PC-GAN-based power consumption small sample enhanced prediction method, system and equipment for extreme scene

Abstract

A PC-GAN-based power consumption small sample enhanced prediction method, system and device for an extreme scene, and the method comprises the steps: firstly obtaining a power consumption load-meteorological data set of a target region, recognizing an extreme scene sample, and constructing a multi-dimensional feature vector; then, constructing a physical constraint generative adversarial network architecture model, and implementing a three-stage constraint strategy by taking the multi-dimensional feature vector as input to obtain a high-quality synthetic sample; and finally, combining the extreme scene sample and the high-quality synthetic sample into an enhanced training data set so as to realize electrical load prediction of the target region. In the application of the design, through a three-stage constraint implementation strategy and a physical and business dual constraint mechanism, and by matching with a scene classification auxiliary task, a synthetic sample is ensured to be in accordance with a physical mechanism and a business rule, scene consistency can be maintained, and step-by-step correction and layered quality inspection are performed by means of a multi-dimensional evaluation index system, so that the quality of a product is improved. And high quality and diversity of data are fully guaranteed.

Description

Technical Field

[0001] This invention relates to data processing methods, belonging to the field of data processing technology, and particularly to a method, system, and device for small-sample augmented prediction of electricity consumption in extreme scenarios based on PC-GAN. Background Technology

[0002] Although extreme scenarios occur with a low probability during the operation of the power system, their impact on the safe and stable operation of the power grid and the reliability of power supply is extremely significant. These low-probability, high-impact events cover a variety of types, including extreme weather, major emergencies, and power grid failures. Their core characteristic is data scarcity. Furthermore, due to the extremely small sample size of historical data for extreme scenarios, it is difficult to meet the training requirements of deep learning models, which leads to a severe imbalance in the data distribution between normal and extreme scenarios.

[0003] Traditional statistical methods and machine learning models suffer from insufficient generalization ability under small sample conditions and are prone to overfitting. Furthermore, the scarcity of training samples in extreme scenarios further weakens the robustness of risk prediction models, making it difficult to effectively support decision-making.

[0004] Existing simple sampling methods perform only linear interpolation operations in the feature space, failing to accurately capture the complex nonlinear features inherent in extreme scenarios. Samples generated by traditional generative adversarial networks lack physical interpretability and may contradict actual business logic and the operational laws of power systems. Numerical simulation methods rely on precise physical model construction, which is not only computationally expensive but also struggles to comprehensively cover various extreme cases. Therefore, there is an urgent need for a method to generate high-quality power system synthesis samples for extreme scenarios that conform to physical laws and business constraints, in order to address the aforementioned shortcomings of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned defects and problems in the prior art and provide a PC-GAN-based method, system and device for small sample augmentation prediction of electricity consumption in extreme scenarios, so as to generate high-quality samples and meet the prediction requirements.

[0006] To achieve the above objectives, the technical solution of this invention is: a small-sample augmentation prediction method for electricity consumption in extreme scenarios based on PC-GAN, comprising:

[0007] Obtain the electricity load-meteorological dataset for the target region, identify extreme scenario samples in the electricity load-meteorological dataset, and construct a multidimensional feature vector based on the electricity load-meteorological dataset;

[0008] Construct a physically constrained generative adversarial network (PC-GAN) model that includes a generator, discriminator, and scene classifier;

[0009] The PC-GAN model, based on the physical constraint generative adversarial network architecture, implements a three-stage constraint strategy with multi-dimensional feature vectors as input to obtain high-quality synthetic samples.

[0010] Extreme scenario samples and high-quality synthetic samples are merged into an enhanced training dataset, and a load prediction model is trained based on the enhanced training dataset to achieve power load prediction for the target area.

[0011] Optionally, the implementation of the three-stage constraint strategy specifically includes:

[0012] During the training phase, constraints are set for general physical loss, specific scenario loss, feature distribution matching loss, temporal consistency loss, generator total loss, and discriminator total loss. The adversarial network architecture model is trained alternately to learn samples and obtain initial candidate samples.

[0013] During the generation phase, for the initial candidate samples, the general physical loss, power consistency constraint, and specific scenario loss are subjected to progressively forced lightweight post-processing corrections and decentralized processing to obtain post-processing candidate samples.

[0014] During the validation phase, for post-processing candidate samples, the pass rate of each constraint and the overall quality score are calculated to perform stratified quality checks and obtain high-quality synthetic samples.

[0015] Optionally, the general physical loss includes:

[0016] Temperature Relationship Constraints Its loss function is as follows:

[0017] ;

[0018] in: The lowest temperature; Average temperature; This is the highest temperature;

[0019] Rainfall Relationship Constraints Its loss function is as follows:

[0020] ;

[0021] in: This represents the maximum rainfall. This refers to cumulative rainfall.

[0022] Wind speed relationship constraints Its loss function is as follows:

[0023] ;

[0024] in: Average wind speed; Maximum wind speed;

[0025] Humidity Relationship Constraints Its loss function is as follows:

[0026] ;

[0027] in: Humidity range;

[0028] Power Consistency Constraint Its expression is as follows:

[0029] ;

[0030] in: This represents the total amount of electricity consumed during the time-of-use period. Daily electricity consumption; Assign time segment numbers;

[0031] The specific scenario loss includes:

[0032] Business constraints in extreme high temperature scenarios Its loss function is as follows:

[0033] ;

[0034] in: This is the highest temperature;

[0035] Temperature-load correlation constraints Its loss function is as follows:

[0036] ;

[0037] ;

[0038] in: This is the sum of the time-of-use electricity consumption; Temperature sensitivity coefficient; Humidity sensitivity coefficient; This is the temperature response function; As the baseline load; This is the actual temperature; For reference temperature;

[0039] Business constraints in extreme rainstorm scenarios Its loss function is as follows:

[0040] ;

[0041] in: Daily rainfall;

[0042] Rainfall-load correlation constraints Its loss function is as follows:

[0043] ;

[0044] in: This is the rainfall load suppression coefficient; for Rainfall at any given moment; This is for additional power load during heavy rain scenarios;

[0045] Business constraints in extreme cold wave scenarios Its loss function is as follows:

[0046] ;

[0047] in: The temperature drop over 24 hours;

[0048] Cold wave-load correlation constraints Its loss function is as follows:

[0049] ;

[0050] in: This refers to the cold wave load coefficient. For reference temperature; This is the coefficient representing the impact of the temperature drop on the load.

[0051] Optionally, the feature distribution matching loss Its loss function is as follows:

[0052] ;

[0053] in: , The first The mean of real and synthetic samples with 3D features; , The first The standard deviation of real and synthetic samples for 3D features; For feature dimensions; The index of the feature dimension;

[0054] The timing consistency loss Its loss function is as follows:

[0055] ;

[0056] in: This represents the average load variation pattern of the real sample.

[0057] The generator's total loss Its loss function is as follows:

[0058] ;

[0059] ;

[0060] ;

[0061] in: The adversarial loss function is the generator loss; The scene classification loss function; Number of scene categories; Labeling based on real-world scenarios; For scene category number; For mathematical expectation; For discriminator For fake samples With conditions The output;

[0062] The total loss of the discriminator Its loss function is as follows:

[0063] ;

[0064] ;

[0065] in: The adversarial loss function is the discriminator loss. For discriminator For real samples With conditions The output; Let be the mathematical expectation.

[0066] Optionally, during the generation phase, for the initial candidate samples, a progressively forced lightweight post-processing correction is performed to adjust the general physical loss, power consistency constraint, and specific scenario loss. This specifically includes:

[0067] Level 1 corrections are mandatory corrections to general physical relations, including:

[0068] The following expression is used to force corrections to samples that violate the temperature magnitude relationship:

[0069] ;

[0070] ;

[0071] ;

[0072] in: , , These are correction values ​​for the minimum temperature, maximum temperature, and average temperature, respectively.

[0073] The sample that violates the wind speed relationship is forcibly corrected, and its expression is as follows:

[0074] ;

[0075] in: This is a correction value for the average wind speed;

[0076] The sample that violates the rainfall relationship is forcibly corrected, and its expression is as follows:

[0077] ;

[0078] in: This is a correction value for the maximum rainfall.

[0079] The forced correction for samples that violate the humidity range is expressed as follows:

[0080] ;

[0081] in: This is a correction value for the humidity range;

[0082] The second level of correction addresses energy consistency by calculating the total hourly energy consumption for each sample and applying a forced correction to the deviation from the daily energy consumption. The expression is as follows:

[0083] ;

[0084] in: This is a correction value for the total time-of-use electricity consumption.

[0085] Level 3 corrections are mandatory corrections for specific scenarios, including:

[0086] For extreme high-temperature scenarios, if the maximum temperature is not within the range of 35-40℃, a forced correction is applied, expressed as follows:

[0087] ;

[0088] in: This is a correction value for extreme high temperatures;

[0089] For samples with temperature-load correlation anomalies, the theoretical load value is calculated using the following expression:

[0090] ;

[0091] in: This represents the theoretical load value for temperature-load correlation anomalies.

[0092] If the actual load deviates from the theoretical load by more than 20%, a weighted average is used for mandatory correction, the expression of which is as follows:

[0093] ;

[0094] in: This is a correction value for the theoretical load value of the temperature-load correlation anomaly;

[0095] For heavy rain scenarios, if the daily cumulative rainfall is less than 50 mm, a mandatory correction is applied, expressed as follows:

[0096] ;

[0097] in: This is a corrected value for the daily cumulative rainfall;

[0098] For samples with anomalies in the rainfall-load correlation, the theoretical load value is calculated using the following expression:

[0099] ;

[0100] in: This represents the theoretical load value for the rainfall-load correlation anomaly. This is the rainfall saturation threshold;

[0101] If the actual load deviates from the theoretical load by more than 25%, a weighted average is used for mandatory correction, the expression of which is as follows:

[0102] ;

[0103] in: This is a correction value for the theoretical load value of the rainfall-load correlation anomaly;

[0104] For cold wave scenarios, ensure the minimum temperature does not exceed 4℃, as expressed below:

[0105] ;

[0106] in: This is a correction value for the lowest temperature;

[0107] For samples with cold wave-load correlation anomalies, the theoretical load value is calculated using the following expression:

[0108] ;

[0109] in: This represents the theoretical load value for the cold wave-load correlation anomaly. The cooling rate sensitivity coefficient;

[0110] If the actual load deviates from the theoretical load by more than 20%, a weighted average is used for mandatory correction, the expression of which is as follows:

[0111] ;

[0112] in: This is a correction value for the theoretical load value of the cold wave-load correlation anomaly.

[0113] Optionally, the step of calculating the pass rate of each constraint for post-processing candidate samples during the verification phase specifically includes:

[0114] Pass rate for calculating temperature relationships Its expression is as follows:

[0115] ;

[0116] in: This represents the total number of temperature samples.

[0117] Calculating the pass rate of rainfall relationship Its expression is as follows:

[0118] ;

[0119] in: This represents the total number of rainfall samples.

[0120] Calculate the pass rate in relation to wind speed Its expression is as follows:

[0121] ;

[0122] in: This represents the total number of wind speed samples.

[0123] Calculate the power consistency pass rate Its expression is as follows:

[0124] ;

[0125] in: This represents the total number of electricity samples.

[0126] Calculate humidity range pass rate Its expression is as follows:

[0127] ;

[0128] in: This represents the total number of humidity samples.

[0129] Based on the pass rates of temperature, rainfall, wind speed, electrical consistency, and humidity range, the overall pass rate of general physical relationships is calculated. Its expression is as follows:

[0130] ;

[0131] Based on defined business constraints for extreme high temperatures, heavy rain, and cold waves, the corresponding pass rates are calculated, specifically including:

[0132] Calculate the pass rate of business constraints in extreme high temperature scenarios Its expression is as follows:

[0133] ;

[0134] in: This represents the total number of samples from extreme high temperatures.

[0135] Calculate temperature-load associated throughput Its expression is as follows:

[0136] ;

[0137] in: Total number of temperature-load samples;

[0138] Calculate the overall pass rate of business constraints in extreme high temperature scenarios. Its expression is as follows:

[0139] ;

[0140] Calculate the pass rate of business constraints in extreme rainstorm scenarios Its expression is as follows:

[0141] ;

[0142] in: This represents the total number of samples from rainstorm scenarios.

[0143] Calculate the rainfall-load correlation pass rate Its expression is as follows:

[0144] ;

[0145] in: This represents the total number of rainfall-load samples.

[0146] Calculate the overall pass rate of business constraints in extreme rainstorm scenarios. Its expression is as follows:

[0147] ;

[0148] Calculate business constraints in extreme cold wave scenarios Its expression is as follows:

[0149] ;

[0150] in: This represents the total number of samples from cold wave scenarios;

[0151] Calculate the cold wave-load correlation pass rate Its expression is as follows:

[0152] ;

[0153] in: This represents the total number of cold wave-load samples.

[0154] Calculate the overall pass rate of business constraints in extreme cold wave scenarios. Its expression is as follows:

[0155] .

[0156] Optionally, the calculation method for the comprehensive quality score specifically includes:

[0157] The distribution coverage of generated samples and real samples in the post-processing candidate samples is calculated, and the feature space is divided into grids. The grid cells covered by real samples and generated samples are statistically analyzed. The expression is as follows:

[0158] ;

[0159] in: For distribution coverage; To generate a set of grid cells covering the sample; The set of grid cells covered by the real sample;

[0160] Calculating the Wasserstein distance measures the similarity of distributions. Its expression is as follows:

[0161] ;

[0162] in: The distribution of the real samples; The distribution of the generated samples; It is a joint distribution; for Real Samples With generated samples Expected distance; To obtain the minimum expected distance under all possible joint distributions;

[0163] Calculate the matching degree between real samples and generated samples. Its expression is as follows:

[0164] ;

[0165] in: For the first real sample Step moment; For the generation of the sample Step moment;

[0166] Calculate the overall pass rate Its expression is as follows:

[0167] ;

[0168] in: The total number of samples; To constrain the total number; For the first The number of times a sample violates the constraint;

[0169] A weighted comprehensive quality score is calculated based on the overall pass rate of general physical relationships, the overall pass rate of specific scenario constraints, the distribution coverage, and the matching degree. Its expression is as follows:

[0170] ;

[0171] in: The overall pass rate for general physical relationships; Constrain the overall pass rate for a specific scenario; The normalized index for the Wasserstein distance is... Distance between the real sample and the generated sample distribution; This represents the upper limit of the distance.

[0172] A PC-GAN-based small-sample augmentation prediction system for extreme scenarios' electricity consumption, which applies the aforementioned method, includes:

[0173] The multidimensional feature vector construction module is used to acquire the electricity load-meteorological dataset of the target area, identify extreme scene samples in the electricity load-meteorological dataset, and construct multidimensional feature vectors based on the electricity load-meteorological dataset.

[0174] The network architecture model building module is used to construct a physically constrained generative adversarial network (PC-GAN) model that includes a generator, a discriminator, and a scene classifier.

[0175] The synthetic sample acquisition module is used to implement a three-stage constraint strategy on the PC-GAN model based on physical constraints, using multi-dimensional feature vectors as input to obtain high-quality synthetic samples.

[0176] The data augmentation prediction module is used to merge extreme scenario samples with high-quality synthetic samples into an augmented training dataset, and to train a load prediction model based on the augmented training dataset to achieve power load prediction for the target area.

[0177] An extreme scenario power consumption small sample augmentation prediction device based on PC-GAN, the device includes a processor and a memory;

[0178] The memory is used to store computer program code and to transmit the computer program code to the processor;

[0179] The processor is used to execute the above-described PC-GAN-based extreme scenario power consumption small sample augmentation prediction method according to the instructions in the computer program code.

[0180] A computer-readable storage medium storing computer-executable instructions, which, when executed on a computer, implement the aforementioned PC-GAN-based method for small-sample power consumption enhancement prediction in extreme scenarios.

[0181] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0182] This invention discloses a PC-GAN-based method, system, and device for small-sample augmented prediction of electricity consumption in extreme scenarios. The method first acquires an electricity load-meteorological dataset for the target region, identifies extreme scenario samples, and constructs a multi-dimensional feature vector. Then, it constructs a physically constrained generative adversarial network architecture model and implements a three-stage constraint strategy using the multi-dimensional feature vector as input to obtain high-quality synthetic samples. Finally, it merges the extreme scenario samples and the high-quality synthetic samples into an augmented training dataset to achieve electricity load prediction for the target region. In application, this design, through a three-stage constraint implementation strategy and a dual constraint mechanism of physical and business aspects, coupled with a scenario classification auxiliary task, ensures that the synthetic samples conform to both physical mechanisms and business rules while maintaining scenario consistency. Furthermore, it utilizes a multi-dimensional evaluation index system for step-by-step correction and hierarchical quality checks, fully guaranteeing the high quality and diversity of the data. Attached Figure Description

[0183] Figure 1 This is a flowchart of the method of the present invention.

[0184] Figure 2 This is a diagram of the physical constraint generative adversarial network model architecture in Embodiment 1 of the present invention.

[0185] Figure 3 This is an overall framework diagram of the three-stage constraint strategy in Embodiment 1 of the present invention.

[0186] Figure 4 This is a system structure diagram of the present invention.

[0187] Figure 5 This is a structural diagram of the device of the present invention.

[0188] In the diagram: 1. Multidimensional feature vector construction module; 2. Network architecture model construction module; 3. Synthetic sample acquisition module; 4. Data augmentation prediction module; 5. Processor; 6. Memory; 7. Computer program code. Detailed Implementation

[0189] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0190] Example 1:

[0191] See Figure 1 A small-sample augmentation prediction method for electricity consumption in extreme scenarios based on PC-GAN includes:

[0192] Obtain the electricity load-meteorological dataset for the target region, identify extreme scenario samples in the electricity load-meteorological dataset, and construct a multidimensional feature vector based on the electricity load-meteorological dataset;

[0193] In this embodiment, an electricity load-meteorological dataset is first obtained from historical data of the target region, and extreme scenario samples are identified from it. The identification criteria for extreme scenarios include multiple dimensions such as electricity load exceeding a statistical threshold, abnormal meteorological indicators, and abnormal electricity volatility. Specifically, when the electricity load exceeds the historical average plus a certain number of standard deviations, or when meteorological indicators such as temperature, wind speed, and rainfall exceed the normal range, or when the time derivative of the electricity load exceeds a set volatility threshold, the data for that period is marked as an extreme scenario sample.

[0194] After identifying extreme scenario samples, the data is standardized to map all feature values ​​to a uniform numerical range, eliminating the influence of units. The expression is as follows:

[0195] ;

[0196] in: Standardized values; These are the original eigenvalues; , These are the maximum and minimum values ​​of the original feature values, respectively;

[0197] Subsequently, feature engineering was performed to construct a multi-dimensional feature vector that includes time-series features, meteorological correlation features, and electricity consumption pattern features. The time-series features include periodic information such as week type and seasonal features. The meteorological correlation features include temperature, humidity, wind speed, and rainfall. The electricity consumption pattern features include derived indicators such as load change rate and cumulative electricity consumption.

[0198] Construct a Physics-Constrained-GAN (PC-GAN) model that includes a generator, discriminator, and scene classifier;

[0199] See Figure 2 In this embodiment, the generator receives random noise vectors and condition vectors as inputs and generates synthetic samples through multi-layer neural network transformation. The condition vectors contain prior information such as meteorological conditions and time features, which guide the generator to generate samples that conform to a specific scenario. The generator network adopts a multi-layer fully connected structure or a long short-term memory network structure, extracts and transforms features layer by layer, and finally outputs synthetic data with the same dimensions as the real samples.

[0200] The generator network adopts a three-layer fully connected structure with hidden layer dimensions of 128, 256, and 512. The forward propagation calculation of the network is as follows:

[0201] ; ;

[0202] ;

[0203] The discriminator receives real or synthetic samples and their corresponding conditional vectors. It employs a three-layer fully connected structure with a 20-dimensional input (15-dimensional features plus a 5-dimensional conditional vector) and outputs the probability of authenticity. Beyond distinguishing between real and synthetic samples, the discriminator also assists with a scene classifier task. Through a multi-task learning mechanism, it enhances the understanding of features from different extreme scenes, improving the scene consistency and feature representation accuracy of generated samples. The scene classification loss is:

[0204] ;

[0205] in: The scene classification loss function; Number of scene categories; Labeling based on real-world scenarios; For scene category number;

[0206] The network architecture incorporates two types of constraint mechanisms. The first type is general physical constraints, including constraints related to temperature, rainfall, wind speed, power consistency, and humidity range. These constraints are based on fundamental physical laws and data logic relationships and are applicable to all scenario types. The second type is scenario-specific constraints, set according to the specific physical mechanisms of different extreme scenarios. Examples include temperature-load nonlinear correlation constraints for extreme high-temperature scenarios, rainfall-load suppression constraints for heavy rain scenarios, and cooling rate-load response constraints for cold wave scenarios. These constraints ensure that the generated samples conform to the business logic and physical characteristics of the specific scenario.

[0207] The PC-GAN model, based on the physical constraint generative adversarial network architecture, implements a three-stage constraint strategy with multi-dimensional feature vectors as input to obtain high-quality synthetic samples.

[0208] See Figure 3 The three-stage constraint strategy refers to setting general physical loss, specific scene loss, feature distribution matching loss, temporal consistency loss, generator total loss and discriminator total loss as constraints during the training phase, and alternately training the adversarial network architecture model to learn samples and obtain initial candidate samples.

[0209] In this embodiment, a soft constraint strategy is implemented during the training phase, following the core principles of weight constraints, scenario learning, and data-driven learning. By introducing a weighted constraint term into the loss function, the generator gradually satisfies physical and business constraints while learning the data distribution, using gradient guidance rather than hard restrictions.

[0210] Furthermore, a comprehensive loss function is designed, comprising six components: adversarial loss, scene classification loss, general physical loss, scene-specific constraint loss, data distribution matching loss, and temporal consistency loss. The adversarial loss drives the game-like learning between the generator and discriminator; the scene classification loss enhances the model's ability to identify features from different extreme scenarios; the general physical loss prevents generated data from violating basic physical laws; the scene-specific loss ensures that generated samples conform to the business logic of a specific scenario; the data distribution matching loss guarantees that the statistical characteristics of generated samples are consistent with real data; and the temporal consistency loss maintains the temporal correlation of the load sequence.

[0211] Each loss term is balanced by a weight coefficient. The weight setting reflects the soft constraint concept, that is, the weight of the constraint term is moderate, which can guide the model's learning direction without excessively restricting the model's ability to explore the data space. The training process adopts an alternating optimization strategy. In each training batch, the discriminator parameters are updated first to improve the discrimination ability, and then the generator parameters are updated to improve the generation quality. Through repeated iterations, the generator gradually masters the ability to generate high-quality samples under the constraint conditions.

[0212] In this embodiment, an alternating optimization strategy is adopted, and each training batch is executed according to the following steps:

[0213] First, real samples with a batch size of 32 are sampled from the real dataset. Then, 32 100-dimensional noise vectors are sampled from the standard normal distribution. Finally, synthetic samples are generated using a generator.

[0214] Calculate the discriminator loss and update the discriminator parameters using the Adam optimizer, with the learning rate set to 0.0001 and the momentum parameter... =0.5, =0.999.

[0215] Then, with the discriminator parameters fixed, new noise vectors are sampled to generate synthetic samples. The total loss of the generator is calculated, and the generator parameters are updated using the Adam optimizer. The learning rate is set to 0.0002. The total number of training rounds is 5000, and the quality of the generated samples is evaluated every 500 rounds.

[0216] Furthermore, the general physical loss includes:

[0217] Temperature Relationship Constraints (Preset weight 0.8), the lowest temperature is no greater than the average temperature; the average temperature is no greater than the highest temperature; and the lowest temperature is no lower than -10℃, and the highest temperature is no higher than 40℃. The loss function is as follows:

[0218] ;

[0219] in: The lowest temperature; Average temperature; This is the highest temperature;

[0220] Rainfall Relationship Constraints (Preset weight 0.5), the maximum daily rainfall does not exceed the cumulative daily rainfall, and its loss function is as follows:

[0221] ;

[0222] in: This represents the maximum rainfall. This refers to cumulative rainfall.

[0223] Wind speed relationship constraints (With a preset weight of 0.5), the average wind speed is no greater than the maximum wind speed, and its loss function is as follows:

[0224] ;

[0225] in: Average wind speed; Maximum wind speed;

[0226] Humidity Relationship Constraints (Preset weight 0.5), with humidity between 0 and 1, the loss function is as follows:

[0227] ;

[0228] in: Humidity range;

[0229] Power Consistency Constraint (Preset weight 1.0), daily electricity consumption equals the sum of hourly electricity consumption, and its expression is as follows:

[0230] ;

[0231] in: This represents the total amount of electricity consumed during the time-of-use period. Daily electricity consumption; Assign time segment numbers;

[0232] The specific scenario loss includes:

[0233] Business constraints in extreme high temperature scenarios (Preset weight 0.8), generally assuming the highest temperature is between 35-40℃, the loss function is as follows:

[0234] ;

[0235] in: This is the highest temperature;

[0236] Temperature-load correlation constraints (With a preset weight of 0.6), its loss function is as follows:

[0237] ;

[0238] The temperature response function is presented in a piecewise form, and its expression is as follows:

[0239] ;

[0240] in: This is the sum of the time-of-use electricity consumption; For temperature sensitivity coefficient, = 0.8 megawatts per degree Celsius; Humidity sensitivity coefficient = 0.3 MW / percentage point; This is the temperature response function; As the baseline load; This is the actual temperature; For reference temperature;

[0241] Business constraints in extreme rainstorm scenarios (With a preset weight of 0.5), its loss function is as follows:

[0242] ;

[0243] in: 50 represents the daily rainfall; 50 is the rainstorm threshold (i.e., when the daily rainfall is ≥50mm, it is considered a rainstorm scenario).

[0244] Rainfall-load correlation constraints (With a preset weight of 0.5), its loss function is as follows:

[0245] ;

[0246] in: This is the rainfall load suppression coefficient; for Rainfall at any given moment; This is for additional power load during heavy rain scenarios;

[0247] Business constraints in extreme cold wave scenarios (With a preset weight of 0.5), its loss function is as follows:

[0248] ;

[0249] in: The temperature drop over 24 hours;

[0250] Cold wave-load correlation constraints (With a preset weight of 0.6), its loss function is as follows:

[0251] ;

[0252] in: This refers to the cold wave load coefficient. For reference temperature; This is the coefficient representing the impact of the temperature drop on the load.

[0253] in: This is the cold wave load coefficient. = 1.2 MW / °C (the increase in load for every 1°C drop in temperature below the reference temperature); For reference temperature; The coefficient representing the impact of the temperature drop on the load. = 0.3 megawatts per degree Celsius.

[0254] To enhance data authenticity, feature distribution matching loss is introduced. Its loss function is as follows:

[0255] ;

[0256] in: , The first The mean of real and synthetic samples with 3D features; , The first The standard deviation of real and synthetic samples for 3D features; For feature dimensions; The index of the feature dimension;

[0257] The timing consistency loss Its loss function is as follows:

[0258] ;

[0259] in: This represents the average load variation pattern of the real sample.

[0260] The generator's total loss Its loss function is as follows:

[0261] ;

[0262] ;

[0263] ;

[0264] in: The adversarial loss function is the generator loss; The scene classification loss function; Number of scene categories; Labeling based on real-world scenarios; For scene category number; For mathematical expectation; For discriminator For fake samples With conditions The output;

[0265] The total loss of the discriminator Its loss function is as follows:

[0266] ;

[0267] ;

[0268] in: The adversarial loss function is the discriminator loss. For discriminator For real samples With conditions The output; Let be the mathematical expectation.

[0269] During the generation phase, for the initial candidate samples, the general physical loss, power consistency constraint, and specific scenario loss are subjected to progressively forced lightweight post-processing corrections and decentralized processing to obtain post-processing candidate samples.

[0270] In this embodiment, after training is completed, the generation phase begins, where the trained generator is used to generate initial candidate samples in batches. Lightweight post-processing corrections are then applied to these initial candidate samples, employing a step-by-step forced correction strategy to ensure that anomalous data that seriously deviates from the constraints are corrected.

[0271] Post-processing corrections are divided into three levels: The first level is basic physical relationship correction, which targets samples that violate basic physical logic such as temperature magnitude relationships, wind speed relationships, rainfall relationships, and humidity ranges. It forces the constraints to be met through simple numerical adjustments, such as reordering temperatures by magnitude and truncating humidity values ​​that exceed the range to a reasonable interval. The specifics are as follows:

[0272] Level 1 corrections are mandatory corrections to general physical relations, including:

[0273] The following expression is used to force corrections to samples that violate the temperature magnitude relationship:

[0274] ;

[0275] ;

[0276] ;

[0277] in: , , These are correction values ​​for the minimum temperature, maximum temperature, and average temperature, respectively.

[0278] The sample that violates the wind speed relationship is forcibly corrected, and its expression is as follows:

[0279] ;

[0280] in: This is a correction value for the average wind speed;

[0281] The sample that violates the rainfall relationship is forcibly corrected, and its expression is as follows:

[0282] ;

[0283] in: This is a correction value for the maximum rainfall.

[0284] The forced correction for samples that violate the humidity range is expressed as follows:

[0285] ;

[0286] in: This is a correction value for the humidity range.

[0287] The second level is power consistency correction. For samples where the total hourly power consumption does not match the daily power consumption, the deviation ratio is calculated. If the deviation exceeds a set threshold, the load value at each time point is adjusted proportionally to ensure that the power conservation constraint is strictly satisfied and to guarantee the physical rationality of the generated data in terms of energy balance. The specific details are as follows:

[0288] The second level of correction addresses energy consistency by calculating the total hourly energy consumption for each sample and applying a forced correction to the deviation from the daily energy consumption. The expression is as follows:

[0289] ;

[0290] in: This is a correction value for the total time-of-use electricity consumption.

[0291] The third level involves scenario-specific constraint corrections, addressing the specific requirements of different extreme scenarios. For extreme high-temperature scenarios, if the maximum temperature falls outside the specified range, boundary limit corrections are applied; if the temperature-load correlation deviation is too large, the theoretical load value is calculated and weighted averaged with the actual value for correction. For heavy rain scenarios, the daily cumulative rainfall is ensured to meet the minimum requirements, and the rainfall time series distribution is adjusted to avoid unreasonable abrupt changes. For cold wave scenarios, the temperature drop magnitude and minimum temperature are verified to meet the standards, and the temperature series is adjusted to maintain a continuous downward trend, as detailed below:

[0292] Level 3 corrections are mandatory corrections for specific scenarios, including:

[0293] For extreme high-temperature scenarios, if the maximum temperature is not within the range of 35-40℃, a forced correction is applied, expressed as follows:

[0294] ;

[0295] in: This is a correction value for extreme high temperatures;

[0296] For samples with temperature-load correlation anomalies, the theoretical load value is calculated using the following expression:

[0297] ;

[0298] in: This represents the theoretical load value for temperature-load correlation anomalies.

[0299] If the actual load deviates from the theoretical load by more than 20%, a weighted average is used for mandatory correction, the expression of which is as follows:

[0300] ;

[0301] in: This is a correction value for the theoretical load value of the temperature-load correlation anomaly;

[0302] For heavy rain scenarios, if the daily cumulative rainfall is less than 50 mm, a mandatory correction is applied, expressed as follows:

[0303] ;

[0304] in: This is a corrected value for the daily cumulative rainfall;

[0305] For samples with anomalies in the rainfall-load correlation, the theoretical load value is calculated using the following expression:

[0306] ;

[0307] in: This represents the theoretical load value for the rainfall-load correlation anomaly. This is the rainfall saturation threshold; beyond this value, the inhibitory effect no longer increases. As the baseline load; This is the rainfall load suppression coefficient;

[0308] If the actual load deviates from the theoretical load by more than 25%, a weighted average is used for mandatory correction, the expression of which is as follows:

[0309] ;

[0310] in: This is a correction value for the theoretical load value of the rainfall-load correlation anomaly;

[0311] For cold wave scenarios, ensure the minimum temperature does not exceed 4℃, as expressed below:

[0312] ;

[0313] in: This is a correction value for the lowest temperature;

[0314] For samples with cold wave-load correlation anomalies, the theoretical load value is calculated using the following expression:

[0315] ;

[0316] in: This represents the theoretical load value for the cold wave-load correlation anomaly. The cooling rate sensitivity coefficient; This is a reference temperature, usually a comfortable temperature. This refers to the temperature sensitivity coefficient during cold waves. The cooling rate;

[0317] If the actual load deviates from the theoretical load by more than 20%, a weighted average is used for mandatory correction, the expression of which is as follows:

[0318] ;

[0319] in: This is a correction value for the theoretical load value of the cold wave-load correlation anomaly.

[0320] After the three-level correction is completed, data decentralization is implemented to prevent the generated samples from being overly concentrated in certain regions of the feature space. In this scheme, the kernel density estimation method is used to calculate the density distribution of candidate samples in the feature space. Samples in regions with excessively high density are selectively retained or regenerated to ensure that the final dataset has good coverage and diversity in the feature space, avoiding new data skew problems during model training. The specific details are as follows:

[0321] For the corrected samples, a diversity enhancement strategy is adopted, and the sample density is calculated using kernel density. Its expression is as follows:

[0322] ;

[0323] in: The Gaussian kernel function; For bandwidth parameters; These are the sample points for which the density is to be calculated; For the known first One sample point; The total number of samples; The feature dimension of the sample;

[0324] Set density threshold For samples whose density exceeds the threshold, some are randomly removed, and perturbation is added to regenerate some samples to enhance sample diversity and obtain post-processing candidate samples.

[0325] During the validation phase, for post-processing candidate samples, the pass rate of each constraint and the overall quality score are calculated to perform stratified quality checks and obtain high-quality synthetic samples.

[0326] In this embodiment, a systematic hierarchical quality check is performed on the post-processed candidate samples to comprehensively evaluate the sample quality from three levels: general physical constraints, scene-specific constraints, and data statistical distribution.

[0327] The general physical constraint verification layer checks whether each sample meets five basic constraints: temperature relationship, rainfall relationship, wind speed relationship, power consistency, and humidity range. It calculates the pass rate for each constraint and the overall pass rate for general physical constraints. This layer ensures that the generated data conforms to basic physical laws and avoids errors that obviously violate common sense. The scenario-specific constraint verification layer checks whether the samples meet the corresponding specific constraints based on the scenario type. For example, for extreme high-temperature scenarios, it verifies the maximum temperature range and the rationality of the temperature-load correlation; for rainstorm scenarios, it verifies the daily cumulative rainfall and the rainfall-load correlation; for cold wave scenarios, it verifies the temperature drop, minimum temperature, and cold wave load response.

[0328] By calculating the pass rate of specific constraints in each scenario and the overall pass rate, we ensure that the generated data conforms to the business logic and physical mechanism of the specific scenario, as follows:

[0329] Pass rate for calculating temperature relationships Its expression is as follows:

[0330] ;

[0331] in: This represents the total number of temperature samples.

[0332] Calculating the pass rate of rainfall relationship Its expression is as follows:

[0333] ;

[0334] in: This represents the total number of rainfall samples.

[0335] Calculate the pass rate in relation to wind speed Its expression is as follows:

[0336] ;

[0337] in: This represents the total number of wind speed samples.

[0338] Calculate the power consistency pass rate Its expression is as follows:

[0339] ;

[0340] in: This represents the total number of electricity samples.

[0341] Calculate humidity range pass rate Its expression is as follows:

[0342] ;

[0343] in: This represents the total number of humidity samples.

[0344] Based on the pass rates of temperature, rainfall, wind speed, electrical consistency, and humidity range, the overall pass rate of general physical relationships is calculated. Its expression is as follows:

[0345] ;

[0346] Based on defined business constraints for extreme high temperatures, heavy rain, and cold waves, the corresponding pass rates are calculated, specifically including:

[0347] Calculate the pass rate of business constraints in extreme high temperature scenarios Its expression is as follows:

[0348] ;

[0349] in: This represents the total number of samples from extreme high temperatures.

[0350] Calculate temperature-load associated throughput Its expression is as follows:

[0351] ;

[0352] in: This is the theoretical load; Total number of temperature-load samples;

[0353] Calculate the overall pass rate of business constraints in extreme high temperature scenarios. Its expression is as follows:

[0354] ;

[0355] Calculate the pass rate of business constraints in extreme rainstorm scenarios Its expression is as follows:

[0356] ;

[0357] in: This represents the total number of samples from rainstorm scenarios.

[0358] Calculate the rainfall-load correlation pass rate Its expression is as follows:

[0359] ;

[0360] in: This represents the total number of rainfall-load samples.

[0361] Calculate the overall pass rate of business constraints in extreme rainstorm scenarios. Its expression is as follows:

[0362] ;

[0363] Calculate business constraints in extreme cold wave scenarios Its expression is as follows:

[0364] ;

[0365] in: This represents the total number of samples from cold wave scenarios;

[0366] Calculate the cold wave-load correlation pass rate Its expression is as follows:

[0367] ;

[0368] in: This represents the total number of cold wave-load samples.

[0369] Calculate the overall pass rate of business constraints in extreme cold wave scenarios. Its expression is as follows:

[0370] .

[0371] The data statistical distribution validation layer evaluates the consistency between the generated samples and the real samples from a macroscopic perspective, as follows:

[0372] The distribution coverage of generated samples and real samples in the post-processing candidate samples is calculated, and the feature space is divided into grids. The grid cells covered by real samples and generated samples are counted to evaluate whether the distribution area of ​​real samples is sufficiently covered. The expression is as follows:

[0373] ;

[0374] in: For distribution coverage; To generate a set of grid cells covering the sample; The set of grid cells covered by the real sample;

[0375] Calculating the Wasserstein distance measures the similarity of distributions. The distance between two distributions measures the degree of difference between them; the smaller the distance, the more similar the distributions are. Its expression is as follows:

[0376] ;

[0377] in: The distribution of the real samples; The distribution of the generated samples; It is a joint distribution; for Real Samples With generated samples Expected distance; To obtain the minimum expected distance under all possible joint distributions;

[0378] The matching degree between the real sample and the generated sample is calculated. The moment matching degree is used to evaluate the consistency of statistical moments such as mean, variance, skewness, and kurtosis, ensuring that the generated sample is similar not only in local features but also in overall statistical properties to the real data. The expression is as follows:

[0379] ;

[0380] in: For the first real sample Step moment; For the generation of the sample Step moment.

[0381] Based on the three-layer verification results, the comprehensive verification index is calculated as follows:

[0382] First, count the total number of constraints violated for each sample, and then calculate the overall pass rate. This indicator reflects the overall compliance of a sample across all constraint dimensions, and its expression is as follows:

[0383] ;

[0384] in: The total number of samples; To constrain the total number; For the first The number of times a sample violates the constraint;

[0385] Then calculate the weighted overall quality score. The overall pass rate of general physical relationships, the overall pass rate of specific scenario constraints, the distribution coverage rate, and the matching degree are weighted and summed to obtain a single scoring index that comprehensively reflects the sample quality. Its expression is as follows:

[0386] ;

[0387] in: The overall pass rate for general physical relationships; Constrain the overall pass rate for a specific scenario; The normalized index for the Wasserstein distance is... Distance between the real sample and the generated sample distribution; This represents the upper limit of the distance.

[0388] By setting an overall pass rate threshold and a comprehensive quality score threshold, samples that simultaneously meet both conditions are selected as the final high-quality synthetic samples. This dual selection mechanism ensures that the samples meet the necessary constraints and have good statistical characteristics, thereby obtaining augmented data that conforms to physical laws and closely approximates the real distribution.

[0389] Extreme scenario samples and high-quality synthetic samples are merged into an augmented training dataset, and a load prediction model is trained based on this dataset to predict the electricity load of the target area. For example, a complete augmented training dataset is constructed with an extreme scenario to normal scenario ratio of 1:4, and the electricity load of the target area is predicted based on this dataset.

[0390] The augmented dataset includes real historical extreme scenario data to ensure data authenticity, and also includes a large amount of high-quality synthetic extreme scenario data to expand the sample size, which significantly improves the problem of extreme scenario data scarcity and provides sufficient and reliable data support for subsequent risk prediction model training.

[0391] In this embodiment, a long short-term memory network is used to construct the load prediction model. The network contains two LSTM layers, each with 128 hidden units, and a Dropout layer (dropout rate of 0.2). The model is trained for 200 epochs using an augmented dataset with a batch size of 64 and a learning rate of 0.001. The performance is then evaluated on an independent test set, and the mean absolute percentage error (MAPE) is calculated. It can be seen that the MAPE of the model before augmentation was 18.5%, which was reduced to 11.2% after augmentation, resulting in an error reduction of 39.5%.

[0392] In this embodiment, based on the proposed method, data augmentation is performed on rainstorm scenarios with a daily cumulative rainfall of not less than 50 mm, and a three-stage constraint strategy is implemented.

[0393] The soft constraints during training included operational constraints for extreme rainstorm scenarios and rainfall-load correlation constraints (both with a preset weight of 0.5). The generation phase included adjustments to ensure a minimum daily cumulative rainfall of 50 mm and to adjust the rainfall time-series distribution to avoid abrupt changes. Stratified checks during the validation phase showed a pass rate of 0.95 for general physical constraints, 0.93 for rainstorm scenario constraints, and an overall pass rate of 0.93. Finally, 850 high-quality synthetic samples were selected, reducing the model's MAPE from 21.3% to 13.7%.

[0394] In this embodiment, based on the proposed method, a three-stage constraint strategy is implemented for cold wave scenarios where the temperature drop is greater than or equal to 8 degrees Celsius in 24 hours and the minimum temperature is less than or equal to 4 degrees Celsius.

[0395] The soft constraints during the training phase include business constraints for extreme cold wave scenarios (preset weight 0.5) and cold wave-load correlation constraints (preset weight 0.6).

[0396] During the generation phase, corrections were made to ensure a continuous downward trend in temperature, forcibly adjusting temperature sequences that did not conform to the characteristics of a cold wave. The validation phase showed a pass rate of 0.94 for general physical constraints, 0.92 for cold wave scenario constraints, and an overall pass rate of 0.92. Finally, 600 high-quality synthetic samples were selected, reducing the model's MAPE from 25.1% to 14.8%.

[0397] The above embodiments demonstrate that the three-stage constraint implementation strategy of the present invention effectively balances the flexibility of model learning with the reliability of physical constraints, and significantly improves the quality of data augmentation and model prediction performance in extreme scenarios.

[0398] Example 2:

[0399] See Figure 4 A small-sample augmentation prediction system for electricity consumption in extreme scenarios based on PC-GAN, which applies the method described in Example 1, the system comprising:

[0400] Multidimensional feature vector construction module 1 is used to obtain the electricity load-meteorological dataset of the target area, identify extreme scenario samples in the load-meteorological dataset, and construct multidimensional feature vectors based on the electricity load-meteorological dataset;

[0401] Network architecture model building module 2 is used to build a physically constrained generative adversarial network (PC-GAN) model that includes a generator, discriminator, and scene classifier.

[0402] The synthetic sample acquisition module 3 is used to implement a three-stage constraint strategy for the physical constraint generative adversarial network architecture PC-GAN model with multi-dimensional feature vectors as input to obtain high-quality synthetic samples.

[0403] Data augmentation prediction module 4 is used to merge extreme scenario samples with high-quality synthetic samples into an augmented training dataset, and to train a load prediction model based on the augmented training dataset to achieve power load prediction for the target area.

[0404] Furthermore, the specific implementation steps of the multidimensional feature vector construction module 1, network architecture model construction module 2, synthetic sample acquisition module 3, and data augmentation prediction module 4 can be found in the corresponding description in Example 1, and will not be repeated here.

[0405] Example 3:

[0406] See Figure 5 A small-sample augmentation prediction device for power consumption in extreme scenarios based on PC-GAN, the device including a processor 5 and a memory 6;

[0407] The memory 6 is used to store computer program code 61 and transmit the computer program code 61 to the processor 5;

[0408] The processor 5 is used to execute the PC-GAN-based small sample power consumption enhancement prediction method for extreme scenarios as described in Embodiment 1, according to the instructions in the computer program code 61.

[0409] In this embodiment, a computer-readable storage medium is also included, which stores computer-executable instructions. When the computer-executable instructions are executed on a computer, the PC-GAN-based method for small-sample power consumption enhancement prediction in extreme scenarios described in Embodiment 1 is implemented.

[0410] Generally, the computer instructions for implementing the method of the present invention can be carried on any combination of one or more computer-readable storage media. Non-transitory computer-readable storage media can include any computer-readable medium except for the signal itself, which is temporarily propagating.

[0411] Computer-readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EKROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0412] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smarttalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. In particular, Python, suitable for neural network computation, and platform frameworks such as TensorFlow and PyTorch can be used. The program code can be executed 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 remote computers, the remote computer can be connected to the user's computer or to an external computer (e.g., via the Internet using an Internet service provider) through any type of network, including a local area network (LAN) or a wide area network (WAN).

[0413] For details regarding the aforementioned devices and non-transitory computer-readable storage media, please refer to the specific description of a PC-GAN-based method for small-sample power consumption enhancement prediction in extreme scenarios and its beneficial effects, which will not be repeated here.

[0414] Although embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for small-sample augmentation prediction of electricity consumption in extreme scenarios based on PC-GAN, characterized in that, include: Obtain the electricity load-meteorological dataset for the target region, identify extreme scenario samples in the electricity load-meteorological dataset, and construct a multidimensional feature vector based on the electricity load-meteorological dataset; Construct a physically constrained generative adversarial network (PC-GAN) model that includes a generator, discriminator, and scene classifier; The PC-GAN model, based on the physical constraint generative adversarial network architecture, implements a three-stage constraint strategy with multi-dimensional feature vectors as input to obtain high-quality synthetic samples. Extreme scenario samples and high-quality synthetic samples are merged into an enhanced training dataset, and a load prediction model is trained based on the enhanced training dataset to achieve power load prediction for the target area.

2. The PC-GAN-based method for small-sample augmentation prediction of electricity consumption in extreme scenarios according to claim 1, characterized in that: The implementation of the three-stage constraint strategy specifically includes: During the training phase, constraints are set for general physical loss, specific scenario loss, feature distribution matching loss, temporal consistency loss, generator total loss, and discriminator total loss. The adversarial network architecture model is trained alternately to learn samples and obtain initial candidate samples. During the generation phase, for the initial candidate samples, the general physical loss, power consistency constraint, and specific scenario loss are subjected to progressively forced lightweight post-processing corrections and decentralized processing to obtain post-processing candidate samples. During the validation phase, for post-processing candidate samples, the pass rate of each constraint and the overall quality score are calculated to perform stratified quality checks and obtain high-quality synthetic samples.

3. The PC-GAN-based method for small-sample augmentation prediction of electricity consumption in extreme scenarios according to claim 1, characterized in that: The general physical loss includes: Temperature relationship constraints Its loss function is as follows: ； in: The lowest temperature; Average temperature; This is the highest temperature; Rainfall Relationship Constraints Its loss function is as follows: ； in: This represents the maximum rainfall. This refers to cumulative rainfall. Wind speed relationship constraints Its loss function is as follows: ； in: Average wind speed; Maximum wind speed; Humidity Relationship Constraints Its loss function is as follows: ； in: Humidity range; Power Consistency Constraint Its expression is as follows: ； in: This represents the total amount of electricity consumed during the time-of-use period. Daily electricity consumption; Assign time segment numbers; The specific scenario loss includes: Business constraints in extreme high temperature scenarios Its loss function is as follows: ； in: This is the highest temperature; Temperature-load correlation constraints Its loss function is as follows: ； ； in: This is the sum of the time-of-use electricity consumption; Temperature sensitivity coefficient; Humidity sensitivity coefficient; This is the temperature response function; As the baseline load; This refers to the actual temperature. For reference temperature; Business constraints in extreme rainstorm scenarios Its loss function is as follows: ； in: Daily rainfall; Rainfall-load correlation constraints Its loss function is as follows: ； in: This is the rainfall load suppression coefficient; for Rainfall at any given moment; This is for additional power load during heavy rain scenarios; Business constraints in extreme cold wave scenarios Its loss function is as follows: ； in: The temperature drop over 24 hours; Cold wave-load correlation constraints Its loss function is as follows: ； in: This refers to the cold wave load coefficient. For reference temperature; This is the coefficient representing the impact of the temperature drop on the load.

4. The PC-GAN-based method for small-sample augmentation prediction of electricity consumption in extreme scenarios according to claim 3, characterized in that: The feature distribution matching loss Its loss function is as follows: ； in: , The first The mean of real and synthetic samples with 3D features; , The first The standard deviation of real and synthetic samples for 3D features; For feature dimensions; The index of the feature dimension; The timing consistency loss Its loss function is as follows: ； in: This represents the average load variation pattern of the real sample. The generator's total loss Its loss function is as follows: ； ； ； in: The adversarial loss function is the generator loss; The scene classification loss function; Number of scene categories; Labeling based on real-world scenarios; For scene category number; For mathematical expectation; For discriminator For fake samples With conditions The output; The total loss of the discriminator Its loss function is as follows: ； ； in: The adversarial loss function is the discriminator loss. For discriminator For real samples With conditions The output; Let be the mathematical expectation.

5. The PC-GAN-based method for small-sample augmentation prediction of electricity consumption in extreme scenarios according to claim 4, characterized in that: The generation phase involves progressively forced lightweight post-processing corrections to the general physical loss, power consistency constraints, and specific scenario losses for the initial candidate samples. Specifically, this includes: Level 1 corrections are mandatory corrections to general physical relations, including: The following expression is used to force corrections to samples that violate the temperature magnitude relationship: ； ； ； in: , , These are correction values ​​for the minimum temperature, maximum temperature, and average temperature, respectively. The sample that violates the wind speed relationship is forcibly corrected, and its expression is as follows: ； in: This is a correction value for the average wind speed; The sample that violates the rainfall relationship is forcibly corrected, and its expression is as follows: ； in: This is a correction value for the maximum rainfall; The forced correction for samples that violate the humidity range is expressed as follows: ； in: This is a correction value for the humidity range; The second level of correction addresses energy consistency by calculating the total hourly energy consumption for each sample and applying a forced correction to the deviation from the daily energy consumption. The expression is as follows: ； in: This is a correction value for the total time-of-use electricity consumption; Level 3 corrections are mandatory corrections for specific scenarios, including: For extreme high-temperature scenarios, if the maximum temperature is not within the range of 35-40℃, a forced correction is applied, expressed as follows: ； in: This is a correction value for extreme high temperatures; For samples with temperature-load correlation anomalies, the theoretical load value is calculated using the following expression: ； in: This represents the theoretical load value for temperature-load correlation anomalies. If the actual load deviates from the theoretical load by more than 20%, a weighted average is used for mandatory correction, the expression of which is as follows: ； in: This is a correction value for the theoretical load value of the temperature-load correlation anomaly; For heavy rain scenarios, if the daily cumulative rainfall is less than 50 mm, a mandatory correction is applied, expressed as follows: ； in: This is a corrected value for the daily cumulative rainfall; For samples with anomalies in the rainfall-load correlation, the theoretical load value is calculated using the following expression: ； in: This represents the theoretical load value for the rainfall-load correlation anomaly. This is the rainfall saturation threshold; If the actual load deviates from the theoretical load by more than 25%, a weighted average is used for mandatory correction, the expression of which is as follows: ； in: This is a correction value for the theoretical load value of the rainfall-load correlation anomaly; For cold wave scenarios, ensure the minimum temperature does not exceed 4℃, as expressed below: ； in: This is a correction value for the lowest temperature; For samples with cold wave-load correlation anomalies, the theoretical load value is calculated using the following expression: ； in: This represents the theoretical load value for the cold wave-load correlation anomaly. The cooling rate sensitivity coefficient; If the actual load deviates from the theoretical load by more than 20%, a weighted average is used for mandatory correction, the expression of which is as follows: ； in: This is a correction value for the theoretical load value of the cold wave-load correlation anomaly.

6. The PC-GAN-based method for small-sample augmentation prediction of electricity consumption in extreme scenarios according to claim 5, characterized in that: The calculation of the pass rate for each constraint for post-processing candidate samples during the verification phase specifically includes: Pass rate for calculating temperature relationships Its expression is as follows: ； in: This represents the total number of temperature samples. Calculating the pass rate of rainfall relationship Its expression is as follows: ； in: This represents the total number of rainfall samples. Calculate the pass rate in relation to wind speed Its expression is as follows: ； in: This represents the total number of wind speed samples. Calculate the power consistency pass rate Its expression is as follows: ； in: This represents the total number of electricity samples. Calculate humidity range pass rate Its expression is as follows: ； in: This represents the total number of humidity samples. Based on the pass rates of temperature, rainfall, wind speed, electrical consistency, and humidity range, the overall pass rate of general physical relationships is calculated. Its expression is as follows: ； Based on defined business constraints for extreme high temperatures, heavy rain, and cold waves, the corresponding pass rates are calculated, specifically including: Calculate the pass rate of business constraints in extreme high temperature scenarios Its expression is as follows: ； in: This represents the total number of samples from extreme high temperatures. Calculate temperature-load associated throughput Its expression is as follows: ； in: Total number of temperature-load samples; Calculate the overall pass rate of business constraints in extreme high temperature scenarios. Its expression is as follows: ； Calculate the pass rate of business constraints in extreme rainstorm scenarios Its expression is as follows: ； in: This represents the total number of samples from rainstorm scenarios. Calculate the rainfall-load correlation pass rate Its expression is as follows: ； in: This represents the total number of rainfall-load samples. Calculate the overall pass rate of business constraints in extreme rainstorm scenarios. Its expression is as follows: ； Calculate business constraints in extreme cold wave scenarios Its expression is as follows: ； in: This represents the total number of samples from cold wave scenarios; Calculate the cold wave-load correlation pass rate Its expression is as follows: ； in: This represents the total number of cold wave-load samples. Calculate the overall pass rate of business constraints in extreme cold wave scenarios. Its expression is as follows: 。 7. The PC-GAN-based method for small-sample augmentation prediction of electricity consumption in extreme scenarios according to claim 6, characterized in that: The calculation method for the comprehensive quality score specifically includes: The distribution coverage of generated samples and real samples in the post-processing candidate samples is calculated, and the feature space is divided into grids. The grid cells covered by real samples and generated samples are statistically analyzed. The expression is as follows: ； in: For distribution coverage; To generate a set of grid cells covering the sample; A set of grid cells covering the real sample; Calculating the Wasserstein distance measures the similarity of distributions. Its expression is as follows: ； in: The distribution of the real samples; The distribution of the generated samples; It is a joint distribution; for Real Samples With generated samples Expected distance; To obtain the minimum expected distance under all possible joint distributions; Calculate the matching degree between real samples and generated samples. Its expression is as follows: ； in: For the first real sample Step moment; For the generation of the sample Step moment; Calculate the overall pass rate Its expression is as follows: ； in: The total number of samples; To constrain the total number; For the first The number of times a sample violates the constraint; A weighted comprehensive quality score is calculated based on the overall pass rate of general physical relationships, the overall pass rate of specific scenario constraints, the distribution coverage, and the matching degree. Its expression is as follows: ； in: The overall pass rate of general physical relationships; Constrain the overall pass rate for a specific scenario; The normalized index of the Wasserstein distance. Distance between the real sample and the generated sample distribution; This represents the upper limit of the distance.

8. A small-sample augmented prediction system for electricity consumption in extreme scenarios based on PC-GAN, characterized in that, The system is applied to the method according to any one of claims 1-7, the system comprising: The multidimensional feature vector construction module (1) is used to obtain the electricity load-meteorological dataset of the target area, identify extreme scene samples in the electricity load-meteorological dataset, and construct multidimensional feature vectors based on the electricity load-meteorological dataset; The network architecture model building module (2) is used to build a physically constrained generative adversarial network (PC-GAN) model containing a generator, discriminator and scene classifier. The synthetic sample acquisition module (3) is used to implement a three-stage constraint strategy with multi-dimensional feature vectors as input to obtain high-quality synthetic samples; The data augmentation prediction module (4) is used to merge extreme scenario samples with high-quality synthetic samples into an augmented training dataset, and to train a load prediction model based on the augmented training dataset to achieve power load prediction for the target area.

9. A small-sample augmentation prediction device for power consumption in extreme scenarios based on PC-GAN, characterized in that: The device includes a processor (5) and a memory (6); The memory (6) is used to store computer program code (61) and to transmit the computer program code (61) to the processor (5). The processor (5) is used to execute the PC-GAN-based small sample power consumption enhancement prediction method for extreme scenarios according to the instructions in the computer program code (61) as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which, when executed on a computer, implement the PC-GAN-based method for small-sample power consumption enhancement prediction in extreme scenarios as described in any one of claims 1-7.

Images