Power system daily load curve prediction method considering distributed power supply

Abstract

The invention provides an electric power system daily load curve prediction method considering a distributed power supply, and relates to the technical field of electric power load prediction. The method comprises the following steps: acquiring environment data of each type of distributed power supply in a historical period; based on the environment data, performing data screening and expansion in an extreme environment, and constructing extreme sample data; based on the environment data, performing data screening in a normal environment, and constructing normal sample data; constructing a weighted loss function by using the scarcity of the extreme sample data, the prediction error criticality of different environments and the load fluctuation ratio, and carrying out model training under a complete environment condition based on the extreme sample data and the normal sample data by taking the minimum weighted loss function as a target to obtain a daily output prediction model of each type of distributed power supply; and predicting a daily load curve of the power system based on the daily output prediction model of each type of distributed power supply. The accuracy of daily load curve prediction of the power system can be improved.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a method for forecasting the daily load curve of a power system that takes into account distributed generation. Background Technology

[0002] The widespread integration of distributed generation has restructured the traditional power system's source-grid-load structure, transforming the power load forecasting logic from a single user-side forecast to a net load forecast that couples user load with distributed generation output. Accurate daily load curve forecasting serves as a core basis for grid dispatch optimization, reserve capacity allocation, and power supply reliability assurance. Its accuracy directly impacts the safety and economy of grid operation, and fluctuations under extreme weather conditions significantly increase, posing a severe challenge to the grid's regulation capabilities.

[0003] The output characteristics of distributed power sources are highly coupled with environmental conditions, and output fluctuations under extreme weather conditions are more sudden and severe. However, due to the generally short connection time of distributed power sources and the low frequency of extreme weather (such as typhoons, continuous rain, and heavy snow), the number of extreme scenario samples in historical operating data is extremely small. This means that the dataset used for model training can only cover normal scenarios, naturally lacking the data dimension of extreme scenarios, resulting in insufficient dataset completeness.

[0004] Traditional forecasting models rely on a large amount of historical data to fit patterns. When the dataset lacks sufficient extreme samples, the model cannot learn the nonlinear mapping relationship between extreme environments and output, resulting in inaccurate load forecasts. Summary of the Invention

[0005] This invention provides a method for predicting the daily load curve of a power system that takes into account distributed generation, in order to solve the problem of low load prediction accuracy.

[0006] In a first aspect, embodiments of the present invention provide a method for predicting the daily load curve of a power system that takes into account distributed generation, comprising: acquiring environmental data of various types of distributed generation in historical periods; based on the environmental data, performing data screening and expansion under extreme environments to construct extreme sample data; based on the environmental data, performing data screening under normal environments to construct normal sample data; constructing a weighted loss function based on the scarcity of extreme sample data, the severity of prediction errors in different environments, and load fluctuation rate; based on the extreme sample data and normal sample data, performing model training under complete environmental conditions with the goal of minimizing the weighted loss function to obtain daily power generation prediction models for various types of distributed generation; and predicting the daily load curve of the power system based on the daily power generation prediction models for various types of distributed generation.

[0007] In one possible implementation, the daily load curve of the power system is predicted based on the daily power generation prediction model of various types of distributed power sources. This includes: acquiring environmental data of various types of distributed power sources for the day to be predicted; obtaining the daily power generation time series curve of each type of distributed power source based on the environmental data and the daily power generation prediction model of each type of distributed power source; calculating the daily power generation time series curve of each type of distributed power source by power supply area based on the daily power generation time series curve, and obtaining the total power output curve of the distributed power source; and calculating the difference between the pre-stored power load prediction curve and the total power output curve to determine the daily load prediction curve of the power supply area.

[0008] In one possible implementation, extreme sample data is constructed by filtering and expanding data under extreme conditions based on environmental data, including: determining extreme data in the environmental data based on environmental data and preset extreme thresholds; expanding the extreme data based on the extreme data, environmental output mapping model and data augmentation model to obtain the original extreme data; and summarizing and filtering the original extreme data to obtain the extreme sample data.

[0009] In one possible implementation, the data augmentation model includes an extrapolation sub-model, an adversarial sub-model, and a physical rule-constrained interpolation sub-model. Based on extreme data, an environmental output mapping model, and the data augmentation model, the extreme data is expanded to obtain original extreme data, including: extrapolating the extreme data based on the extreme data, the extrapolation sub-model, and the environmental output mapping model to determine the extrapolated extreme data; conducting adversarial training on the extreme data based on the extreme data, the adversarial sub-model, and the environmental output mapping model to determine the synthetic extreme data; interpolating the extreme data based on the extreme data, the physical rule-constrained interpolation sub-model, and the environmental output mapping model to determine the interpolated extreme data; and integrating the extrapolated extreme data, the synthetic extreme data, and the interpolated extreme data to obtain the original extreme data.

[0010] In one possible implementation, extreme data is extrapolated based on extreme data, an extrapolation sub-model, and an environmental output mapping model to determine the extrapolated extreme data. This includes: extracting environmental characteristics of various distributed power sources from the extreme data, where the environmental characteristics are key environmental parameters affecting the output of the distributed power sources; expanding the environmental characteristics based on preset extreme thresholds and preset step sizes to obtain expanded environmental characteristics; obtaining the predicted output of the distributed power sources based on the expanded environmental characteristics and the environmental output mapping model; and determining valid data based on the expanded environmental characteristics, predicted output, and physical constraints of the distributed power sources, using the valid data as the extrapolated extreme data.

[0011] In one possible implementation, adversarial training is performed on extreme data based on extreme data, an adversarial sub-model, and an environmental output mapping model to determine synthetic extreme data. This includes: training based on extreme data, output characteristic data of distributed power sources in a normal environment, and an adversarial sub-model; wherein the adversarial sub-model is constrained by the theoretical output range of the environmental output mapping model; when the distribution difference between the obtained samples and the extreme samples reaches a preset condition, the adversarial training stops, and synthetic extreme data is obtained; interpolation is performed on the extreme data based on extreme data, a physical rule-constrained interpolation sub-model, and the environmental output mapping model to determine interpolated extreme data. This includes: selecting the two ends of the extreme data's continuous extreme time period; obtaining the physical correlation coefficient between environmental characteristics and output within the continuous extreme time period based on the two ends of the data and the environmental output mapping model; performing linear interpolation calculation on the time period between the two ends of the data using the physical rule-constrained interpolation sub-model based on the physical correlation coefficient and the equipment efficiency decay law under continuous extreme environment to obtain the environmental characteristics and output data of the middle time period of the two ends of the data points; and obtaining the interpolated extreme data based on the environmental characteristics and output data of the middle time period of the two ends of the data points.

[0012] In one possible implementation, a weighted loss function is constructed based on the scarcity of extreme sample data, the severity of prediction errors in different environments, and load volatility. This includes: counting the number of extreme samples and normal samples for each type of distributed power source to obtain the total number of samples for complete environmental data; calculating the proportion of extreme sample data based on the number of extreme samples for each type of distributed power source and the total number of samples for complete environmental data; determining the scarcity weight of each extreme sample data based on the proportion of extreme sample data; determining the error severity weight for different environments based on the prediction error and error threshold of distributed power source output in normal and extreme environments; calculating the load volatility weight for each time period based on the load volatility and maximum load volatility for each time period; and determining the weighted loss function based on the scarcity weight, error severity weight, and load volatility weight.

[0013] In one possible implementation, a weighted loss function is determined based on scarcity weight, error severity weight, and load fluctuation weight. This includes: multiplying the scarcity weight, error severity weight, and load fluctuation weight for a single sample data to obtain the comprehensive weight of the single sample data; and determining the weighted loss function based on the comprehensive weight, the total number of samples of complete environmental data, and the prediction error of distributed power output.

[0014] Secondly, embodiments of the present invention provide a power system daily load curve prediction device that takes into account distributed power sources, comprising: a communication module for acquiring environmental data of various types of distributed power sources in historical periods; a processing module for performing data filtering and expansion under extreme environments based on the environmental data to construct extreme sample data; performing data filtering under normal environments based on the environmental data to construct normal sample data; constructing a weighted loss function based on the scarcity of extreme sample data, the severity of prediction errors in different scenarios, and load fluctuation rate; training the model under complete environmental conditions based on extreme sample data and normal sample data, with the goal of minimizing the weighted loss function, to obtain daily power generation prediction models for various types of distributed power sources; and predicting the daily load curve of the power system based on the daily power generation prediction models for various types of distributed power sources.

[0015] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation thereof.

[0016] In this embodiment of the invention, by filtering and expanding data under extreme weather conditions and filtering data under normal weather conditions, a complete training dataset covering both conventional and extreme scenarios is constructed. This addresses the problem of scarce and unevenly distributed distributed power generation output data under extreme scenarios, laying a data foundation for model learning. During the model training phase, a weighted loss function is constructed based on the scarcity of extreme samples, the severity of prediction errors in different scenarios, and load fluctuation rate. The function is optimized with the goal of minimizing loss. The scarcity of extreme samples represents the weight of extreme samples, amplifying their impact during training and preventing insufficient model fitting of extreme scenario patterns due to insufficient sample size. The severity of prediction errors in different scenarios represents the error penalty coefficient, forcibly correcting prediction biases under extreme scenarios through high penalty strength. Load fluctuation rate represents the sudden changes in output of extreme samples, strengthening the feature weights during periods of high fluctuation and improving the model's ability to capture sudden changes in output patterns under extreme scenarios. The complete dataset and the multi-dimensional loss function work synergistically, enabling the trained daily power generation prediction models for various types of distributed power sources to not only grasp the environmental-output patterns of conventional scenarios but also accurately capture the special characteristics of extreme scenarios, reducing overall prediction bias and ultimately improving the accuracy of daily load prediction. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the implementation of the daily load curve prediction method for power systems considering distributed power sources provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the power system daily load curve prediction device considering distributed power sources provided in an embodiment of the present invention; Figure 3It is a schematic diagram of the electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0018] Next, the embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0019] Refer to Figure 1 , which shows the implementation flowchart of the daily load curve prediction method for a power system considering distributed power sources provided by an embodiment of the present invention, and is described in detail as follows: Step 101: Obtain the environmental data of various types of distributed power sources during the historical period.

[0020] In some embodiments, various types of distributed power sources refer to small-scale power generation devices deployed on the user side or near the load center and connected to the power system in a decentralized manner. Different from traditional centralized large-scale power plants (such as large thermal power plants and hydropower plants), their output is significantly affected by environmental factors and is a power-side variable that needs to be considered key in the prediction of the daily load curve of the power system.

[0021] In some embodiments, various types of distributed power sources include new energy types, traditional energy types, waste heat / pressure types.

[0022] Exemplarily, new energy types include photovoltaic power plants (distributed rooftop photovoltaics, small-scale ground photovoltaics), small wind turbines (onshore wind farms with a single-unit capacity ≤ 1.5 MW), and small energy storage systems (electrochemical energy storage supporting photovoltaics / wind power, user-side energy storage).

[0023] Exemplarily, traditional energy types include small gas turbines (mostly used as backup power sources for industrial and commercial users), and small biomass power generation, such as rural straw gasification power generation and biogas power generation in farms.

[0024] Exemplarily, waste heat / pressure types include waste heat power generation in industrial enterprises (such as waste heat boiler power generation in steel mills and cement plants), and pressure recovery power generation (such as pressure recovery power generation of mine ventilation fans).

[0025] In some embodiments, environmental data refers to the basic environmental data that affects the output of various types of distributed power sources on the day to be predicted, which is the original measurement data reflecting the environmental conditions, without data cleaning, fusion or enhancement processing, and is the basis for generating complete environmental data later.

[0026] In some embodiments, environmental data includes data of photovoltaic power plants, small wind turbines, small energy storage systems, and small gas turbines.

[0027] Exemplarily, data of photovoltaic power plants includes light intensity, environmental temperature, relative humidity, and solar altitude angle.

[0028] Exemplarily, data of small wind turbines includes wind speed, wind direction, and air density.

[0029] For example, data from a small energy storage system may include ambient temperature, outdoor humidity, and atmospheric pressure (which affect battery heat dissipation efficiency).

[0030] For example, data for a small gas engine includes ambient temperature (which affects gas combustion efficiency) and gas composition (such as methane content, which affects calorific value).

[0031] Step 102: Based on environmental data, perform data filtering and expansion under extreme environments to construct extreme sample data.

[0032] In some embodiments, extreme environments refer to extreme conditions rather than normal operating conditions. The data sample size in such extreme environments is usually small, but it has a critical impact on the accuracy of load forecasting. For example, extreme rainstorms can cause a sharp drop in photovoltaic output, and extreme winds may trigger wind turbine shutdowns.

[0033] For example, in extreme heavy rainfall and sandstorm scenarios of photovoltaic power plants, the irradiance (<50W / ㎡), output value (<10% of rated power), and shutdown status (such as inverter protection shutdown signal).

[0034] For example, in extreme wind (greater than 25 m / s) and no-wind scenarios for small wind turbines, wind speed, turbine output (>1.2 times the rated power or <5% of the rated power), yaw system fault signal, and downtime.

[0035] For example, in extreme high temperature (greater than 40 degrees Celsius) scenarios, there are abnormal fluctuations in energy storage battery temperature, charging and discharging power limits (such as the discharge power dropping to 80% of the rated value at high temperatures), and SOC (state of charge).

[0036] As one possible implementation, step 102 can be specifically implemented as steps A11-A13.

[0037] A11: Based on environmental data and preset extreme thresholds, determine the extreme data in the environmental data.

[0038] A12: Based on extreme data, environmental output mapping model and data augmentation model, the extreme data is expanded to obtain the original extreme data.

[0039] A13: Based on the original extreme data, we summarize and filter to obtain extreme sample data.

[0040] In some embodiments, the preset extreme threshold is a critical value for determining whether environmental data belongs to an extreme scenario. It is used to filter out samples that exceed the normal range from the environmental data and is determined by technicians based on equipment characteristics, industry standards, or historical data statistics.

[0041] In some embodiments, the preset extreme thresholds include thresholds for photovoltaic, wind power, and gas turbines.

[0042] For example, the thresholds for photovoltaics include light intensity < 50 W / m² (extremely low light) and temperature > 45°C (extremely high temperature).

[0043] For example, wind power thresholds include wind speed > 25 m / s (cutoff wind speed, wind turbine shutdown) and wind speed < 0.5 m / s (extreme no wind).

[0044] For example, the threshold for a gas turbine is a gas pressure <0.6MPa (extremely low pressure, unable to operate stably).

[0045] In some embodiments, extreme data refers to sample data in environmental data that exceeds a preset extreme threshold, reflecting the original operating state of distributed power sources under extreme environments.

[0046] In some embodiments, extreme data includes extreme wind scenarios and extreme high temperature scenarios.

[0047] For example, extreme wind scenarios include wind speeds of 28 m / s (exceeding the threshold of 25 m / s) and wind turbine output of 0 MW (due to shutdown).

[0048] For example, in extreme high-temperature scenarios such as an ambient temperature of 48°C (exceeding the threshold of 45°C), photovoltaic output drops to 60% of the rated value (due to component efficiency degradation).

[0049] In some embodiments, the environmental output mapping model is a set of mathematical models, algorithms, or rules that describe the quantitative or qualitative correlation between environmental data affecting the output of distributed power sources and the actual output of distributed power sources. Essentially, it is an abstract expression of the physical process or data pattern of how environmental conditions determine power output. It can output corresponding predicted power output values ​​based on input environmental data, or infer corresponding environmental characteristics based on known output values.

[0050] For example, for a photovoltaic power station, irradiance and ambient temperature are the core environmental data. The environmental output mapping model can be expressed as photovoltaic output = f(irradiance, ambient temperature), where f() is the mathematical expression of the mapping relationship.

[0051] In some embodiments, the environmental output mapping model, in conventional scenarios, can directly output hourly output predictions of distributed power sources based on real-time or predicted environmental data (such as the next day's sunshine and wind speed predicted by a weather station). During extreme environment data screening and post-data augmentation integration, the model can act as a physical rule checker. For example, if an extreme data point shows that photovoltaic output reaches 50% of rated power under a light intensity of 50W / m² (extremely low light), the environmental output mapping model calculates that the theoretical output under this light intensity should be <10% of the rated power. Therefore, this sample can be identified as an outlier and removed, ensuring the rationality of the final extreme environment data. In data augmentation and daily output prediction model training, the model provides the physical output boundaries of the distributed power source. When training the daily output prediction model, the output of the environmental output mapping model can be used as prior knowledge to correct the model's prediction results. For example, if the model predicts that the photovoltaic output exceeds the physical limit, it is forcibly corrected to the rated power.

[0052] In some embodiments, the classification of environmental output mapping models includes physical models, data-driven models, and hybrid models.

[0053] Examples include physical models, such as formulas based on device principles, like photovoltaic output = irradiance × module conversion efficiency × area × temperature coefficient.

[0054] Examples of data-driven models include neural networks (inputting light and temperature, outputting photovoltaic power output) and random forests (inputting wind speed and wind direction, outputting wind power output).

[0055] Examples include hybrid models, such as models that combine physical rules (e.g., maximum output limits) and data fitting (e.g., modified LSTM models).

[0056] In some embodiments, data augmentation models are tools used to expand the amount of extreme sample data to address the problem of sample scarcity in extreme scenarios.

[0057] In some embodiments, the original extreme data is an initial set of extreme samples obtained after processing by a data augmentation model, containing extended data of the original extreme samples, but which has not yet been filtered and verified.

[0058] In some embodiments, extreme sample data are valid extreme samples selected from extreme data that conform to physical laws and prediction requirements.

[0059] As one possible implementation, embodiments of the present invention filter extreme data based on environmental data and preset extreme thresholds, separating extreme data from regular data through threshold definition; the original extreme data is obtained by expanding the data based on a data augmentation model, solving the industry pain point of small sample size and incomplete coverage of extreme scenarios, such as the scarcity of historical data for extreme rainstorms and super typhoons, and expanding the sample size through technical means; the extreme sample data is summarized and filtered to ensure the validity of the final data, remove noise or outliers that do not conform to physical laws, and provide high-quality input for subsequent model training.

[0060] In some embodiments, the data augmentation model includes an extrapolation sub-model, an adversarial sub-model, and a physical rule-constrained interpolation sub-model.

[0061] As one possible implementation, step A12 can be specifically implemented as steps B11-B14.

[0062] B11: Based on extreme data, extrapolation sub-models, and environmental output mapping models, extrapolate extreme data to determine the extrapolated extreme data.

[0063] B12: Based on extreme data, adversarial sub-models, and environmental output mapping models, adversarial training is performed on extreme data to determine synthetic extreme data.

[0064] B13: Based on extreme data, physical rule-constrained interpolation sub-models and environmental output mapping models, interpolation is performed on extreme data to determine the interpolated extreme data.

[0065] B14: Integrate extrapolated extreme data, synthesized extreme data, and interpolated extreme data to obtain the original extreme data.

[0066] In some embodiments, the essence of the environmental output mapping model is a mathematical model that quantifies the correlation between environmental data and distributed power output. For example, the photovoltaic light intensity-temperature-output relationship model and the wind speed-power curve model of wind turbines. Its core value is to transform abstract environmental data into output data that conforms to the physical characteristics of the equipment, while filtering out abnormal data that does not conform to the rules. This characteristic enables it to play a dual role in the generation of original extreme samples, including correlation calculation and constraint verification. Moreover, constraints are the core of ensuring the validity of the data.

[0067] In some embodiments, the environmental output mapping model serves as a key constraint to ensure that extreme data generated through extrapolation, adversarial methods, interpolation, etc., conforms to the physical operating laws of distributed power sources, thus avoiding the occurrence of invalid data that is detached from reality.

[0068] In some embodiments, extrapolated extreme data refers to newly generated extreme samples after the range of existing extreme data is expanded through extrapolation sub-models. The core is to extend the range of environmental data to more extreme environments based on the trend of existing extreme samples, and generate extreme samples that are less likely to occur in real scenarios but may occur.

[0069] In some embodiments, the extrapolated extreme data includes the expanded extreme data and the corresponding output data.

[0070] For example, in a photovoltaic power plant, existing extreme data such as 50W / m² irradiance and 8% of rated output can be extrapolated to extreme data such as: 40W / m² irradiance (weaker) and 6.4% of rated output; and 30W / m² irradiance and 4.8% of rated output (extrapolated along a linear trend).

[0071] For example, in the case of existing extreme data for small wind turbines, such as wind speed of 25m / s (cut-out wind speed), the output is 0; the extrapolated extreme data are wind speed of 27m / s (stronger) and output of 0; wind speed of 30m / s and output of 0 (the wind turbine is continuously shut down, and the output remains at 0).

[0072] In some embodiments, synthetic extreme data refers to synthesizing novel extreme scenario samples (especially composite extreme scenario samples that are rare in real-world scenarios) by combining existing extreme data with conventional environmental output characteristic data through adversarial sub-models. The core is to make the generated samples conform to the distribution pattern of extreme scenarios and satisfy physical constraints through adversarial training.

[0073] In some embodiments, the composite extreme data includes photovoltaic composite extreme data, wind turbine composite extreme data, and energy storage composite extreme data.

[0074] For example, extreme data for photovoltaic composites include extreme low light (40W / ㎡ irradiance), extreme low temperature (-10℃), and 5% of the rated output (low temperature will cause a slight decrease in module efficiency, with output 1.4% lower than in a simple low light scenario).

[0075] For example, wind turbine composite extreme data include extreme wind (wind speed 28m / s), extreme low temperature (-20℃), zero output (wind turbine shutdown), and yaw system fault signal (low temperature causes mechanical parts to jam).

[0076] In some embodiments, interpolated extreme data refers to the intermediate extreme data supplemented by the physical laws of the device between the two valid data points at both ends of a continuous extreme time period by constraining the interpolation sub-model through physical rules. The core is to solve the problem of sample fragmentation caused by data loss in continuous extreme time periods and ensure the temporal continuity of data.

[0077] In some embodiments, interpolated extreme data includes continuous extreme time periods, data points at both ends, interpolated extreme data (intermediate time periods), and data features.

[0078] For example, in a continuous extreme time period such as 10:00-13:00 (3 hours in total), only the two ends of the data point (10:00 and 13:00) are known.

[0079] For example, the two data points are 10:00 (sunlight 60W / ㎡, power output 7.2%) and 13:00 (sunlight 40W / ㎡, power output 4.8%).

[0080] For example, interpolated extreme data (mid-term) such as 11:00: irradiance 53.3W / ㎡ (linear decay), output 6.4% (combined with the slow decay of component efficiency by 0.2% during the rainstorm period); 12:00: irradiance 46.7W / ㎡, output 5.6% (efficiency further decays by 0.2%).

[0081] For example, data characteristics such as environmental data and output change continuously according to physical rules (linear decay, efficiency decay) without jumps, which conforms to the actual operating rules.

[0082] In some embodiments, the original extreme data refers to the set of extreme data obtained after integrating and preliminarily verifying extrapolated extreme data, synthetic extreme data, and interpolated extreme data. It is the basis for subsequent summarization and screening to obtain the final extreme sample data, and has not yet undergone strict validity screening (such as removing individual outliers that do not conform to physical laws).

[0083] In some embodiments, the original extreme data can be obtained by weighted summation of extrapolated extreme data, synthetic extreme data, and interpolated extreme data.

[0084] In some embodiments, the raw extreme data includes a set of raw extreme samples from photovoltaic systems and a set of raw extreme samples from wind turbine systems.

[0085] For example, the original photovoltaic extreme data set includes extrapolated low light samples, synthetic low light samples, low temperature samples, and interpolated samples from periods of heavy rain.

[0086] For example, the original extreme data set of the wind turbine includes extrapolated strong wind samples, synthetic strong wind samples, low temperature samples, and interpolated continuous windless period samples.

[0087] As one possible implementation, this invention embodiment can clearly define the composition of the data augmentation model, breaking it down into an extrapolation sub-model, an adversarial sub-model, and a physical rule-constrained interpolation sub-model, corresponding to three core requirements: sample range expansion, novel sample synthesis, and missing data imputation, respectively. The application scenarios of each sub-model are refined: the extrapolation sub-model is for expanding existing extreme samples to more extreme ranges (e.g., from wind speed of 25 m / s to 30 m / s), the adversarial sub-model is for generating samples for composite extreme scenarios (e.g., rare scenarios of strong winds and low temperatures), and the interpolation sub-model is for supplementing missing data for continuous extreme periods (e.g., extreme rainstorms lasting 3 hours but with 1 hour of missing data in the middle).

[0088] As one possible implementation, step B11 can be specifically implemented as steps C11-C14.

[0089] C11: Extract environmental characteristics of various distributed power sources from extreme data. These environmental characteristics are key environmental parameters that affect the output of distributed power sources.

[0090] C12: Based on preset extreme thresholds and preset step sizes, environmental features are extended to obtain extended environmental features.

[0091] C13: Based on the extended environmental characteristics and environmental output mapping model, the predicted output of distributed power sources is obtained.

[0092] C14: Based on the extended environmental characteristics, predicted output, and physical constraints of distributed power sources, determine the effective data and use the effective data as extrapolated extreme data.

[0093] In some embodiments, environmental features refer to key environmental data features extracted from extreme data that can significantly affect the output of distributed power sources. They serve as the basic input for extrapolation and reflect the core attributes of extreme environments.

[0094] In some embodiments, environmental characteristics include those of photovoltaic power plants, wind turbines, and energy storage systems.

[0095] For example, environmental characteristics of a photovoltaic power plant include light intensity, ambient temperature, and atmospheric transparency (reflecting the degree of dust / haze).

[0096] In some embodiments, the environmental characteristics of the fan include wind speed, wind direction angle, and air density.

[0097] In some embodiments, the environmental characteristics of the energy storage system include ambient temperature, humidity, and atmospheric pressure.

[0098] In some embodiments, the preset step size refers to the incremental interval when extrapolating environmental features, which is used to control the fineness of the expansion and ensure that the generated samples are uniformly distributed within an extreme range.

[0099] In some embodiments, the extended environmental features refer to the new extreme environmental data obtained by expanding the range of the original environmental features through preset extreme thresholds and preset step sizes, and are the intermediate output of the extrapolation sub-model.

[0100] In some embodiments, the extended environmental characteristics are as follows: original extreme sample: light intensity 50W / ㎡ (weak light); extended: 45W / ㎡, 40W / ㎡, ..., 10W / ㎡ (weaker light); original extreme sample: wind speed 25m / s (strong wind); extended: 26m / s, 27m / s, ..., 35m / s (stronger wind).

[0101] In some embodiments, the physical constraints of distributed power sources refer to the output limits or operating rules determined by the device characteristics and operating principles of the distributed power sources. These constraints are used to screen the validity of extrapolated samples and avoid generating data that violates physical laws.

[0102] In some embodiments, valid data refers to the extended environmental features and corresponding predicted output matching data retained after verification by the physical constraints of the distributed power source, which is the final output of the extrapolation sub-model.

[0103] As one possible implementation, embodiments of the present invention can extract core environmental data strongly correlated with power output (such as solar irradiance and wind speed of wind turbines) from extreme samples to clarify the basic variables for extrapolation; based on preset extreme thresholds (such as wind speed limit of 35 m / s) and preset step size (such as 1 m / s), expand environmental features to a more extreme range to clarify the scope and granularity of extrapolation; through an environmental power output mapping model, transform the expanded environmental features into corresponding power output to establish a new environment-new power output relationship; and use distributed power source physical constraints (such as wind turbine power output being 0 after cutting off wind speed) to eliminate invalid samples to ensure the rationality of extrapolated data.

[0104] As one possible implementation, step B12 can be specifically implemented as steps D11-D12.

[0105] D11: Training is performed based on the output characteristic data of distributed power sources in extreme and normal environments, and the adversarial sub-model; wherein the adversarial sub-model is constrained by the theoretical output range of the environmental output mapping model.

[0106] D12: When the distribution difference between the obtained sample and the extreme sample reaches the preset condition, the adversarial process stops, and synthetic extreme data is obtained.

[0107] In some embodiments, the output characteristic data of distributed power sources in normal environments refers to the matching data of environmental data and corresponding output when the distributed power source is running in a non-extreme environment (meeting a preset normal threshold), reflecting the output pattern under normal operating conditions.

[0108] In some embodiments, the output characteristic data of distributed power sources in conventional environments includes wind turbine scenarios and photovoltaic scenarios.

[0109] For example, photovoltaic scenarios include output data when the irradiance is 200-1000W / ㎡ and the temperature is 15-35℃ (e.g., 500W / ㎡ of irradiance corresponds to 50% of the rated output).

[0110] For example, the wind turbine scenario includes output data at wind speeds of 3-15 m / s (within the rated wind speed range) (e.g., a wind speed of 10 m / s corresponds to 70% of the rated output).

[0111] In some embodiments, the theoretical output range of the environmental output mapping model refers to the output range that the distributed power source can physically reach, calculated according to the environmental output mapping model, and is the reasonable boundary for generating samples by the adversary model.

[0112] In some embodiments, the theoretical output range of the environmental output mapping model includes photovoltaic scenarios and wind turbine scenarios.

[0113] For example, the output range in a photovoltaic scenario is [0, rated power × 1.1] (considering the possibility of exceeding the rated output during short periods of intense sunlight).

[0114] For example, in a wind turbine scenario, the output range is [0, rated power] (the output remains at the rated value after the wind speed exceeds the rated wind speed, until the cut-off wind speed).

[0115] In some embodiments, the distribution difference between the sample and the extreme sample reaching a preset condition means that the similarity between the sample generated by the adversarial sub-model and the original extreme sample in data distribution (such as environmental data distribution, output distribution) meets a preset standard, indicating that adversarial training can be stopped.

[0116] In some embodiments, synthetic extreme data refers to new extreme samples generated by adversarial models that conform to the distribution patterns of extreme scenarios and satisfy physical constraints, especially for composite extreme scenarios that are rare in real data.

[0117] As one possible implementation, embodiments of the present invention can use extreme data and normal environmental output feature data as training basis to allow the generator to learn the difference features between extreme scenarios and normal scenarios, and then synthesize composite extreme samples that do not exist in real data; during the adversarial process, the theoretical output range of the environmental output mapping model is used as a constraint to avoid generating false samples that violate physical laws.

[0118] As one possible implementation, step B13 can be specifically implemented as steps E11-E14.

[0119] E11: Filters data points at both ends of a continuous extreme time period for extreme data.

[0120] E12: Based on the data points at both ends and the environmental output mapping model, the physical correlation coefficient between environmental characteristics and output during this continuous extreme period is obtained.

[0121] E13: Based on the physical correlation coefficient and the equipment efficiency decay law under continuous extreme environment, the interpolation sub-model constrained by physical rules is used to perform linear interpolation calculation on the time period between the two data points to obtain the environmental characteristics and output data of the middle time period between the two data points.

[0122] E14: Based on the environmental characteristics and output data of the middle period between the two data points, interpolated extreme data are obtained.

[0123] In some embodiments, the physical correlation coefficient between environmental characteristics and output refers to a coefficient that quantitatively describes the causal relationship between changes in environmental data and changes in output over a continuous extreme period, reflecting the laws governing physical influence.

[0124] In some embodiments, the equipment efficiency degradation law under continuous extreme environments refers to the degradation trend of output efficiency of distributed power sources over time due to equipment aging and performance degradation in long-term extreme environments.

[0125] As one possible implementation, this embodiment of the invention combines the physical correlation coefficient between the environment and output (e.g., photovoltaic output decreases by 1.2% for every 10W / ㎡ decrease in sunlight) and the equipment efficiency decay law (e.g., energy storage efficiency decreases by 0.5% per hour during continuous high temperatures) to ensure that the supplemented data conforms to the actual operating characteristics of the equipment; it expands the original two isolated data points at both ends into continuous time-series data covering the entire extreme period (e.g., 11 intermediate points are supplemented at a 15-minute granularity for a 3-hour period), so that the extreme samples change from static points to dynamic sequences.

[0126] Step 103: Based on environmental data, perform data filtering under normal conditions to construct normal sample data.

[0127] In some embodiments, normal sample data refers to environmental data that does not exceed a preset extreme threshold, reflecting the original operating state of the distributed power source under normal conditions.

[0128] Step 104: Construct a weighted loss function based on the scarcity of extreme sample data, the severity of prediction errors in different environments, and load fluctuation. Based on extreme sample data and normal sample data, train the model under complete environmental conditions with the goal of minimizing the weighted loss function to obtain daily power generation prediction models for various types of distributed power sources.

[0129] In some embodiments, extreme sample data (including environmental and output data) and normal sample data are fused by unifying the format, associating data, removing redundancy, and verifying anomalies, transforming the two types of data into a dataset with consistent structure and comprehensive coverage.

[0130] In some embodiments, integration includes format unification, data association, redundancy removal, and anomaly verification.

[0131] For example, format unification standardizes the time granularity (e.g., 15 minutes / data entry), parameter units (e.g., light intensity is unified to W / m², wind speed to m / s), and data format of the two types of data.

[0132] For example, data association is achieved by ensuring that the structure of the sample data is consistent with that of the extreme sample data.

[0133] For example, redundancy removal deletes duplicate records in two types of data.

[0134] For example, the anomaly check is based on the theoretical range of the environmental output mapping model and removes invalid data (such as anomaly records in the original environmental data where the illumination is 0W / ㎡ but the corresponding output is 50% of the rated value).

[0135] In this embodiment, fusing extreme sample data and normal sample data to obtain complete environmental data is a key data processing step in the entire power system daily load curve prediction method. Its core function is to solve the problems of data fragmentation, incomplete scenario coverage, and inconsistent formats, and to provide high-quality, full-scenario basic data support for subsequent model training.

[0136] In some embodiments, the scarcity of extreme sample data refers to the degree of scarcity of extreme scenarios.

[0137] In some embodiments, the prediction error is the difference between the predicted output of the power generation prediction model and the actual monitored output, used to measure the accuracy of the model prediction.

[0138] In some embodiments, the error severity weight is a weight calculated based on the relationship between the prediction error and the error threshold, used to quantify the degree of harm of the prediction error to the power grid under different scenarios (the severity of the error is higher in extreme scenarios).

[0139] In some embodiments, load volatility refers to the variation in load (or distributed generation output) in a power system during a certain period with the load (or output) in an adjacent period. It is used to measure the stability of power system operation. The greater the volatility, the greater the challenge to grid dispatch.

[0140] In some embodiments, the daily power output prediction model refers to a machine learning model that is trained to predict the output of distributed power sources at different times of the day based on input environmental data (such as sunlight and wind speed), and is a core tool for load curve prediction.

[0141] As one possible implementation, this invention is based on complete environmental data, with more and richer samples. The model can learn the patterns of low-probability events, and the trained model covers both normal and extreme scenarios, making it more consistent with reality and thus improving the accuracy of model predictions. Furthermore, a loss function constructed using the proportion of extreme environmental data, prediction error, and load fluctuation rate is used for training to obtain daily power generation prediction models for various types of distributed power sources. This is the core technology hub of the entire power system's daily load curve prediction method. Its role is to solve the shortcomings of traditional prediction models, such as low accuracy in extreme scenarios and poor adaptability during periods of high fluctuation. Through customized training logic, a high-precision power output prediction model that covers all scenarios and adapts to the characteristics of the power system is constructed. The model trained by this invention is more consistent with actual scenarios, avoiding the neglect of extreme scenarios. Therefore, using the daily power generation prediction model trained by this invention for prediction is more realistic, thereby improving the accuracy of daily load prediction.

[0142] As one possible implementation, step 104 can be specifically implemented as steps A31-A36.

[0143] A31: Count the number of extreme samples and normal samples of each type of distributed power source to obtain the total number of samples for complete environmental data.

[0144] A32: The proportion of extreme sample data is calculated based on the number of extreme samples of each type of distributed power source and the total number of samples of complete environmental data.

[0145] A33: Determine the scarcity weight of each extreme sample data based on the proportion of extreme sample data.

[0146] A34: Based on the prediction error and error threshold of distributed power output in normal and extreme environments, determine the error severity weight for different environments.

[0147] A35: Based on the load fluctuation rate and maximum load fluctuation rate for each time period, the load fluctuation weight for each time period is calculated.

[0148] A36: Determine the weighted loss function based on scarcity weight, error severity weight, and load volatility weight.

[0149] In some embodiments, the extreme sample data ratio refers to the ratio of the number of extreme samples to the total number of complete environmental data samples, which is used to reflect the scarcity of extreme samples (the lower the ratio, the higher the scarcity).

[0150] In some embodiments, the sample scarcity weight is a weight value calculated based on the proportion of extreme sample data and used to increase the importance of extreme samples in the loss function. Essentially, it is a weighted compensation for scarce samples (extreme samples).

[0151] In some embodiments, the prediction error of distributed power output in normal and extreme scenarios refers to the difference between the predicted output and the actual output of the model for normal and extreme scenario samples respectively during the training process, which is used to measure the prediction accuracy of the model in different scenarios.

[0152] In some embodiments, the error severity weight is a weight calculated based on the relationship between the prediction error and the error threshold, used to quantify the degree of harm of the prediction error to the power grid under different scenarios (the severity of the error is higher in extreme scenarios).

[0153] In some embodiments, the load volatility of each time period refers to the variation of the output of distributed generation in a power system during a certain time period (e.g., 1 hour) with the output of adjacent time periods, and is used to measure the stability of the output (the higher the volatility, the greater the difficulty of grid dispatch).

[0154] In some embodiments, the maximum load volatility refers to the maximum value of the load volatility across all time periods in the complete environmental data. It represents the most severe fluctuation in the output of distributed power sources and serves as a benchmark for measuring volatility risk.

[0155] In some embodiments, the load volatility weight is a weight calculated based on the ratio of the load volatility of each period to the maximum load volatility, used to increase the importance of high volatility periods in the loss function (the higher the volatility, the greater the weight).

[0156] As one possible implementation, embodiments of the present invention can address the problems of traditional loss functions not paying enough attention to extreme scenarios and being out of touch with power system demands by quantifying the scarcity of extreme samples, the severity of errors, and the risk of load fluctuations. Ultimately, a customized loss function adapted to the characteristics of the power system can be constructed, ensuring that the training direction of the daily power generation prediction model is deeply bound to the requirements of safe operation of the power grid.

[0157] As one possible implementation, step A36 can be specifically implemented as steps B31-B32.

[0158] B31: For a single sample data, multiply the scarcity weight, error severity weight, and load volatility weight to obtain the comprehensive weight of the single sample data.

[0159] B32: Determine the weighted loss function based on the comprehensive weight, the total number of samples of complete environmental data, and the prediction error of distributed power output.

[0160] In some embodiments, a single sample data point refers to an independent environmental data point—an actual output record—within the complete environmental data set. It is the smallest data unit for model training and can belong to either a normal scenario or an extreme scenario.

[0161] In some embodiments, a single sample data includes a timestamp, environmental data, the actual output of the corresponding distributed power source, and a scenario identifier.

[0162] In some embodiments, the total number of samples in the complete environmental data refers to the total number of all samples in the complete environmental data (regular samples, extreme samples), which is used to normalize the overall loss.

[0163] As one possible implementation, embodiments of the present invention can solve the problem that traditional loss functions cannot accurately adapt to the priority requirements of power systems through the calculation logic of single-sample weighting and total loss aggregation.

[0164] Step 105: Based on the daily power generation prediction model of various types of distributed power sources, predict the daily load curve of the power system.

[0165] In some embodiments, the daily power output prediction model is a trained, specialized prediction model for different types of distributed power sources, capable of outputting a time-series curve of power output for the next day based on environmental data.

[0166] In some embodiments, the power generation prediction model includes a photovoltaic power generation prediction model, a wind turbine power generation prediction model, an energy storage power generation prediction model, and other distributed power generation models.

[0167] In some embodiments, the daily power generation prediction model is the core prediction tool, responsible for transforming environmental data into specific power generation time series curves, and is the key hub connecting environmental data and load curves.

[0168] As one possible implementation, step 105 can be specifically implemented as steps A41-A44.

[0169] A41: Obtain environmental data for various types of distributed power sources on the day to be predicted. The environmental data is the key environmental data affecting the output of distributed power sources.

[0170] A42: Based on environmental data and daily power generation prediction models for various types of distributed power sources, the daily power generation time series curves of each type of distributed power source are obtained.

[0171] A43: Based on the daily power output time-series curve, the daily power output time-series curves of each type of distributed power source are calculated according to the power supply area to obtain the total power output curve of the distributed power source.

[0172] A44: Calculate the difference between the pre-stored electricity load forecast curve and total output curve to determine the daily load forecast curve for the power supply area.

[0173] In some embodiments, the environmental data of various types of distributed power sources on the date to be predicted refers to the key environmental factors that affect the output of different types of distributed power sources (such as photovoltaic, wind turbines, energy storage, etc.) on the target date (date to be predicted) when load forecasting is required, and it is the input basis of the prediction model.

[0174] In some embodiments, the daily output time series curve of a distributed power source refers to a continuous curve formed by the predicted output values ​​of a certain type of distributed power source arranged at a fixed time granularity (such as 15 minutes or 1 hour) on the day to be predicted, reflecting the output change pattern of the power source within a day.

[0175] In some embodiments, the daily power output time-series curve of the corresponding distributed power source includes a time dimension and a power output value.

[0176] For example, the time dimension covers the period from 00:00 to 24:00 of the day to be predicted, divided according to a preset granularity (e.g., 96 data points / day, one point every 15 minutes).

[0177] For example, the output value is the predicted output for each point in time, such as 300kW of photovoltaic power at 8:00, 800kW at 12:00, and 0kW at night.

[0178] In some embodiments, the total output curve of a distributed power source refers to the curve formed by superimposing the daily output time-series curves of all types of distributed power sources in a certain power supply area along the time dimension, reflecting the overall power supply capacity of distributed power sources in that area.

[0179] In some embodiments, the electricity load forecast curve refers to the curve showing the change in electricity demand of all users (residential, industrial, commercial, etc.) within the power supply area over time on the forecast day, reflecting the electricity consumption trend of the area.

[0180] As one possible implementation, embodiments of the present invention can ensure that the prediction process is quantifiable and reproducible by clearly defining the complete logical chain of input, calculation, and output, while adapting to the actual needs of power system dispatch.

[0181] In this embodiment of the invention, by filtering and expanding data under extreme weather conditions and filtering data under normal weather conditions, a complete training dataset covering both conventional and extreme scenarios is constructed. This addresses the problem of scarce and unevenly distributed distributed power generation output data under extreme scenarios, laying a data foundation for model learning. During the model training phase, a weighted loss function is constructed based on the scarcity of extreme samples, the severity of prediction errors in different scenarios, and load fluctuation rate. The function is optimized with the goal of minimizing loss. The scarcity of extreme samples represents the weight of extreme samples, amplifying their impact during training and preventing insufficient model fitting of extreme scenario patterns due to insufficient sample size. The severity of prediction errors in different scenarios represents the error penalty coefficient, forcibly correcting prediction biases under extreme scenarios through high penalty strength. Load fluctuation rate represents the sudden changes in output of extreme samples, strengthening the feature weights during periods of high fluctuation and improving the model's ability to capture sudden changes in output patterns under extreme scenarios. The complete dataset and the multi-dimensional loss function work synergistically, enabling the trained daily power generation prediction models for various types of distributed power sources to not only grasp the environmental-output patterns of conventional scenarios but also accurately capture the special characteristics of extreme scenarios, reducing overall prediction bias and ultimately improving the accuracy of daily load prediction.

[0182] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0183] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0184] Figure 2 A schematic diagram of a power system daily load curve prediction device considering distributed generation provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 2 As shown, the power system daily load curve prediction device 2, which takes into account distributed generation, includes: Communication module 21 is used to acquire environmental data of various types of distributed power sources in historical periods; Processing module 22 is used to filter and expand data under extreme environments based on environmental data to construct extreme sample data; to filter data under normal environments based on environmental data to construct normal sample data; to construct a weighted loss function based on the scarcity of extreme sample data, the severity of prediction errors in different scenarios, and load fluctuation rate; to train the model under complete environmental conditions based on extreme sample data and normal sample data with the goal of minimizing the weighted loss function, thereby obtaining daily power generation prediction models for various types of distributed power sources; and to predict the daily load curve of the power system based on the daily power generation prediction models for various types of distributed power sources.

[0185] Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 3 As shown, the electronic device 3 of this embodiment includes a processor 30 and a memory 31. The memory 31 stores a computer program 32. When the processor 30 executes the computer program 32, it implements the steps in the various method embodiments described above. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the various device embodiments described above.

[0186] For example, computer program 32 may be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in electronic device 3.

[0187] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 3 may also include input / output devices, network access devices, buses, etc.

[0188] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0189] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0190] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for predicting the daily load curve of a power system considering distributed generation, characterized in that, include: Acquire environmental data for various types of distributed power sources during historical periods; Based on the environmental data, data filtering and expansion are performed under extreme environments to construct extreme sample data; Based on the environmental data, data is filtered under normal conditions to construct normal sample data; A weighted loss function is constructed based on the scarcity of extreme sample data, the severity of prediction errors in different environments, and load fluctuation. Based on the extreme sample data and the normal sample data, the model is trained under complete environmental conditions with the goal of minimizing the weighted loss function, and daily power generation prediction models for various types of distributed power sources are obtained. Based on the daily power generation prediction model of various types of distributed power sources, the daily load curve of the power system is predicted.

2. The method for predicting the daily load curve of a power system considering distributed generation as described in claim 1, characterized in that, The daily power generation prediction model based on various types of distributed power sources is used to predict the daily load curve of the power system, including: Obtain environmental data for various types of distributed power sources on the day to be predicted. The environmental data is key environmental data that affects the output of distributed power sources. Based on environmental data and daily power generation prediction models for various types of distributed power sources, the daily power generation time series curves of each type of distributed power source are obtained. Based on the daily power output time-series curve, the daily power output time-series curves of each type of distributed power source are calculated according to the power supply area to obtain the total power output curve of the distributed power source. The daily load forecast curve for the power supply area is determined by calculating the difference between the pre-stored power load forecast curve and the total output curve.

3. The method for predicting the daily load curve of a power system considering distributed generation as described in claim 1, characterized in that, The process of filtering and expanding data under extreme environments based on the environmental data to construct extreme sample data includes: Based on environmental data and preset extreme thresholds, determine the extreme data in the environmental data; Based on extreme data, environmental output mapping model and data augmentation model, the extreme data is expanded to obtain the original extreme data; Based on the original extreme data, the extreme sample data is obtained by summarizing and filtering.

4. The method for predicting the daily load curve of a power system considering distributed generation as described in claim 3, characterized in that, The data augmentation model includes an extrapolation sub-model, an adversarial sub-model, and a physical rule-constrained interpolation sub-model. The extreme data, based on the extreme data, environmental output mapping model, and data augmentation model, is expanded to obtain the original extreme data, including: Based on extreme data, extrapolation sub-models, and environmental output mapping models, the extreme data are extrapolated to determine the extrapolated extreme data. Based on extreme data, adversarial sub-models, and environmental output mapping models, adversarial training is performed on extreme data to determine synthetic extreme data. Based on extreme data, physical rule-constrained interpolation sub-models, and environmental output mapping models, extreme data are interpolated to determine the interpolated extreme data. The extrapolated extreme data, the synthesized extreme data, and the interpolated extreme data are integrated to obtain the original extreme data.

5. The method for predicting the daily load curve of a power system considering distributed generation as described in claim 4, characterized in that, The method of extrapolating extreme data based on extreme data, extrapolation sub-models, and environmental output mapping models to determine extrapolated extreme data includes: Environmental characteristics of various distributed power sources are extracted from extreme data. These environmental characteristics are key environmental parameter features that affect the output of distributed power sources. The environmental features are expanded based on preset extreme thresholds and preset step sizes to obtain the expanded environmental features. Based on the extended environmental characteristics and environmental output mapping model, the predicted output of distributed power sources is obtained. Based on the extended environmental characteristics, predicted power output, and physical constraints of distributed power sources, valid data is determined and used as extrapolated extreme data.

6. The method for predicting the daily load curve of a power system considering distributed generation as described in claim 4, characterized in that, The method of conducting adversarial training on extreme data based on extreme data, adversarial sub-models, and environmental output mapping models to determine synthetic extreme data includes: The system is trained based on the output characteristics of distributed power sources in extreme and normal environments, and an adversarial sub-model; wherein the adversarial sub-model is constrained by the theoretical output range of the environmental output mapping model. When the distribution difference between the obtained sample and the extreme sample reaches a preset condition, the confrontation stops, and synthetic extreme data is obtained; The interpolation sub-model based on extreme data, physical rule constraints, and environmental output mapping model interpolates the extreme data to determine the interpolated extreme data, including: Filter data points at both ends of a consecutive extreme time period; Based on the data points at both ends and the environmental output mapping model, the physical correlation coefficient between environmental characteristics and output during this continuous extreme period is obtained. Based on the physical correlation coefficient and the equipment efficiency decay law under continuous extreme environment, the interpolation sub-model constrained by physical rules is used to perform linear interpolation calculation on the time period between the two data points to obtain the environmental characteristics and output data of the middle time period between the two data points. Based on the environmental characteristics and output data of the intermediate time period between the two data points, interpolated extreme data are obtained.

7. The method for predicting the daily load curve of a power system considering distributed generation as described in claim 1, characterized in that, The weighted loss function, constructed based on the scarcity of extreme sample data, the severity of prediction errors in different environments, and load volatility, includes: The total number of samples for complete environmental data is obtained by counting the number of extreme samples and normal samples for each type of distributed power source. The proportion of extreme sample data is calculated based on the number of extreme samples of each type of distributed power source and the total number of samples of complete environmental data. Based on the proportion of extreme sample data, determine the scarcity weight of each extreme sample data; Based on the prediction error and error threshold of distributed power output in normal and extreme environments, the error severity weights for different environments are determined. Based on the load fluctuation rate and maximum load fluctuation rate for each time period, the load fluctuation weight for each time period is calculated. The weighted loss function is determined based on the scarcity weight, the error severity weight, and the load fluctuation weight.

8. The method for predicting the daily load curve of a power system considering distributed generation as described in claim 7, characterized in that, The determination of the weighted loss function based on scarcity weight, error severity weight, and load volatility weight includes: For a single sample data, the scarcity weight, error severity weight, and load volatility weight are multiplied together to obtain the comprehensive weight of the single sample data; The weighted loss function is determined based on the comprehensive weight, the total number of samples of complete environmental data, and the prediction error of distributed power output.

9. A power system daily load curve prediction device considering distributed generation, characterized in that, include: The communication module is used to acquire environmental data of various types of distributed power sources during historical periods. The processing module is used to: filter and expand data under extreme environments based on the environmental data to construct extreme sample data; filter data under normal environments based on the environmental data to construct normal sample data; construct a weighted loss function based on the scarcity of extreme sample data, the severity of prediction errors in different scenarios, and load fluctuation rate; train the model under complete environmental conditions based on the extreme sample data and the normal sample data, with the goal of minimizing the weighted loss function, to obtain daily power generation prediction models for various types of distributed power sources; and predict the daily load curve of the power system based on the daily power generation prediction models for various types of distributed power sources.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 8.

Images