Load operation scenario generation method and system based on boundary condition constraint guidance

Abstract

A load operation scene generation method and system based on boundary condition constraint guidance, comprising: collecting historical source load data and meteorological data; dividing the operation scene into extreme scenes and normal scenes, and setting scene labels; classifying the preprocessed data based on the divided operation scenes; constructing a CGAN-based operation scene generation model, introducing scene labels as conditional inputs in the generator and discriminator, and constructing a physical constraint loss function including load and meteorological matching loss, new energy output and meteorological matching loss, and supply and demand balance constraint loss; training the model and generating each operation scene data; using an index system including structural similarity index, KL divergence and mean square error to evaluate the generated data, and optimizing the operation scene generation model based on the evaluation result. The present application can effectively solve the technical difficulties of the existing load operation scene generation method, such as the lack of physical rationality and incomplete full-scene coverage in extreme scenes.

Description

Technical Field

[0001] This invention belongs to the field of power system load management and operation optimization, and in particular relates to a method and system for generating load operation scenarios based on boundary condition constraints. Background Technology

[0002] In the development of power system load management and operation scenario construction, early research mainly served load forecasting and dispatch optimization under centralized power supply structures, employing static operating conditions and rule-based modeling methods, which were insufficient to cope with sudden and complex coupled extreme operating conditions. With the increasing penetration of renewable energy, new extreme scenarios such as "extreme heat without wind," "no solar power during evening peak hours," "continuous wind and solar power shortages," and primary energy scarcity are frequently occurring, posing bottlenecks to traditional models in scenario description and operational strategy response. In recent years, the development of artificial intelligence, especially deep learning, has provided a new technological path for the intelligent construction of load management operation data scenarios. From the early construction of typical daily load curves to the introduction of generative models to enhance the simulation of diverse and rare scenarios, the technological paradigm has gradually evolved into an intelligent generative framework centered on data-driven approaches and aiming for scenario completeness.

[0003] Currently, both domestic and international research has gradually established a data-driven technological path centered on generative models for constructing extreme load scenarios in power systems, encompassing key aspects such as deep learning, multimodal fusion, transfer learning, and sample quality assessment. International research leads in generative model innovation, scenario diversity, and generalization capabilities, emphasizing algorithm performance and adaptability to complex scenarios; while domestic research focuses on close integration with engineering applications to improve model practicality and deployment efficiency, driving the rapid development of scenario data platforms and intelligent generation systems.

[0004] Chinese patent application CN119988925A discloses a method for generating extreme weather scenarios based on the CGAN algorithm. It extracts meteorological indicators and power grid operation status indicators under different types of extreme weather; establishes statistical characteristic indicators for risk scenarios; and selects extreme weather events affecting power grid safety and stability to form a training set. It also establishes a conditional generative adversarial network model capable of generating extreme weather scenarios. However, it still has the following shortcomings: data preprocessing lacks systematic optimization, and spatiotemporal consistency and physical correlation are insufficiently guaranteed. The training set is constructed only by screening historical extreme weather data, without systematically addressing the spatiotemporal misalignment and short-term random disturbances between source-load data and meteorological data. Its data processing does not involve spatiotemporal alignment operations, and it does not employ targeted filtering algorithms to preserve the inherent physical correlation of the data. Simple screening alone cannot eliminate the interference of measurement noise on the characteristics of extreme scenarios, thus affecting the realism of the generated scenarios. Scenario division only focuses on two types of weather, resulting in a single dimension and incomplete coverage. The CGAN model does not embed physical constraint loss terms, leading to insufficient physical compliance of the generated scenarios. Verification methods lack multi-dimensional quantitative evaluation, resulting in inaccurate judgment of scenario effectiveness.

[0005] Chinese patent application CN117350161A discloses a method for generating extreme wind power scenarios based on conditional generative adversarial networks (CGANs). It enhances extreme wind power data, categorizes common real-world extreme situations, and derives three extreme metrics: high wind output, low wind output, and large peak-to-valley difference. A CGAN model is then built to generate and evaluate these scenarios. However, it still has the following shortcomings: scenario coverage is limited to wind power alone, failing to address multi-energy and load composite scenarios; data preprocessing does not handle spatiotemporal alignment and short-term disturbances, resulting in noise interference in the enhanced data; CGAN lacks a physical constraint loss term, requiring repeated parameter adjustments for scenario generation, leading to low efficiency; and the validation metrics only focus on power error, lacking assessments of structural similarity and distribution consistency.

[0006] In summary, current scene segmentation methods largely rely on a single threshold, failing to incorporate multi-dimensional constraints such as physical thresholds, supply and demand indicators, and weather types. Single-threshold segmentation struggles to accurately define scene boundaries, blurring the distinction between extreme and normal scenarios, and resulting in generated scenes lacking clear physical meaning and interpretability. Furthermore, existing technologies employing generative models like CGAN do not embed targeted physical constraint loss terms into the loss function. Relying solely on adversarial loss terms only ensures the statistical realism of the generated data, failing to constrain the physical logic between variables, leading to insufficient compliance of the generated data. Moreover, current technologies lack diverse evaluation metrics for scene effectiveness verification.

[0007] Therefore, there are still significant challenges in terms of stable modeling capabilities with few samples, accuracy of long-term construction across scenarios, and complete representation of high-dimensional feature coupling states. It is necessary to further strengthen research in areas such as multi-source data fusion, reliability assessment of generated results, and dynamic response modeling under extreme operating conditions to support the construction of a load management system with higher reliability and stronger resilience under new power systems. Summary of the Invention

[0008] To address the shortcomings of existing technologies in terms of stable modeling capabilities with few samples, accuracy of cross-scenario long-term construction, and complete representation of high-dimensional feature coupling states, this invention provides a load operation scenario generation method and system based on boundary condition constraints.

[0009] The present invention adopts the following technical solution.

[0010] In a first aspect, the present invention discloses a method for generating load operation scenarios based on boundary condition constraints, the method comprising the following steps: Step 1: Collect historical source load data and corresponding meteorological data at the same time, and perform preprocessing; Step 2: Divide the operating scenarios into extreme scenarios and normal scenarios, and set scenario labels; classify the preprocessed historical data based on the divided operating scenarios to obtain historical data under each operating scenario; Step 3: Construct a CGAN-based scenario generation model, introduce scenario labels as conditional inputs into the generator and discriminator, and construct a physical constraint loss function that includes load and weather matching loss, new energy output and weather matching loss, and supply and demand balance constraint loss. Step 4: Train the model based on historical data and corresponding scene labels for each operating scenario, and use the trained model to generate data for each operating scenario; Step 5: The generated data for each running scenario are evaluated using a three-dimensional indicator system, including structural similarity index, KL divergence, and mean squared error. The running scenario generation model is then optimized based on the evaluation results.

[0011] More preferably, In step 2, the extreme scenarios include extremely hot and windless conditions, evening peak hours without sunlight, continuous windless and sunless conditions, shortage of primary energy supply, sudden changes in new energy output, and continuous low wind and solar power output; the normal scenarios are all other scenarios except for the extreme scenarios.

[0012] More preferably, The CGAN-based scenario generation model has a total loss function that is a weighted sum of adversarial loss, physical constraint loss, and regularization loss, specifically:

[0013] in, To combat the losses; This is the loss due to physical constraints, used to ensure physical compliance; This is the regularization loss; The weighting coefficients used to combat loss can be dynamically adjusted according to the model distribution fitting requirements; , These are the weighting coefficients for physical constraint loss and regularization loss, respectively, which can be determined based on the priority of physical constraints and the generalization requirements of the model.

[0014] More preferably, In step 3, the load-meteorological matching loss is determined as follows:

[0015] in, Losses due to load matching with weather conditions; The number of time steps; for Temperature at any given time; This is the high-temperature threshold, which can be adjusted by region; for Constant load; The high-temperature load threshold is set when the temperature is higher than a certain value. The 90th percentile of the load during historically high-temperature periods; It is a proportionality coefficient, and satisfies This is used to adjust the degree of influence of temperature and load deviation on losses; Let be the fault tolerance coefficient, and satisfy . It is used to add tolerance space to physical constraints and avoid excessive penalties and numerical instability during model training.

[0016] More preferably, In step 3, the loss due to matching the power output of the new energy source with weather conditions is determined in the following manner:

[0017] in, To match the losses incurred in generating new energy sources with meteorological conditions, for Wind speed at all times; The low wind speed threshold for wind power; for Wind power output at all times; To determine the low wind power output threshold, a wind speed lower than [missing information] is used. The 10th percentile of wind power output during historically low wind speed periods; for Light intensity at any given time; This represents the low light intensity threshold for photovoltaic applications. for Photovoltaic power output at all times; To determine the low photovoltaic output threshold, the light intensity is taken as below... The 10th percentile of photovoltaic output during historically low sunshine periods; It is a proportionality coefficient, and satisfies ; Let be the fault tolerance coefficient, and satisfy . .

[0018] More preferably, In step 3, the supply and demand balance constraint loss is determined in the following manner:

[0019] in, Losses are constrained by supply and demand imbalance; Number of scene types For the first Load ramp-up rate for similar scenarios; For the first The ramp-up rate threshold for similar scenarios is determined based on the 85th percentile of the ramp-up rate for similar scenarios in history. The tolerance range for the ramp rate, and satisfying the following conditions. .

[0020] More preferably, In step 5, the generated runtime scenario data is evaluated using structural similarity metrics, specifically including: For historical electrical load time series data and generated electrical load time series data of the same scenario and time period, a sliding window is used to extract the data, resulting in a one-to-one corresponding historical sliding window dataset and generated sliding window dataset. The local structural similarity index value of each sliding window is calculated, and the average value of the local structural similarity index values ​​of all sliding windows is taken as the final structural similarity index value of the generated running scenario data. The local structural similarity index value is determined in the following manner:

[0021] in, For data within the historical sliding window, Generate the corresponding data within the sliding window; , They are respectively , The local mean; , They are respectively , The local variance, for and Local covariance; , All are stability constants, and satisfy... , .

[0022] Secondly, the present invention discloses a load operation scenario generation system based on the aforementioned method and guided by boundary condition constraints, including a data acquisition module, a scenario division and historical data classification module, an operation scenario generation model construction module, an operation scenario data generation module, and an operation scenario generation data evaluation module. The data acquisition module is used to collect historical source load data and meteorological data at the same time, and to perform preprocessing. The scenario segmentation and historical data classification module is used to divide the operating scenarios into extreme scenarios and normal scenarios, and set scenario labels; the preprocessed historical data is classified based on the segmented operating scenarios to obtain the historical data under each operating scenario; The runtime scenario generation model building module is used to build a runtime scenario generation model based on CGAN. It introduces scenario labels as conditional inputs into the generator and discriminator, and constructs a physical constraint loss function that includes load and weather matching loss, new energy output and weather matching loss, and supply and demand balance constraint loss. The runtime scenario data generation module is used to train the model based on historical data and corresponding scenario labels in each runtime scenario, and to generate runtime scenario data using the trained model. The runtime scenario generation data evaluation module is used to evaluate the generated runtime scenario data using a three-dimensional index system including structural similarity index, KL divergence and mean squared error, and to optimize the runtime scenario generation model based on the evaluation results.

[0023] Thirdly, the present invention provides a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of the first aspects of the present invention.

[0024] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects of the present invention.

[0025] The beneficial effects of this invention are compared with those of the prior art: 1. It achieves precise definition of normal and extreme scenarios, so that each generated scenario corresponds to clear operational characteristics and physical meaning, solving the problems of vague scenario boundaries and lack of interpretability in traditional technologies; 2. Traditional CGAN relies solely on adversarial loss to optimize distribution fitting, which can easily lead to generated data that violates the physical laws of the power system. This invention introduces a physical constraint loss function into the model, integrating three core physical constraints: load and weather matching, renewable energy output and weather matching, and supply and demand balance. At the same time, it ensures the consistency of statistical distribution through adversarial loss, achieving a dual guarantee of statistical realism and physical compliance, so that the generated data conforms to the actual operating logic of the power system. 3. The total loss function adopts a weighted sum of adversarial loss, physical constraint loss, and regularization loss. The weight coefficients of the three types of loss can be dynamically adjusted according to the scenario priority, model generalization requirements, and numerical magnitude balance. In normal scenarios, it can balance distribution fitting and physical constraints, in extreme scenarios, it can strengthen the priority of physical constraints, and in small dataset scenarios, it can increase the regularization weight to prevent overfitting. It can flexibly adapt to the engineering application requirements of different power grids and has strong versatility. 4. Construct a three-dimensional indicator system including structural similarity index, KL divergence and mean squared error to comprehensively evaluate the quality of generated data from three dimensions: consistency of temporal structure, goodness of fit of statistical distribution and accuracy of numerical error. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the load operation scenario generation method based on boundary condition constraints of the present invention. Figure 2 This is a diagram of the CGAN network structure in Embodiment 1 of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0028] like Figure 1 As shown, this invention discloses a method for generating load operation scenarios based on boundary condition constraints, the method comprising the following steps: Step 1: Collect historical source load data and corresponding meteorological data at the same time, and perform preprocessing; Step 2: Divide the operating scenarios into extreme scenarios and normal scenarios, and set scenario labels; classify the preprocessed historical data based on the divided operating scenarios to obtain historical data under each operating scenario; The extreme scenarios include extremely hot and windless conditions, evening peak hours without sunlight, continuous periods without wind or sunlight, shortages in primary energy supply, sudden changes in new energy output, and sustained low wind and solar power output; the normal scenarios are all other scenarios except for the extreme scenarios.

[0029] Step 3: Construct a CGAN-based scenario generation model, introduce scenario labels as conditional inputs into the generator and discriminator, and construct a physical constraint loss function that includes load and weather matching loss, new energy output and weather matching loss, and supply and demand balance constraint loss. The CGAN-based scenario generation model has a total loss function that is a weighted sum of adversarial loss, physical constraint loss, and regularization loss, specifically:

[0030] in, To combat the losses; This is the loss due to physical constraints, used to ensure physical compliance; This is the regularization loss; The weighting coefficients used to combat loss can be dynamically adjusted according to the model distribution fitting requirements; , These are the weighting coefficients for physical constraint loss and regularization loss, respectively, which can be determined based on the priority of physical constraints and the generalization requirements of the model.

[0031] The load-meteorological matching loss is determined in the following manner:

[0032] in, Losses due to load matching with weather conditions; The number of time steps; for Temperature at any given time; This is the high-temperature threshold, which can be adjusted by region; for Constant load; The high-temperature load threshold is set when the temperature is higher than a certain value. The 90th percentile of the load during historically high-temperature periods; It is a proportionality coefficient, and satisfies This is used to adjust the degree of influence of temperature and load deviation on losses; Let be the fault tolerance coefficient, and satisfy . It is used to add tolerance space to physical constraints and avoid excessive penalties and numerical instability during model training.

[0033] The loss due to matching the power output of the new energy source with weather conditions shall be determined in the following manner:

[0034] in, To match the losses incurred in generating new energy sources with meteorological conditions, for Wind speed at all times; The low wind speed threshold for wind power; for Wind power output at all times; To determine the low wind power output threshold, a wind speed lower than [missing information] is used. The 10th percentile of wind power output during historically low wind speed periods; for Light intensity at any given time; This represents the low light intensity threshold for photovoltaic applications. for Photovoltaic power output at all times; To determine the low photovoltaic output threshold, the light intensity is taken as below... The 10th percentile of photovoltaic output during historically low sunshine periods; It is a proportionality coefficient, and satisfies ; Let be the fault tolerance coefficient, and satisfy . .

[0035] The supply-demand balance constraint loss is determined in the following manner:

[0036] in, Losses are constrained by supply and demand imbalance; Number of scene types For the first Load ramp-up rate for similar scenarios; For the first The ramp-up rate threshold for similar scenarios is determined based on the 85th percentile of the ramp-up rate for similar scenarios in history. The tolerance range for the ramp rate, and satisfying the following conditions. .

[0037] Step 4: Train the model based on historical data and corresponding scene labels for each operating scenario, and use the trained model to generate data for each operating scenario; Step 5: The generated data for each running scenario are evaluated using a three-dimensional indicator system, including structural similarity index, KL divergence, and mean squared error. The running scenario generation model is then optimized based on the evaluation results.

[0038] The generated runtime scenario data is evaluated using structural similarity metrics, specifically including: For historical electrical load time series data and generated electrical load time series data of the same scenario and time period, a sliding window is used to extract the data, resulting in a one-to-one corresponding historical sliding window dataset and generated sliding window dataset. The local structural similarity index value of each sliding window is calculated, and the average value of the local structural similarity index values ​​of all sliding windows is taken as the final structural similarity index value of the generated running scenario data. The local structural similarity index value is determined in the following manner:

[0039] in, For data within the historical sliding window, Generate the corresponding data within the sliding window; , They are respectively , The local mean; , They are respectively , The local variance, for and Local covariance; , All are stability constants, and satisfy... , .

[0040] Example 1: like Figure 1 As shown, this invention discloses a method for generating load operation scenarios based on boundary condition constraints, comprising the following steps: Step 1: Collect historical source load data and corresponding meteorological data at the same time, and perform preprocessing; The historical source-load data includes wind power output, photovoltaic power output, and electricity load data; the meteorological data includes, but is not limited to, temperature, wind speed, and solar intensity.

[0041] More preferably, the historical source-load data and accompanying meteorological data are smoothed using an extended Kalman filter; The smoothing process of historical source load data and accompanying meteorological data using extended Kalman filtering specifically includes the following steps: Step 1.1: Collect historical data on electricity load, wind power output, photovoltaic power output, and related meteorological conditions at one sampling point every 15 minutes; Step 1.2: Perform extended Kalman filtering on historical source loads and corresponding meteorological data. The core is to filter out measurement noise and short-term random disturbances through local linearization iteration of the nonlinear system, while preserving the inherent physical correlation of the data.

[0042] First, the data is preprocessed. In this embodiment, the 3σ criterion or box plot is preferably used to identify and remove outliers such as sensor malfunctions and instantaneous pulses, and missing data is filled using linear interpolation. Based on timestamps and spatial region identifiers, spatiotemporal alignment of source load data and meteorological data is achieved. Second, for the physical characteristics of different data types, specific nonlinear state equations and observation equations are constructed to quantify the correlation between the true data values, observed values, and meteorological factors. Then, the system state vector and noise covariance matrix are initialized, and the state estimate is gradually optimized through prediction and update iteration to achieve noise filtering and data smoothing. Finally, the filtered dataset is subjected to boundary constraint verification to ensure that the smoothed data is free from noise interference and strictly follows the physical laws of source load-meteorological coupling.

[0043] Step 2: Divide the operating scenarios into extreme scenarios and normal scenarios, and set scenario labels; classify the preprocessed historical data based on the divided operating scenarios to obtain historical data under each operating scenario; The scenarios are divided into extreme and normal categories based on physical thresholds, supply and demand indicators, etc., and the specific steps include the following: Step 2.1: Divide the extreme scenarios; The extreme scenarios include extreme heat and no wind, no light during the evening peak, continuous no wind and no light, shortage of primary energy supply, sudden changes in new energy output, and continuous low wind and solar output; the normalized scenarios are all other scenarios except for the extreme scenarios.

[0044] The extreme scenarios described consider multiple factors, including meteorological conditions, renewable energy output characteristics, and primary energy supply. Each scenario is defined by specific physical thresholds, supply and demand indicators, and representative weather conditions. These indicators vary depending on the data source for different regions. The specific criteria for scenario classification for a particular region are as follows: (1) Extremely hot and windless scene The scenario of extreme heat and no wind is defined as one that simultaneously meets both high temperature meteorological conditions and extremely low wind power output conditions. Specifically, the following criteria must be met simultaneously: 1) Regional temperature > 25℃; 2) Regional wind power equivalent wind speed < 3.0m / s; 3) Regional wind power output < 20MW. The representative weather condition for the extremely hot and windless scenario is a heat wave.

[0045] (2) Evening peak without light A scenario that simultaneously meets the low photovoltaic output condition and is during the evening peak period of 18:00-22:00 is defined as an evening peak no-light scenario. Specifically, the low photovoltaic output condition must simultaneously meet the following criteria: 1) Regional photovoltaic equivalent irradiance < 40W / m². If there is no irradiance monitoring data, the photovoltaic power is multiplied by 10 to convert it into equivalent irradiance for judgment; 2) Regional photovoltaic power < 5MW. The representative weather conditions for the dark evening peak scene are: dense fog and heavy rain.

[0046] (3) Continuous windless and lightless scene A continuous no-wind and no-solar scenario is defined as one that simultaneously meets the conditions of extremely low photovoltaic output and extremely low wind power output, and the continuous operating time is ≥3 hours. Specifically, the continuous no-wind and no-solar scenario must simultaneously meet the following criteria: 1) The equivalent irradiance of the photovoltaic system in the area is <15W / m². If there is no irradiance monitoring data, the photovoltaic power is multiplied by 10 to convert it into equivalent irradiance for judgment; 2) The equivalent wind speed of the wind power system in the area is <2.0m / s; 3) The continuous operating time of the above conditions is ≥3 hours. The representative weather condition for the continuous windless and lightless scenario is: anticyclonic stagnation.

[0047] (4) Scenarios with a shortage of primary energy supply A simultaneous and significant decline in both wind and solar power output constitutes a scenario of primary energy supply shortage. Specifically, the scenario of primary energy supply shortage must simultaneously meet the following criteria: 1) Wind power output decline rate > 2% or absolute wind power output decline > 2MW; 2) Solar power output decline rate > 2% or absolute solar power output decline > 1MW. The representative weather conditions for the aforementioned primary energy supply shortage scenario are: hail and thunderstorms.

[0048] (5) Scenarios of sudden changes in new energy output When the time period is not the evening peak period of 18:00-22:00, and meets any of the following conditions, it is considered a sudden change scenario in new energy output: 1) Photovoltaic power change rate > 2% or absolute value of photovoltaic power change > 1MW; 2) Wind power change rate > 2% or absolute value of wind power change > 1MW; The representative weather conditions for the sudden changes in the output of new energy sources are: solar eclipse, cold wave, and typhoon.

[0049] (6) Scenarios with sustained low power output of wind and solar power A scenario characterized by continuous low wind and solar power output is defined as one where both wind and solar power output are simultaneously met, and the operating conditions are continuous for ≥4 hours. Specifically, the following criteria must be met simultaneously: 1) Equivalent wind speed in the region <2.5m / s; 2) Equivalent solar irradiance in the region <80W / m². If there is no irradiance monitoring data, the equivalent irradiance is determined by multiplying the solar power by 10; 3) The above operating conditions are continuous for ≥4 hours. The representative weather conditions for the scenario of sustained low wind and solar power output are: continuous light wind and the rainy season.

[0050] Step 2.2: Classification of Normal Scenarios; Apart from the six extreme scenarios mentioned in step 2.1, all other scenarios can be considered normal scenarios.

[0051] Step 3: Construct a CGAN-based scenario generation model, introduce scenario labels as conditional inputs into the generator and discriminator, and construct a physical constraint loss function that includes load and weather matching loss, new energy output and weather matching loss, and supply and demand balance constraint loss. By combining constraints from extreme and normal operating scenarios, physical constraints are embedded into a condition-based generative adversarial network, specifically including the following steps: Step 3.1: Construct a runtime scenario generation model based on CGAN; like Figure 2 As shown, Conditional Generative Adversarial Networks (CGANs) are a variant of generative adversarial networks. These networks take conditional variables as input and provide deep learning models with the ability to generate new data with specific characteristics. In CGANs, both the generator and discriminator receive additional conditional information as input, allowing the generated data to satisfy these specific conditions.

[0052] Compared to ordinary generative adversarial networks (GANs), conditional generative adversarial networks (GANs) consider additional conditional variables when generating data, greatly improving the specificity and diversity of the generated data, making it an ideal data augmentation tool. The generator, in addition to simulating the distribution of real data, must also satisfy specific conditional constraints, while the discriminator needs to determine whether a given sample is real and whether it meets the given conditions.

[0053] Conditional generative adversarial networks (GANs) have proven highly effective in addressing the challenge of enhancing operational data in the renewable energy market. By setting various typical operational scenarios as conditional variables, the generator can learn from the existing data distribution and generate renewable energy operational data that conforms to specific scenarios based on different conditions, providing a large number of highly diverse simulation samples. Probability distribution characteristics, renewable energy output, and other conditions can be precisely incorporated into the generated data, thereby effectively achieving data balancing and enhancement.

[0054] The application of this machine learning model effectively solves the problem of insufficient data in the current electricity market, making it difficult to describe and process complex operating conditions. Furthermore, through techniques such as overlay, fusion, truncation, and distributed sampling, the generated data can be further optimized to better meet actual operating needs, enhancing the breadth and effectiveness of the data and improving the operating efficiency and accuracy of the new energy power system.

[0055] The loss function of this invention is a ternary structure of "adversarial loss + physical constraint loss + regularization loss", in which the physical constraint loss term is fully quantized and directly corresponds to the physical logic of extreme scenarios, thus solving the core defect of "no physical constraints" in the existing technology.

[0056] Total loss function expression: (1) in, To mitigate losses and ensure statistical fidelity; This is the loss due to physical constraints, used to ensure physical compliance; This is the regularization loss, used to prevent overfitting; To counteract the loss, the weighting coefficients can be dynamically adjusted according to the model distribution fitting requirements. In this embodiment, preferably, ; , The weighting coefficients for physical constraint loss and regularization loss are determined based on the priority of physical constraints, model generalization requirements, and the balance of the magnitudes of each loss term. In this embodiment, preferably... , ; Existing technologies only use Wasserstein distance but lack physical constraints. This invention optimizes adversarial loss by incorporating scene conditions. : (2) in, For the true data distribution, The random noise distribution, typically a standard normal distribution, serves as the random source of input to the generator. The scenario condition distribution represents the distribution of condition variables in the current task scenario; Input distribution for gradient penalty term; The generator output is the given scene. and noise Then, generate pseudo-data that matches the scenario; This indicates that under scenario condition y, the discriminator outputs to the generator. Authenticity rating; For the discriminator output, i.e., in a given scene Next, for the input data Authenticity assessment; For the interpolated samples of the gradient penalty term, and , The interpolation coefficients are and satisfy the following conditions: , This is the gradient penalty coefficient, whose value can be determined based on model training stability, convergence speed, and generated data quality. In this embodiment, preferably, ; The discriminator's expected output is for real samples, measuring the discriminator's average score on real data; To generate the discriminator's expected output for the samples, measure the discriminator's average score on the pseudo-data; This is a gradient penalty term used to constrain the gradient norm of the discriminator at the interpolated samples to be close to 1. For the discriminator to interpolate samples The gradient L2 norm.

[0057] Step 3.2: Construct a scene generation model using a conditional generative adversarial network framework, and introduce scene labels as conditional inputs into the generator and discriminator; at the same time, embed a physical constraint loss term into the model's total loss function, and incorporate the correlation between source load and meteorological physics under extreme scenarios (such as high temperature scenarios corresponding to high load, and windless and lightless scenarios corresponding to low new energy output) as a constraint penalty term into the training process to ensure that the generated data is statistically realistic while strictly satisfying the physical laws and boundary constraints under extreme scenarios; For the variable relationships in extreme scenarios, the physical constraint loss is broken down into three computable sub-items, which are directly embedded into the loss function, specifically: (3) in, Losses due to load matching with weather conditions, To match the losses incurred in generating new energy sources with meteorological conditions, Losses are due to supply and demand imbalance constraints.

[0058] Furthermore, each loss function term can be determined in the following way: The load and weather matching loss (For example, high temperatures can lead to high loads), specifically: (4) in, For the number of time steps, for Temperature at all times The high temperature threshold can be adjusted according to the region; in this embodiment, it is preferably 35°C. for Constant load, The high-temperature load threshold is set when the temperature is higher than a certain value. The 90th percentile of the load during historically high-temperature periods; It is a proportionality coefficient, and satisfies This is used to adjust the degree of influence of temperature and load deviation on losses, and can be determined according to the regional load and the strength of temperature correlation. In this embodiment, it is preferably 0.1. Let be the fault tolerance coefficient, and satisfy . This is used to add tolerance space to physical constraints, avoiding excessive penalties and numerical instability during model training. In this embodiment, it is preferably 0.05; when and When this occurs, the loss value increases, and this loss item is used to penalize unreasonable scenarios such as "high temperature and low load".

[0059] The loss due to the matching of new energy output with weather conditions (For example, low wind speeds can lead to low wind power output), specifically: (5) in, for Wind speed at all times The minimum wind speed threshold for wind power is preferably 3 m / s; for Wind power output at all times To determine the low wind power output threshold, a wind speed lower than [missing information] is used. The 10th percentile of wind power output during historically low wind speed periods; for Constant light intensity The preferred value is the low light intensity threshold for photovoltaic applications. , for Photovoltaic power output at all times To determine the low photovoltaic output threshold, the light intensity is taken as below... The 10th percentile of photovoltaic output during historically low sunshine periods; It is a proportionality coefficient, and satisfies In this embodiment, the value is preferably 0.02. Let be the fault tolerance coefficient, and satisfy . This is used to add tolerance space to physical constraints and avoid excessive penalties and numerical instability during model training. In this embodiment, it is preferably 0.05. This loss term is used to penalize scenarios that violate physical laws, such as "low wind speed and high wind power output" and "low sunlight and high photovoltaic power output".

[0060] The supply and demand balance constraint loss (For example, in a dark evening rush hour scenario, the constraint is that the load ramp-up rate must meet the standard), specifically: (6) in, Number of scene types (7 categories). For the first Load ramp-up rate for similar scenarios (e.g., ramp-up rate during the evening peak hours when there is no light, from 18:00 to 22:00). For the first The ramp-up rate threshold for similar scenarios is determined based on the 85th percentile of the ramp-up rate for similar scenarios in history. The tolerance range for the ramp rate, and satisfying the following conditions. In this embodiment, preferably, This loss term is used to ensure that the supply and demand trends in extreme scenarios are consistent with reality (e.g., the load ramp-up rate in the evening peak no-light scenario must be higher than the normal rate).

[0061] Furthermore, the regularization loss (L2 regularization, to prevent model overfitting), determined as follows: (7) in, , These are the sets of all trainable parameters for the generator and discriminator, respectively, with the parameter size constrained by the L2 norm. is a single trainable parameter in a neural network.

[0062] Step 4: Train the model based on historical data and corresponding scene labels for each operating scenario, and use the trained model to generate data for each operating scenario; Step 5: The generated data for each running scenario are evaluated using a three-dimensional index system including structural similarity index, KL divergence and mean square error, and the running scenario generation model is optimized based on the evaluation results.

[0063] Specifically, the steps include the following: In scene generation, various statistical metrics can be used to evaluate the performance and output quality of the generative model. First, the Structural Similarity Index (SSIM) is a commonly used metric to measure the similarity between real small sample data and generated scene data, suitable for assessing the structural similarity between generated and target data. Second, the Kullback-Leibler (KL) divergence measures the difference between two probability distributions and can be used to analyze the deviation between the distribution of generated data and the distribution of real data. Furthermore, Pearson correlation coefficient analysis can be used to assess the linear correlation between various internal data. To comprehensively analyze the results of Pearson correlation coefficient analysis, mean squared error (MSE) is used. By calculating the average of the squared differences between the Pearson analysis results of generated scene data and real scene data for each metric, a method to quantify the overall performance of the model is provided, improving the accuracy and practicality of scene generation.

[0064] Step 5.1: Calculate the Structural Similarity Index (SSIM) value; The Structural Similarity Index Measure (SSIM) is a metric used to measure the similarity between images, and it is widely used in image quality assessment and image processing. By comparing the brightness, contrast, and structural features of images, SSIM can more closely approximate human visual perception, compensating for the shortcomings of mean squared error (MSE) and peak signal-to-noise ratio (PSNR).

[0065] In this invention, the structural similarity index is creatively applied to the trend structure verification of power time series data. By comparing the peak and valley positions, ramp rate, and fluctuation rhythm of the curves, the structural consistency between the generated data and the real data is verified.

[0066] The local structural similarity value between historical sliding window data and generated sliding window data is determined as follows: (8) Among them, subscript Data within a sliding window representing historical power time-series data, index The sliding window data represents the data generated in the same scene and at the same time as the historical power time series data. The two are paired comparison data groups. In this embodiment, it is further preferred to obtain the data by a sliding window based on a window of 1 hour (containing 4 sampling points) and a step size of 1 sampling point. It is the local mean of the data within the historical window, representing the average load level of the historical electricity load time series data within that window; To generate a local mean of the data within the window, representing the average load level of the generated electrical load time series data within that window; The local variance of the data within the historical window; To generate the local variance of the data within the window, The local covariance between the data within the historical window and the data within the generated window is used to reflect the degree of linear correlation between the two data trends; , All are stability constants, and satisfy... , This parameter is used to avoid denominators of zero in the formula and to reduce the distortion of calculation results caused by excessively small local means and variances of the data. It is a fixed correction parameter in the calculation of structural similarity index, with no requirements for scenario adaptation adjustment. It is only used to ensure the effectiveness of the formula calculation. In this embodiment, preferably, The value is 0.001. The value is 0.0009.

[0067] In this embodiment, the calculation process of the structural similarity index value specifically includes: Step 5.1.1: Select input data. In normal scenarios, select 30 sets of historical normal day "load and new energy output time series data" and pair them one by one with 30 sets of generated normal day data. In extreme scenarios, pair them according to 6 scenarios (such as "no light during evening peak" scenario, input the historical 18:00-22:00 load ramp-up curve of this scenario and the generated curve of the same scenario). Step 5.1.2: Set up a sliding window, calculate the local structural similarity index value of each window, and take the average value of all windows as the final structural similarity index value; Preferably, the window size is 1 hour (including 4 sampling points), and the step size is 1 sampling point; Substitute into formula (8) to obtain the SSIM value of each window, and take the average value of all windows as the final structural similarity index value; Step 5.1.3: Determine the validity of the trend structure; In normal scenarios, when the final structural similarity index value is ≥0.85, the structures are considered to be highly consistent and in line with daily operating trends. In extreme scenarios, when the final structural similarity index value is ≥0.80, the structure is considered to be highly consistent and in line with the daily operation trend; (extreme data is scarce, so the requirements can be relaxed appropriately, but core trends such as "evening peak climbing" and "continuous low output" must be consistent).

[0068] In one specific implementation, when the input is the historical load time series of the scenario from 18:00 to 22:00 (with the ramp rate gradually increasing) and the generated load time series of the same period; when the output calculates the final structural similarity index value = 0.83 (≥0.80), it indicates that the ramp rhythm and peak occurrence time of the generated curve are consistent with the history, and the structure is effective.

[0069] Step 5.2: Calculate the KL divergence; Kullback-Leibler divergence (KL Divergence) is an asymmetric metric used to measure the difference between two probability distributions. In information theory, KL divergence is used to describe the difference between two probability distributions. To another distribution Information loss. It is often used in machine learning, such as for the optimization of probabilistic models, distribution fitting, and variational inference.

[0070] The KL divergence is defined for two probability distributions. and

[0071] (9) For continuous distributions: (10) in, The true distribution (usually the target distribution); Approximate distribution or assumed distribution; Indicates from distribution To distribution Information loss.

[0072] Based on the definition of "measuring the difference between two probability distributions" in the selected content, the KL divergence in this invention does not evaluate the distribution of the entire data, but rather calculates it for key indicators of power system dispatch, ensuring that the statistical characteristics of the generated scenario are consistent with reality.

[0073] In this embodiment, the KL divergence calculation process specifically includes: Step 5.2.1: Select core feature indicators based on different operating scenarios; In normal scenarios, the core characteristic indicators selected include the average daily load, the daily fluctuation rate of wind power output, the midday peak value of photovoltaic power, and the correlation coefficient between temperature and load. In extreme scenarios, different indicators are selected for different extreme operating conditions. For example, in extremely hot and windless scenarios, the core characteristic indicators selected include the maximum daily load and the average daily wind power output; in evening peak and no light scenarios, the core characteristic indicators selected include the load ramp-up rate from 18:00 to 22:00; and in continuous windless and no light scenarios, the proportion of extremely low wind and solar power output is selected.

[0074] Step 5.2.2: For each core indicator of the historical data, construct the true distribution using kernel density estimation (KDE). For the same indicator in the generated data, construct a corresponding approximate distribution. ;Calculate the KL divergence of each index according to the continuous distribution KL divergence formula, i.e. formula (10); For each type of operating scenario, the average value of the KL divergence of all core indicators in that scenario is calculated and used as the final KL divergence index value for that scenario.

[0075] Step 5.2.3: Set differentiated thresholds according to the scenario type to verify the consistency of data distribution; For normal scenarios: when the final KL divergence index value is less than or equal to 0.3, it is determined that the generated data has a small difference in distribution from the real data and consistent statistical characteristics; For extreme scenarios: When the final KL divergence index value is less than or equal to 0.4, it is determined that the generated data has a small difference from the distribution of the real data and the statistical characteristics are consistent; to adapt to the large fluctuations of extreme scenarios, reasonable deviations are allowed.

[0076] In one specific implementation, the maximum daily load under extreme scenarios is used as the input index. Historical data distribution and generated data distribution are constructed respectively. The final KL divergence index value is calculated to be 0.29 (≤0.4), which verifies that the load peak distribution of the generated data is consistent with the historical extreme scenarios, conforms to the statistical law of "high temperature corresponds to high load", and ensures the statistical fidelity of the generated data.

[0077] Step 5.3: Calculate the mean square error; Mean Squared Error (MSE) is a commonly used metric to measure the difference between predicted and actual values ​​of data. It is the average of the squares of the prediction error (i.e., the difference between the predicted and actual values).

[0078] The formula for calculating the mean square error is: (11) in, It is the number of observation points. It is the first The actual value of each observation point, and It is the first Predicted values ​​for each observation point.

[0079] Mean squared error (MSE) is an important tool for evaluating model performance and is widely used in regression analysis in fields such as statistics, machine learning, and data science. It provides a simple way to quantify the predictive accuracy of a model, and is particularly suitable for optimizing linear regression and other algorithms that minimize the squared error.

[0080] Based on the definition of the selected content "measures the mean square error between the predicted value and the actual value", this invention is extended to a dual purpose: to verify both the numerical error of a single indicator and the stability of the physical correlation.

[0081] In this embodiment, specifically, the mean square error calculation process includes: Step 5.3.1: Evaluation of numerical error of a single indicator; Data input: Extract numerical sequences of the same scenario and the same indicator from historical and generated data (such as the "midday peak value of photovoltaic power" sequence in normal scenarios). Calculation logic: Substitute the above numerical sequence into the mean square error formula (11) to calculate the numerical error of the index; Judgment criteria: For normal scenarios, when the following conditions are met... At that time, it is determined that the generated data is accurate; for extreme scenarios, when the following conditions are met... At that time, it was determined that the generated data was accurate; among them, This represents the standard deviation of the corresponding historical indicator value series.

[0082] Step 5.3.2: Physical association stability assessment; Data input: Select "meteorological factors-source-load power" pairing indicators with clear physical correlation in the power system, such as typical correlation pairs like temperature-load and wind speed-wind power output; Calculation logic: Calculate the Pearson correlation coefficient for each of the associated pairs in the historical data. The Pearson correlation coefficient of the associated pairs in the generated runtime scenario data. Based on the aforementioned correlation coefficient, the deviation between the two is quantified using MSE, and the calculation method is as follows: ; Judgment criteria: When the following conditions are met At that time, it was determined that the physical relationship of the generated scene data was stable and there were no logical contradictions.

[0083] In one specific implementation, the input data consists of historical normal day ambient temperature time-series data and load power time-series data (corresponding to historical correlation coefficients). Generate ambient temperature time-series data and load power time-series data for normal days (and generate corresponding correlation coefficients). Calculations show that the single-index numerical error MSE = 0.007, which meets the requirements. Normal scenario judgment conditions; physical correlation deviation ,satisfy The correlation stability determination condition verifies the numerical accuracy and physical logic consistency of the generated scene data, indicating that both numerical and physical correlations are valid.

[0084] Step 5.4: Conduct a collaborative evaluation of the three-dimensional indicators; Evaluation order: First, verify "distribution compliance" using KL divergence; then verify "structural consistency" using SSIM; and finally verify "numerical and correlation stability" using MSE. For each indicator exceeding the standard, the following actions will be taken: (1) Handling of excessive KL divergence: When the KL divergence exceeds the scene threshold, adjust the weight coefficient of the adversarial loss term of CGAN in step 3 to optimize the distribution fitting ability. (2) Handling of SSIM exceeding limits: When the SSIM metric exceeds the scenario threshold: strengthen the penalty for physical constraint loss in step 3. Specifically, the weight coefficient of physical constraint loss can be increased to improve the temporal structure consistency of the generated data. (3) Handling of MSE exceeding the standard: When the numerical MSE or physical correlation MSE exceeds the threshold, check the preprocessing effect or model parameters in step 1, including but not limited to the following operations: Adjust the outlier identification threshold of the 3σ criterion or box plot, and re-remove outliers such as sensor malfunctions and transient pulses; Optimize missing value imputation methods and adopt interpolation methods that are more in line with physical laws (such as time series interpolation and physical constraint interpolation) to improve data integrity; Improve the spatiotemporal alignment accuracy between source load data and meteorological data, and revise the matching rules between timestamps and spatial area identifiers.

[0085] Finally, effective scenarios are selected, and only the generated data that meets the criteria of "SSIM compliance + KL divergence compliance + MSE (numerical + correlation) compliance" are retained to form the final set of usable scenarios.

[0086] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0087] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0088] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0089] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for generating load operation scenarios based on boundary condition constraints, characterized in that, The method includes the following steps: Step 1: Collect historical source load data and corresponding meteorological data at the same time, and perform preprocessing; Step 2: Divide the operating scenarios into extreme scenarios and normal scenarios, and set scenario labels; classify the preprocessed historical data based on the divided operating scenarios to obtain historical data under each operating scenario; Step 3: Construct a CGAN-based scenario generation model, introduce scenario labels as conditional inputs into the generator and discriminator, and construct a physical constraint loss function that includes load and weather matching loss, new energy output and weather matching loss, and supply and demand balance constraint loss. Step 4: Train the model based on historical data and corresponding scene labels for each operating scenario, and use the trained model to generate data for each operating scenario; Step 5: The generated data for each running scenario are evaluated using a three-dimensional indicator system, including structural similarity index, KL divergence, and mean squared error. The running scenario generation model is then optimized based on the evaluation results.

2. The load operation scenario generation method based on boundary condition constraints according to claim 1, characterized in that: In step 2, the extreme scenarios include extremely hot and windless conditions, evening peak hours without sunlight, continuous windless and sunless conditions, shortage of primary energy supply, sudden changes in new energy output, and continuous low wind and solar power output; the normal scenarios are all other scenarios except for the extreme scenarios.

3. The load operation scenario generation method based on boundary condition constraints according to claim 1, characterized in that: The CGAN-based scenario generation model has a total loss function that is a weighted sum of adversarial loss, physical constraint loss, and regularization loss, specifically: in, To combat the losses; This is the loss due to physical constraints, used to ensure physical compliance; This is the regularization loss; The weighting coefficients are dynamically adjusted based on the model distribution fitting requirements to counteract the loss. , These are the weighting coefficients for physical constraint loss and regularization loss, respectively, which are determined based on the priority of physical constraints and the generalization requirements of the model.

4. The method for generating load operation scenarios based on boundary condition constraints according to claim 1, characterized in that: In step 3, the load-meteorological matching loss is determined as follows: in, Losses due to load matching with weather conditions; The number of time steps; for Temperature at any given time; This is the high-temperature threshold, which can be adjusted by region; for Constant load; The high-temperature load threshold is set when the temperature is higher than a certain value. The 90th percentile of the load during historically high-temperature periods; It is a proportionality coefficient, and satisfies This is used to adjust the degree of influence of temperature and load deviation on losses; Let be the fault tolerance coefficient, and satisfy . It is used to add tolerance space to physical constraints and avoid excessive penalties and numerical instability during model training.

5. The method for generating load operation scenarios based on boundary condition constraints according to claim 1, characterized in that: In step 3, the loss due to matching the power output of the new energy source with weather conditions is determined in the following manner: in, To match the losses incurred in generating new energy sources with meteorological conditions, for Wind speed at all times; The low wind speed threshold for wind power; for Wind power output at all times; To determine the low wind power output threshold, a wind speed lower than [missing information] is used. The 10th percentile of wind power output during historically low wind speed periods; for Light intensity at any given time; This represents the low light intensity threshold for photovoltaic applications. for Photovoltaic power output at all times; For low photovoltaic output threshold, the light intensity is taken as below The 10th percentile of photovoltaic output during historically low sunshine periods; It is a proportionality coefficient, and satisfies ; Let be the fault tolerance coefficient, and satisfy . .

6. The load operation scenario generation method based on boundary condition constraints according to claim 2, characterized in that: In step 3, the supply and demand balance constraint loss is determined in the following manner: in, Losses are constrained by supply and demand imbalance; Number of scene types For the first Load ramp-up rate for similar scenarios; For the first The ramp-up rate threshold for similar scenarios is determined based on the 85th percentile of the ramp-up rate for similar scenarios in history. The tolerance range for the ramp rate, and satisfying the following conditions. .

7. The method for generating load operation scenarios based on boundary condition constraints according to claim 1, characterized in that: In step 5, the generated runtime scenario data is evaluated using structural similarity metrics, specifically including: For historical electrical load time series data and generated electrical load time series data of the same scenario and time period, a sliding window is used to extract the data, resulting in a one-to-one corresponding historical sliding window dataset and generated sliding window dataset. The local structural similarity index value of each sliding window is calculated, and the average value of the local structural similarity index values ​​of all sliding windows is taken as the final structural similarity index value of the generated running scenario data. The local structural similarity index value is determined in the following manner: in, For data within the historical sliding window, Generate the corresponding data within the sliding window; , They are respectively , The local mean; , They are respectively , The local variance, for and Local covariance; , All are stability constants, and satisfy... , .

8. A load operation scenario generation system based on boundary condition constraints guided by the method of any one of claims 1-7, comprising a data acquisition module, a scenario division and historical data classification module, an operation scenario generation model construction module, an operation scenario data generation module, and an operation scenario generation data evaluation module, characterized in that: The data acquisition module is used to collect historical source load data and meteorological data at the same time, and to perform preprocessing. The scenario segmentation and historical data classification module is used to divide the running scenarios into extreme scenarios and normal scenarios, and to set scenario labels; The preprocessed historical data is classified based on the divided operating scenarios to obtain historical data under each operating scenario; The runtime scenario generation model building module is used to build a runtime scenario generation model based on CGAN. It introduces scenario labels as conditional inputs into the generator and discriminator, and constructs a physical constraint loss function that includes load and weather matching loss, new energy output and weather matching loss, and supply and demand balance constraint loss. The runtime scenario data generation module is used to train the model based on historical data and corresponding scenario labels in each runtime scenario, and to generate runtime scenario data using the trained model. The runtime scenario generation data evaluation module is used to evaluate the generated runtime scenario data using a three-dimensional index system including structural similarity index, KL divergence and mean squared error, and to optimize the runtime scenario generation model based on the evaluation results.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-7.

Images