processed by probability statistics or fuzzy optimization method. The wind and light output is described as a random variable or a fuzzy variable: the random method assumes that it obeys a certain probability distribution, and the probability characteristics of system state and line loss are obtained by Monte Carlo simulation probability power flow calculation; the fuzzy method describes the output uncertainty through membership function, and obtains the fuzzy distribution of system line loss by fuzzy power flow calculation. On this basis, the optimization model is established to minimize the expected line loss or the worst scenario probability, and the scheduling scheme is solved. The disadvantage of this method is that the optimization effect is seriously dependent on the subjective assumption of the uncertainty model. In the probability optimization method, the effectiveness of the solution is completely based on the prior assumption that the wind and light output obeys a certain probability distribution, and it is difficult to obtain an accurate probability distribution model in actual engineering. The optimization based on the wrong distribution assumption will lead to a significant decline in the actual performance of the scheduling scheme. In the fuzzy optimization method, the optimization result is significantly affected by the shape and parameters of the fuzzy membership function set by human beings. Different settings will get very different optimization schemes, and there is a lack of objective and unified evaluation standard.

[0005] Under this background, on the one hand, the uncertainty of distributed power output significantly increases the difficulty of line loss calculation; on the other hand, the existing optimization method is difficult to effectively deal with various possible output fluctuation scenarios while ensuring economy. Therefore, how to realize the accurate calculation and resource optimization scheduling of the line loss of the distribution network of the transformer area under the premise of fully considering the uncertainty of the distributed energy output has become a key technical problem to be solved in the field of transformer area distribution network at present. SUMMARY

[0006] The purpose of the present application is to provide a two-stage distribution robust line loss optimization method considering wind and light uncertainty, which aims to solve the key technical problems in the prior art that the line loss optimization scheme has poor adaptability and insufficient robustness in actual operation due to ignoring the uncertainty of wind and light output or relying on subjective probability assumption.

[0007] In order to achieve the above purpose, the technical scheme adopted by the present application is:

[0008] A two-stage distribution robust line loss optimization method considering wind and light uncertainty is provided. Firstly, a day-ahead-real-time two-stage line loss optimization method is constructed. In the day-ahead stage, a multi-scenario sample set reflecting wind and light output uncertainty is generated by combining the K-means clustering algorithm and the conditional generative adversarial network (CGAN). In the real-time stage, based on the day-ahead optimization result and the updated output information, a multi-scenario distribution robust optimization model is used to realize the joint correction of line loss and economic dispatch. In the optimization solving process, the forward-backward sweep method is embedded into the power flow calculation model to accurately reflect the line transmission loss of the distribution network. Secondly, a distribution robust optimization method based on scenario generation is constructed. A diversified scenario set is generated by using historical wind and light data through K-means clustering and conditional generative adversarial network, and distribution robust optimization planning and correction are carried out under different scenarios.

[0009] The two-stage distribution robust line loss optimization method considering wind and light uncertainty has the following detailed steps:

[0010] Step 1: Wind and light uncertainty scenario generation method based on conditional generative adversarial network

[0011] Step 1.1 Collect historical measured data and day-ahead prediction data of wind power and photovoltaic output as training sample set. The day-ahead prediction data is used as conditional label, and the historical measured wind and light power data is used as real sample.

[0012] Step 1.2 Construct an adversarial game network model containing generator and discriminator ; wherein the generator adopts a convolutional neural network structure, the input is random noise and day-ahead prediction conditional label , and the output is generated wind and light power scenario sample ; the discriminator also adopts a convolutional neural network structure, the input is real sample or generated sample and corresponding label , and its function is to judge whether the input sample is from real historical data or generated by the generator;

[0013] Step 1.3 Define the objective function of the conditional generative adversarial network, which is used to describe the minimax game between the generator and the discriminator. Based on the objective function, the model is trained by using the alternating optimization method. By using the trained generator, a large number of random noise and specific day-ahead prediction data are input to generate a large-scale wind and light output scenario set.

[0014] Step 1.4 verifies the accuracy of the time sequence correlation and uncertainty description of the generated scene, and introduces the autocorrelation coefficient and the partial autocorrelation coefficient correlation analysis is performed;

[0015] Step 1.5 obtains a set of scenes that meet the day-ahead prediction conditions and can fully reflect the uncertainty of wind and light output by training and evaluating the network; the K-means clustering algorithm is used to divide the high-quality scene set generated by CGAN into modes, and the sample allocation and center position are updated iteratively to finally obtain representative typical output modes, and each class of samples corresponds to different weather scene characteristics;

[0016] Step 2: Construct a day-ahead optimal scheduling model considering the line loss of the transformer area

[0017] Step 2.1 formulates a day-ahead scheduling plan based on the day-ahead prediction data of wind and light, with the minimum transformer area line loss and economic scheduling cost as the target;

[0018] Step 2.2 constructs a day-ahead optimal scheduling model considering the line loss of the transformer area, taking into account the operation characteristics of wind, light, energy storage, electric vehicles, and adjustable load, with the minimum transformer area line loss and economic scheduling cost as the objective function;

[0019] Step 2.3 establishes the system active / reactive power balance constraint in the day-ahead stage, the day-ahead stage output constraint of wind power and photovoltaic, the day-ahead stage operation constraint of energy storage device, the day-ahead stage operation constraint of electric vehicle (EV) cluster, transferable load (TL), and interruptible load;

[0020] Step 2.4 embeds the forward-backward sweep method into the optimal scheduling model for accurate solution of power distribution network power flow. After the power flow calculation converges, the system total theoretical line loss value is calculated combined with the branch current value ; which is fed back to the objective function, and the final solution of the model obtains a day-ahead resource scheduling scheme considering economy and low line loss;

[0021] Step 3: Construct a real-time distribution robust optimal scheduling model considering wind and light uncertainty

[0022] Step 3.1, based on the extreme wind and light scenarios obtained by K-means clustering in step 1 and the scheduling scheme in the day-ahead stage, considers wind power uncertainty in the real-time stage to dynamically correct the day-ahead decision; adopts the distribution robust optimization method to construct a probability uncertainty set with extreme scenarios as candidate distributions , and constructs an extreme value distribution through weights, so as to include the system risk under the worst scenario into the optimization target;

[0023] Step 3.2 Establish real-time system active / reactive power balance constraints, real-time output constraints for wind power and photovoltaics, real-time operation constraints for energy storage devices, real-time operation constraints for EV clusters, real-time operation constraints for transferable loads, and real-time operation constraints for interruptible loads.

[0024] Step 3.3 uses K-means clustering to obtain k extreme landscape scenes, each denoted as . (in In the sub-Bruker optimization framework, the worst-case distribution is considered, which is a discrete distribution composed of these extreme scenarios combined with certain probability weights.

[0025] Step 3.4 involves constructing a probabilistically uncertain set. Weights need to be adjusted. Two types of constraints are imposed: First, the weighted mean of all extreme scenarios should lie within a confidence interval centered on the historical data mean; second, the Wasserstein distance between this extreme value distribution and the empirical distribution must not exceed a given ambiguity parameter. ;

[0026] Step 3.5: The Blue Bar optimization model is based on the worst-case distribution. The goal is to find a scheduling strategy that minimizes the sum of expected operating costs and line losses. By solving this model, a robust real-time resource scheduling scheme that takes into account the uncertainties of wind and solar power can be obtained, along with the corresponding theoretical line loss value.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1) This invention proposes a two-stage line loss optimization method (day-ahead / real-time) for economic dispatch and operational optimization of distribution transformer areas. The method performs overall dispatch planning in the day-ahead stage and dynamically adjusts the method based on the latest load and output data in the real-time stage, achieving both line loss minimization and safety constraints. During the optimization process, a forward-backward substitution method is embedded for power flow calculation, accurately simulating bidirectional power flow characteristics, branch currents, and node voltage distribution. By coordinating various distributed resources such as wind power, photovoltaics, energy storage, electric vehicles, and adjustable loads, the method improves the operational efficiency and power supply reliability of the distribution network in the distribution area while ensuring system safety.

[0029] 2) The application constructs a distribution robust optimization method based on scenario generation, which is used to describe the volatility and uncertainty of wind and light output, and to realize the adaptive adjustment of the scheduling scheme to uncertainty. By using historical wind and light data, a diversified scenario set is generated through K-means clustering and conditional generative adversarial network, and distribution robust optimization planning and correction are carried out under different scenarios. This method makes the scheduling scheme maintain low line loss, controllable economy and system robustness under different wind and light output fluctuations, effectively overcoming the limitations of traditional methods on subjective probability assumption or artificial membership function dependence. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a whole flow chart of distribution area line loss optimization.

[0031] Figure 2 is a daily line loss curve.

[0032] Figure 3 is a comparison curve of line loss of four scenarios in real-time stage.

[0033] Figures 4-7 are resource scheduling diagrams of extreme scenarios ①, ②, ③ and ④, respectively. DETAILED DESCRIPTION

[0034] The application will be further described in detail below in combination with embodiments and drawings.

[0035] Embodiment 1

[0036] As shown in Figure 1 , the application provides a distribution area two-stage distribution robust line loss optimization method considering wind and light uncertainty. First, a day-ahead-real-time two-stage optimization framework is constructed. In the day-ahead stage, a multi-scenario sample set reflecting the uncertainty of wind and light output is generated by combining K-means clustering algorithm and conditional generative adversarial network (CGAN); in the real-time stage, based on the day-ahead optimization result and the updated output information, a multi-scenario distribution robust optimization model is used to realize the joint correction of line loss and economic dispatch. In the optimization solving process, the forward-backward sweep method (Forward-Backward Sweep) is embedded into the power flow calculation model to accurately reflect the line transmission loss of the distribution network. The model takes the minimization of distribution area line loss and economic dispatch cost as the objective function, and comprehensively considers the uncertainty of renewable energy output, power flow constraints and resource operation boundaries, to realize the optimal operation of the distribution area with economic and robustness. The detailed steps are as follows:

[0037] Step 1: Wind and light uncertainty scenario generation method based on conditional generative adversarial network

[0038] The historical measured data and day-ahead prediction data of wind power and photovoltaic power output are collected as a training sample set, the day-ahead prediction data is taken as a conditional label, and the historical measured wind and light power data are taken as real samples.

[0039] An adversarial game network model including a generator and a discriminator is constructed. The generator adopts a convolutional neural network structure, the input is random noise subject to a normal distribution and day-ahead prediction conditional labels , and the output is generated wind and light power scene samples . The function is to learn the probability distribution characteristics of wind and light power output and generate as realistic simulation samples as possible. The discriminator also adopts a convolutional neural network structure, the input is real samples or generated samples and corresponding labels , and the function is to judge whether the input sample is from real historical data or generated by the generator.

[0040] The loss function and optimization objective of the model need to be defined. The loss function of the generator aims to maximize the discrimination probability of the discriminator on the generated scene, and the expression is:

[0041]

[0042] In the formula, represents the expected value of the corresponding distribution; represents the probability distribution; represents the discrimination output of the discriminator on “generated samples ” meeting the condition , and the closer the value is to 1, the more it conforms to the real sample.

[0043] The loss function of the discriminator is defined as:

[0044]

[0045] In the formula, represents the probability distribution of the generated sample; is the probability of the discriminator judging that the input sample is a “real sample” under the premise that the condition is given.

[0046] The objective function of the conditional generative adversarial network is defined to describe the minimax game between the generator and the discriminator, and the definition is as follows:

[0047]

[0048] To solve the training instability and mode collapse problems caused by the Jensen-Shannon divergence in the original generative adversarial network, the Wasserstein distance is used to replace the JS divergence as the loss function metric indicator. The indicator is used to calculate the distance between the probability distribution of the wind power output generated by the generator and the probability distribution of the real historical data . The advantage is that it can still provide meaningful distance values and gradient information even if the generated distribution and the real distribution overlap very little (i.e. the generated simulation data is poor). This effectively solves the gradient disappearance problem caused by the inability to accurately measure the distribution distance in traditional methods, thereby avoiding model training failure. The Wasserstein distance is defined as:

[0049]

[0050] In the formula, denotes and the joint probability distribution; denotes the distance measure between scenes.

[0051] To solve the problem of direct calculation of the Wasserstein distance in high-dimensional continuous wind power output data space, the Kantorovich-Rubinstein dual theorem can be used to mathematically equivalently transform the original minimization problem of seeking the optimal joint distribution into a maximization problem on the class of 1-Lipschitz continuous functions. The dual form is represented as:

[0052]

[0053] In the formula, denotes that the discriminator D satisfies the 1-Lipschitz continuity.

[0054] To strictly ensure that the discriminator satisfies the above 1-Lipschitz continuity constraint during training, a gradient penalty term is introduced, thereby constructing the overall objective function of the improved Wasserstein conditional generative adversarial network:

[0055]

[0056] In the formula, denotes the norm of the discriminator gradient, is the penalty function.

[0057] Based on the above objective function, the model is trained using an alternating optimization method.

[0058] The collected historical measured data and current-day forecast data are processed. A high-dimensional noise vector conforming to a standard normal distribution is randomly extracted from the latent space. This noise vector is then vertically concatenated with the conditional data (current-day forecast values) in the training samples to construct the generator's input matrix. This input matrix is ​​then fed into the generator network, and after convolution operations, the generator outputs generated samples containing simulated wind and solar power output.

[0059] To enable the discriminator to learn to distinguish between real and fake data, two sets of inputs are needed. The first set involves concatenating the conditional data from the training set with the corresponding real samples (historical measured values) vertically, and then inputting this concatenation into the discriminator to obtain its discrimination value for the real data. The second set involves concatenating the conditional data from the training set with the generated samples output by the generator vertically, and then inputting this concatenation into the discriminator to obtain its discrimination value for the generated data.

[0060] Select gradient penalty sampling points, calculate the loss functions of the generator and discriminator, and update the network weights of the generator and discriminator respectively. If training is not yet complete, return to the next round of training.

[0061] Once training is complete (i.e., the loss function converges or the maximum number of iterations is reached), the trained generator network in the CGAN model is extracted.

[0062] Using a trained generator, a large set of wind and solar power output scenarios is generated by inputting a large amount of random noise and specific pre-holiday forecast data. To verify the accuracy of the generated scenarios in describing temporal correlation and uncertainty, an autocorrelation coefficient is introduced. With partial autocorrelation coefficient The expression for correlation analysis is as follows:

[0063]

[0064]

[0065]

[0066] In the formula, For real samples or generated samples Output value at any given moment This is the mean of the output sequence. For time intervals, For and between Output value at each moment The correlation.

[0067] To evaluate the effectiveness of the scenario set in characterizing uncertainty, coverage and power range width metrics are defined, with the following expressions:

[0068]

[0069]

[0070] In the formula, For confidence level Coverage below The total number of sampling points in the test set. Indicates confidence level The total number of measured wind and solar power values ​​falling within the predicted confidence interval. For confidence level The average width of the power range below, , The nth sampling point is at a confidence level The maximum and minimum power values ​​under these conditions.

[0071] Through the above network training and evaluation process, a set of high-quality scenarios that meet the day-ahead forecast conditions and can fully reflect the uncertainty of wind and solar power output can be obtained.

[0072] The K-means clustering algorithm was used to classify the high-quality scene set generated by CGAN. Through cluster analysis, typical meteorological and power output patterns can be identified, and a large number of samples can be divided into several classes, thereby generating more representative landscape scenes under different climatic characteristics.

[0073] Let the scene dataset be The number of clusters is The optimization objective of K-means clustering is to minimize the sum of squared errors from sample points to cluster centers, as expressed below:

[0074]

[0075] In the formula, Indicates the first Clusters, This is the center vector of the cluster.

[0076] By iteratively updating the sample allocation and center position, the final result is obtained. Each sample represents a typical output pattern, and each type of sample corresponds to different weather scenario characteristics, providing data support for the subsequent two-stage sub-Bruker optimization.

[0077] Step 2: Construct a day-ahead optimization scheduling model that considers line losses in transformer areas.

[0078] Based on the day-ahead forecast data of wind and solar power, a day-ahead dispatch plan is formulated with the goal of minimizing line loss in the transformer area and economic dispatch costs.

[0079] Taking into account the operational characteristics of wind and solar power, energy storage, electric vehicles, and adjustable loads, and with the objective function of minimizing transformer area line loss and economic dispatch cost, a day-ahead optimization dispatch model considering transformer area line loss is constructed as follows:

[0080]

[0081]

[0082] In the formula, This is the loss cost coefficient; This represents the cost coefficient for interacting with the power grid; This represents the cost coefficient for wind curtailment. , , For the operating cost coefficients of wind power, photovoltaic, and energy storage; , The compensation price coefficient for transferable and interruptible loads; The subsidy cost coefficient for coordinating dispatch of electric vehicles; The current day electricity price; This indicates the power exchanged with the power grid (positive for purchasing electricity, negative for selling electricity). , , , , , , , , , , At the node ,time The curtailed wind power, curtailed solar power, actual wind power output, charging power of energy storage devices, discharging power of energy storage devices, input and output power of transferable loads, interruption power of interruptible loads, charging power of electric vehicles, and discharging power of electric vehicles.

[0083] Establish system active / reactive power balance constraints for the day-ahead phase:

[0084]

[0085] In the formula, , Indicates time node The active and reactive loads.

[0086] Establish day-ahead output constraints for wind and solar power:

[0087] ;

[0088]

[0089]

[0090]

[0091]

[0092]

[0093] In the formula, , This represents the maximum output power of the wind power and photovoltaic units at node s at time t. Indicates the amount of wind power curtailed and the amount of solar power curtailed; , This indicates the predicted power of wind and solar power.

[0094] Establish day-ahead operational constraints for energy storage devices:

[0095]

[0096]

[0097]

[0098]

[0099]

[0100] In the formula, express At every moment The state of charge in which energy is stored; , For charge and discharge efficiency; For time step; , This indicates the lower and upper limits of the state of charge. , This represents the charging and discharging state variables at time t; , This indicates the maximum charging and discharging power.

[0101] Establishing the EV cluster: Current operational constraints

[0102]

[0103]

[0104]

[0105]

[0106] In the formula, , The maximum and minimum capacity of electric vehicles, with the minimum capacity required to meet the travel needs of vehicle owners in emergency situations; , For electric vehicle charging (discharging) efficiency; The capacity of the electric vehicle's battery when leaving home; Let be the daily travel distance of the s-th electric vehicle; The amount of electricity required for an electric vehicle to travel 100km.

[0107] Transferable loads (TL) must meet the following constraints: constant total power before and after transfer, transfer ratio, and transfer status.

[0108]

[0109]

[0110]

[0111]

[0112] Where TL represents the category of transferable load, . and These represent the input power and output power of the TL at time t, respectively. and These are the state variables that can be transferred in and out of TL, respectively; and These represent the maximum power input and maximum power output limits of TL, respectively.

[0113] Interruptible loads (ILs) must meet constraints related to interruption ratio, interruption duration, and number of interruptions:

[0114]

[0115]

[0116]

[0117] Where IL represents the category of interruptible load. . It is an interruptible variable in IL. It is an interruptible state variable of IL. It is the maximum interruptible value of IL. and These represent the longest continuous interruptible time and the maximum number of interruptibles for the IL during the scheduling period, respectively.

[0118] The forward-backward substitution method is embedded in the optimization scheduling model to accurately solve the power flow of the distribution network.

[0119] Perform initialization by setting the initial values ​​for all node voltages, typically the rated voltage, with a phase angle of 0, i.e.:

[0120]

[0121] The superscript (0) indicates the 0th iteration.

[0122] The back-substitution process begins at the load node and, based on Kirchhoff's current law, calculates the complex current of each branch layer upstream. The calculation formula is as follows:

[0123]

[0124] in, , Net active and reactive power is injected into the nodes; For nodes Flow to Node The branch current; , for the node A set of directly connected downstream branches;

[0125] The forward calculation process involves obtaining the complex currents of each branch and, starting from the root node with a known voltage, calculating the complex voltages of each node segment by segment downstream, taking into account the branch impedances. The calculation formula is as follows:

[0126]

[0127] At the same time, the node voltage must meet the safety operation constraints:

[0128]

[0129] in, , Downstream nodes and upstream nodes voltage, Type is a connection node With nodes The branch impedance; and These represent the upper and lower limits of the node voltage amplitude, respectively.

[0130] To determine iterative convergence, first calculate the correction amount for the voltage amplitude of all nodes compared to the voltage amplitude of the previous iteration:

[0131]

[0132] In the formula, , The first The voltage amplitude of the next iteration and the first Next iteration: Voltage amplitude. Find the maximum voltage amplitude correction value among all nodes. Determine the convergence condition ,in This is the preset convergence threshold. If the condition is met, the iteration ends; otherwise, the above back-substitution and forward push processes are repeated using the newly calculated nodal complex voltage values ​​until convergence.

[0133] After the power flow calculation converges, the total theoretical line loss of the system is calculated by combining the branch current values. :

[0134]

[0135] In the formula, This represents the total theoretical line loss in the transformer area. Fixed losses of distribution transformers (including no-load losses, auxiliary component losses, etc.). For nodes To the node The current value of the branch circuit; For nodes To the node The resistance of the branch circuit; This is the calculation period.

[0136] The above line loss values Feedback is fed into the objective function, and the final solution model yields a day-ahead resource scheduling scheme that balances economy and low line loss.

[0137] Step 3: Construct a real-time distributed bar optimization scheduling model that considers wind and solar uncertainties.

[0138] The K-means clustering obtained in step 1 Based on the day-ahead scheduling scheme of the extreme wind scenarios and step 2, the real-time stage considers wind power uncertainties and dynamically adjusts the day-ahead decisions. This stage employs the distributed Blue bar optimization method to construct a set of probabilistic uncertainties with extreme scenarios as candidate distributions. And through weights An extreme value distribution is constructed to incorporate the system risk under the worst-case scenario into the optimization objective. Based on this, the model can achieve coordinated and corrective scheduling of resources such as energy storage, electric vehicles, and adjustable loads, effectively suppressing cost increases caused by extreme fluctuations and further reducing line losses in the transformer area.

[0139] Taking into account the operational characteristics of wind and solar power, energy storage, electric vehicles, and adjustable loads, and with the objective function of minimizing transformer line loss and economic dispatch cost, a real-time distributed bar optimization model considering uncertainties is constructed:

[0140]

[0141] Scene index, used to traverse different scenes; This represents the set of weights for extreme scenarios. For the first Weights for extreme scenarios It is by A set of probabilistic uncertainties constructed; , , , , , , , , , Indicates at time Penalty cost coefficients for wind power output adjustment, photovoltaic power output adjustment, loss adjustment, energy storage equipment charging and discharging adjustment, transferable load output and input adjustment, interruptible load adjustment, and electric vehicle charging and discharging adjustment; , Indicates the first A scene, a moment The predicted output of wind power and solar power; , Indicates the first A scene, a moment The actual output of wind power and solar power; , , Indicates the first A scene, a moment The amount of loss adjustment and the amount of charging and discharging power adjustment of energy storage equipment; , , Indicates the first A scene, a moment Adjustments to the transferable load output, transferable load input, and interruptible load; , This represents the adjustment amount of the charging and discharging power of the electric vehicle at time t in the k-th scenario. Node set; Branch set, for branch Its first node is denoted as The end node is denoted as .

[0142] Establish active / reactive power balance constraints for the system in the real-time phase:

[0143]

[0144] In the formula, , Indicates time node The active and reactive loads.

[0145] Establish real-time stage output constraints for wind and solar power:

[0146]

[0147]

[0148]

[0149]

[0150] In the formula, , Let be the wind power and solar power deviation in the k-th scenario.

[0151] Establish real-time phased operational constraints for energy storage devices:

[0152]

[0153]

[0154]

[0155]

[0156]

[0157] Establish real-time operational constraints for the EV cluster:

[0158]

[0159]

[0160]

[0161]

[0162]

[0163] In the formula, , These are the maximum and minimum values ​​for the adjustment amount of the energy storage device.

[0164] Real-time operational constraints for transferable loads:

[0165]

[0166]

[0167]

[0168]

[0169]

[0170] In the formula, , Input the minimum and maximum values ​​of the adjustment amount for transferable loads; , Input the minimum and maximum values ​​of the adjustment amount for the transferable load.

[0171] Interruptible load real-time phase operation constraints:

[0172]

[0173]

[0174] In the formula, , These are the minimum and maximum values ​​for interruptible load adjustments.

[0175] This invention uses K-means clustering to obtain k extreme landscape scenes, each denoted as . (in In the sub-Bruker optimization framework, the worst-case distribution is considered, which is a discrete distribution composed of these extreme scenarios combined with certain probability weights. This extreme value distribution can be expressed as:

[0176]

[0177] In the formula, The weights corresponding to extreme scenario k satisfy the following conditions: and , It is the Dirac delta function.

[0178] To construct a probabilistically uncertain set Weights need to be adjusted. Two types of constraints are imposed: First, the weighted mean of all extreme scenarios should lie within a confidence interval centered on the historical data mean; second, the Wasserstein distance between this extreme value distribution and the empirical distribution must not exceed a given ambiguity parameter. Based on this, the set of uncertainties is defined as follows:

[0179]

[0180] In the formula, It is the probability weight of the k-th extreme scenario. and These are the lower and upper bounds of the population mean, respectively; It is the observation value of the i-th reference sample. It is a fuzzy parameter of the distribution. It represents the total number of reference samples.

[0181] The optimal model of the blue bar is based on the worst distribution. The goal is to find a scheduling strategy that minimizes the sum of expected operating costs and line losses. By solving this model, a robust real-time resource scheduling scheme that considers the uncertainties of wind and solar power can be obtained, along with the corresponding theoretical line loss value.

[0182] Example 2

[0183] The improved IEEE-15 node distribution area was selected as the test object. The network has a typical radial topology, 14 branches, a total active load of 1226 kW, a total reactive load of 1251 kVA, a rated voltage of 11 kV, and a base capacity of 1 MVA. Day-ahead and real-time scheduling results, comparisons of four extreme scenarios, and comparisons with deterministic models are presented, demonstrating the advantages in reducing line losses.

[0184] The four extreme scenarios were derived as follows: First, quantile statistics were performed on the hourly output data of wind and solar power to divide the output into ranges. Hourly output was divided into three levels: less than 25% (L), 25%–75% (M), and greater than 75% (H). Given that both wind and solar scenarios operate on a 24-hour cycle, they were simplified into six four-hour periods, each corresponding to one of the three levels. Based on the quantile results, S scenarios were coded as combinations of L / M / H. Each scenario code was then matched against rows in an orthogonal array. The S scenarios were ranked according to the matching degree, and rows containing combinations of extreme levels were selected, resulting in S extreme scenarios. The top four scenarios were then selected based on their scores.

[0185] Figure 2 The study demonstrates the intraday distribution characteristics of line losses in a distribution substation under typical daily operation. Peak periods, corresponding to peak load times and periods of concentrated distributed energy output, are the main contributors to line losses. These results provide a basis for subsequent two-stage sub-Bruker optimization.

[0186] Figure 3 The paper presents line loss comparison curves in real-time under four extreme scenarios. Multi-scenario analysis can cover various manifestations of wind power uncertainty, thus avoiding misjudgment of line loss patterns based on a single scenario. Decision-making based on the constructed scenario set achieves an overall reduction in line loss.

[0187] Figures 4-7 The resource scheduling performance of the SDRO model under four scenarios is demonstrated. It shows good results in all scenarios, effectively reducing line losses through flexible adjustments to the coordinated use of resources such as energy storage and electric vehicles. Although the total cost of the model fluctuates slightly, it successfully achieves significant optimization of line loss costs, forming a reasonable trade-off between system operating efficiency and robustness.

[0188] In summary, this invention proposes a two-stage distributed grid optimization scheduling method for wind power, photovoltaics, energy storage, electric vehicles, and adjustable loads under uncertainties in wind power and photovoltaic power distribution networks. This method can effectively reduce line active power losses and suppress cost fluctuations under extreme conditions, and has good engineering application value.

Claims

1. A two-stage distribution robust line loss optimization method for a transformer area considering wind and light uncertainty, characterized in that, Firstly, a day-ahead-real-time two-stage line loss optimization method is constructed. In the day-ahead stage, a multi-scenario sample set reflecting the uncertainty of wind and solar output is generated by combining the K-means clustering algorithm with the conditional generative adversarial network; In the real-time stage, based on the day-ahead optimization results and updated output information, a multi-scenario distribution robust optimization model is used to realize the joint correction of line loss and economic dispatch; In the optimization solving process, the forward-backward substitution method is embedded into the power flow calculation model to accurately reflect the line transmission loss of the distribution network; Secondly, a distribution robust optimization method based on scenario generation is constructed. A diversified scenario set is generated by using historical wind and solar data through K-means clustering and conditional generative adversarial network, and distribution robust optimization planning and correction are carried out under different scenarios. 2.The two-stage distribution robust line loss optimization method considering wind and light uncertainty of a transformer area according to claim 1, wherein, The steps are as follows: Step 1: wind and solar uncertainty scenario generation method based on conditional generative adversarial network Step 1.1 Collect historical measured data and day-ahead prediction data of wind and solar output as training sample set, take day-ahead prediction data as conditional label, and take historical measured wind and solar power data as real sample; Step 1.2 constructing an adversarial game network model comprising a generator and a discriminator ; wherein the generator adopts a convolutional neural network structure, the input being random noise subject to a normal distribution and a day-ahead prediction condition label , and the output being a generated wind-solar power scene sample ; the discriminator also adopts a convolutional neural network structure, the input being a real sample or a generated sample and a corresponding label , and the role thereof being to determine whether the input sample is from real historical data or generated by the generator; Step 1.3 Define the objective function of conditional generative adversarial network to characterize the minimax game between generator and discriminator; Based on the objective function, the model is trained by alternating optimization; Use the trained generator to input a large amount of random noise and specific day-ahead prediction data to generate a large-scale wind and solar output scenario set; Step 1.4 verifies the accuracy of the timing correlation and uncertainty description of the generated scenario, introducing the autocorrelation coefficient with the partial autocorrelation coefficient correlation analysis is performed; Step 1.

5. Obtain a set of scenarios that meet the conditions of day-ahead prediction and can fully reflect the uncertainty of wind and light output by training and evaluating the process through the network; use the K-means clustering algorithm to divide the high-quality scenario set generated by the CGAN into modes, and finally obtain representative typical output modes by iteratively updating the sample allocation and center position, each class of samples corresponding to different weather scene characteristics; Step 2: Construct day-ahead optimization scheduling model considering substation line loss Step 2.1 Based on wind and solar day-ahead prediction data, develop day-ahead scheduling plan with the objective of minimizing substation line loss and economic dispatch cost; Step 2.2 Considering the operating characteristics of wind, solar, energy storage, electric vehicles and adjustable load, construct day-ahead optimization scheduling model considering substation line loss with the objective function of minimizing substation line loss and economic dispatch cost; Step 2.3 Establish system active / reactive power balance constraints in day-ahead stage, wind and solar output constraints in day-ahead stage, energy storage device operation constraints in day-ahead stage, EV cluster operation constraints in day-ahead stage, transferable load (TL), interruptible load; Step 2.

4. Embed the forward-backward substitution method into the optimization scheduling model to accurately solve the power flow of the distribution network. After the power flow calculation converges, calculate the total theoretical line loss value of the system combined with the branch current value ; Feed it back to the objective function to finally solve the model and get the day-ahead resource scheduling scheme considering economy and low line loss; Step 3: Construct real-time distribution robust optimization scheduling model considering wind and solar uncertainty Step 3.1 The extreme wind scenarios obtained by K-means clustering in Step 1 On the basis of the extreme wind scenarios obtained in Step 1 and the scheduling scheme of the day-ahead stage, the real-time stage considers the uncertainty of wind power and dynamically corrects the day-ahead decision; a distribution robust optimization method is used to construct a probability uncertainty set with extreme scenarios as candidate distributions , and an extreme value distribution is constructed through weights , so as to include the system risk in the worst scenario into the optimization objective; Step 3.2 Establish system active / reactive power balance constraints in real-time stage, wind and solar output constraints in real-time stage, energy storage device operation constraints in real-time stage, EV cluster operation constraints in real-time stage, transferable load operation constraints in real-time stage, interruptible load operation constraints in real-time stage; Step 3.

3. Get k extreme weather scenes by K-means clustering, each scene is denoted as (Where ); In the distribution robust optimization framework, consider the worst-case distribution, that is, the discrete distribution composed of these extreme scenes with certain probability weights; Step 3.4 is constructing a probabilistic uncertainty set , imposing two types of constraints on the weights : first, the weighted average of all extreme scenarios should lie within a confidence interval centered on the average of the historical data; second, the Wasserstein distance between the extreme-value distribution and the empirical distribution should not exceed a given fuzziness parameter ; Step 3.5 Robust optimization model under worst distribution Next, the scheduling strategy is sought to minimize the sum of expected operation cost and line loss; by solving this model, the real-time resource scheduling scheme considering wind and solar uncertainty and the corresponding theoretical line loss value are obtained. 3.The two-stage distribution robust line loss optimization method considering wind and light uncertainty of a transformer area according to claim 2, wherein, The objective function of conditional generative adversarial network defined in step 1.3 is used to characterize the minimax game between generator and discriminator, which is defined as follows: 。 4.The two-stage distribution robust line loss optimization method considering wind and light uncertainty of a transformer area according to claim 2, wherein, Step 1.4 verifies the accuracy of the time sequence correlation and uncertainty description of the generated scene, and introduces a self-correlation coefficient With the partial autocorrelation coefficient Correlation analysis is performed, and its expression is as follows: wherein is the output value at time is the mean of the output sequence, is the time interval, is the time interval, is the time interval, is the time interval, is the time interval, is the correlation of the output values at the times . 5.The two-stage distribution robust line loss optimization method considering wind and light uncertainty of a transformer area according to claim 2, wherein, In step 1.5, typical weather and output patterns can be identified by cluster analysis, a large number of samples are divided into several categories, and more representative wind and light scenes are generated under different climate characteristics, and the scene data set is , the number of clusters is , and the optimization goal of K-means clustering is to minimize the sum of squared errors of sample points to cluster centers, expressed as follows: wherein represents the cluster, is the center vector of the cluster. 6.The two-stage distribution robust line loss optimization method considering wind and light uncertainty of a transformer area according to claim 2, wherein, The objective function of the day-ahead optimization scheduling model considering substation line loss in step 2.2 is as follows: In the formula, This is the loss cost coefficient; This represents the cost coefficient for interacting with the power grid; This represents the cost coefficient for wind curtailment. , , For the operating cost coefficients of wind power, photovoltaic, and energy storage; , The compensation price coefficient for transferable and interruptible loads; The subsidy cost coefficient for coordinating dispatch of electric vehicles; The current day electricity price; Indicates the power interacting with the power grid; , , , , , , , , , , At the node ,time The curtailed wind power, curtailed solar power, actual wind power output, charging power of energy storage devices, discharging power of energy storage devices, input and output power of transferable loads, interruption power of interruptible loads, charging power of electric vehicles, and discharging power of electric vehicles.

7. The two-stage distribution robust line loss optimization method considering wind and light uncertainty according to claim 2, wherein, Step 2.

4. After the convergence of the power flow calculation, the system total theoretical line loss value is calculated in combination with the branch current value : In the formula, is the total theoretical line loss of the transformer district; is the fixed loss of the distribution transformer, including the no-load loss and auxiliary element loss; is the node is the node is the current value of the branch; is the node is the node is the resistance of the branch; is the calculation period. 8.The two-stage distribution robust line loss optimization method considering wind and light uncertainty of a transformer area according to claim 2, wherein, The real-time distribution robust optimization model considering uncertainty in step 3.1 is as follows: Scene index, used to traverse different scenes; This represents the set of weights for extreme scenarios. For the first Weights for extreme scenarios It is by A set of probabilistic uncertainties constructed; , , , , , , , , , Indicates at time Penalty cost coefficients for wind power output adjustment, photovoltaic power output adjustment, loss adjustment, energy storage equipment charging and discharging adjustment, transferable load output and input adjustment, interruptible load adjustment, and electric vehicle charging and discharging adjustment; , Indicates the first A scene, a moment The predicted output of wind power and solar power; , Indicates the first A scene, a moment The actual output of wind power and solar power; , , Indicates the first A scene, a moment The amount of loss adjustment and the amount of charging and discharging power adjustment of energy storage equipment; , , Indicates the first A scene, a moment Adjustments to the transferable load output, transferable load input, and interruptible load; , This represents the adjustment amount of the charging and discharging power of the electric vehicle at time t in the k-th scenario; Node set; Branch set, for branch Its first node is denoted as The end node is denoted as . 9.The two-stage distribution robust line loss optimization method considering wind and light uncertainty according to claim 2, wherein, Step 3.

3. Get k extreme weather scenes by K-means clustering, each scene is denoted as (Where ); In the distribution robust optimization framework, consider the worst-case distribution, that is, the discrete distribution composed of these extreme scenes with certain probability weights; this extreme distribution can be expressed as: wherein is the weight corresponding to the extreme scenario k, satisfying and , is the Dirac delta function. 10.The two-stage distribution robust line loss optimization method considering wind and light uncertainty according to claim 2, wherein, In step 3.4, the uncertainty set is defined as follows: wherein is the probability weight of the kth extreme scenario, and are the lower and upper bounds of the population mean, respectively; is the observation of the ith reference sample, is the fuzziness parameter of the distribution, is the total number of reference samples; The distribution robust optimization model is to minimize the sum of the expected operation cost and the line loss Next, the scheduling strategy is sought that minimizes the sum of the expected operation cost and the line loss.

Images