New energy power data enhancement method and related equipment

Abstract

The embodiment of the invention discloses a new energy power data enhancement method and related equipment, and the method comprises the steps: synchronously collecting an accumulated snow coverage rate time sequence and a first power output sequence of a new energy station in a snowy weather; and positioning a target accumulated snow coverage rate time sequence with the accumulated snow coverage change rate exceeding a first preset threshold value in the first preset unit time, and further determining a power data scarce region in the first power output sequence. The sparse starting and ending time period, the snow coverage rate change trend and the first meteorological data of the power data scarce area are obtained, a preset time sequence generative adversarial network is input to generate a second power output sequence, and the second power output sequence can supplement missing and inaccurate data in the first power output sequence; therefore, the third power output sequence obtained by enhancing the first power output sequence by using the second power output sequence has more complete and accurate power output data.

Description

Technical Field

[0001] This invention relates to power data enhancement technology, and more particularly to a method and related equipment for enhancing power data in new energy sources. Background Technology

[0002] With the increasing proportion of renewable energy power generation, accurate power output forecasting is crucial for ensuring the stability of energy supply. However, under extreme weather conditions, such as blizzards, the collection of power output data from renewable energy power plants is significantly affected, such as missing power output data for certain periods or inaccurate power values. This results in low accuracy of power forecasting during blizzards using this power data. Summary of the Invention

[0003] In view of this, the present invention provides a method and related equipment for enhancing new energy power data.

[0004] The specific technical solution of the first embodiment of the present invention is as follows: a method for enhancing new energy power data, the method comprising: synchronously collecting a time series sequence of snow cover rate and a first power output sequence of a new energy power station during a blizzard; obtaining a target snow cover rate time series sequence in the snow cover rate time series where the snow cover change rate exceeds a first preset threshold within a first preset unit time; determining a power data sparse region in the first power output sequence based on the target snow cover rate time series sequence; the power data sparse region being a time range in which the number of power output data collection points is less than a second preset threshold within a preset second unit time; acquiring the sparse start and end time periods, the snow cover rate change trend, and first meteorological data of the power data sparse region, and inputting the sparse start and end time periods, the snow cover rate change trend, and the first meteorological data into a preset time series generative adversarial network to generate a second power output sequence of the power data sparse region; the meteorological data including at least temperature and humidity; and enhancing the first power output sequence using the second power output sequence to obtain an enhanced third power output sequence.

[0005] Preferably, the step of enhancing the first power output sequence with the second power output sequence to obtain the enhanced third power output sequence includes: inserting the second power output sequence into the sparse start and end time period of the first power output sequence to obtain the third power output sequence.

[0006] Preferably, determining the power data sparse region in the first power output sequence based on the target snow coverage time series includes: determining the duration corresponding to the target snow coverage time series; and obtaining the power data sparse region within the duration in the first power output sequence.

[0007] Preferably, the step of inputting the sparse start and end time periods, the snow cover change trend, and the first meteorological data into a preset temporal generative adversarial network to generate a second power output sequence for the power data-scarce region includes: extracting the power output range corresponding to the target time period as a boundary condition in the first power output sequence; the target time period is the time period before and after the sparse start and end time periods after being extended by a preset length; constructing input data based on the boundary conditions, the snow cover change trend, and the first meteorological data; and concatenating the input data with a preset random noise vector and inputting it into the preset temporal generative adversarial network to generate a second power output sequence for the power data-scarce region.

[0008] Preferably, after obtaining the enhanced third power output sequence, the method further includes: acquiring second meteorological data of the new energy power station during a blizzard; fusing the snow cover time series sequence, the third power output sequence, the second meteorological data, and a preset weighted value corresponding to each second meteorological data to obtain a multi-dimensional enhanced output sequence; extracting the fourth power output sequence corresponding to the sparse start and end time periods from the enhanced output sequence, and extracting the fifth power output sequence corresponding to the sparse start and end time periods from the first output sequence; calculating the similarity between the fourth power output sequence and the fifth power output sequence; if the similarity is less than a third preset threshold, then the fourth power output sequence is taken as the power data scarce region, and the method of acquiring the sparse start and end time periods, snow cover change trend, and first meteorological data of the power data scarce region is returned, and the sparse start and end time periods, snow cover change trend, and first meteorological data are input into a preset time series generative adversarial network to generate the second power output sequence of the power data scarce region, until the similarity is not less than the third preset threshold, and the last obtained enhanced third power output sequence is the fourth power output sequence.

[0009] Preferably, after obtaining the fourth power output sequence, the method further includes: obtaining a first dominant frequency and a first dominant frequency amplitude of the fourth power output sequence; the first dominant frequency is the frequency with the largest amplitude in the fourth power output sequence; the first dominant frequency and the first dominant frequency amplitude constitute a first feature vector; obtaining a second dominant frequency and a second dominant frequency amplitude of the first power output sequence; the second dominant frequency is the frequency with the largest amplitude in the first power output sequence; the second dominant frequency and the second dominant frequency amplitude constitute a second feature vector; calculating the Euclidean distance between the first feature vector and the second feature vector; using the least squares method to solve the linear transformation matrix formed by the first feature vector and the second feature vector with the goal of minimizing the Euclidean distance, obtaining feature adjustment parameters of the fourth power output sequence; the feature adjustment parameters are the dominant frequency adjustment parameters and the dominant frequency amplitude adjustment parameters; and optimizing the fourth power output sequence using the feature adjustment parameters to obtain a fifth power output sequence.

[0010] Preferably, the method further includes: obtaining a power decrease output sequence in the first power output sequence; then obtaining the second dominant frequency and the second dominant frequency amplitude of the first power output sequence includes: obtaining the second dominant frequency and the second dominant frequency amplitude of the power decrease output sequence; the second dominant frequency is the frequency with the largest amplitude in the power decrease output sequence.

[0011] The specific technical solution of the second embodiment of the present invention is as follows: a new energy power data enhancement system, the system comprising: a data acquisition module, a target sequence extraction module, a scarce region extraction module, a sequence generation module, and an enhancement module; the data acquisition module is used to simultaneously acquire the snow cover rate time series sequence and the first power output sequence of new energy power stations during blizzard weather; the target sequence extraction module is used to obtain a target snow cover rate time series sequence in the snow cover rate time series sequence where the snow cover change rate exceeds a first preset threshold within a first preset unit time; the scarce region extraction module is used to determine the power data scarce region in the first power output sequence based on the target snow cover rate time series sequence. The power data scarce region is defined as the time range in which the number of power output data collection points within a preset second unit time period is less than a second preset threshold. The sequence generation module is used to acquire the sparse start and end time periods, snow cover rate change trend, and first meteorological data of the power data scarce region, and input the sparse start and end time periods, snow cover rate change trend, and first meteorological data into a preset temporal generative adversarial network to generate a second power output sequence of the power data scarce region. The meteorological data includes at least temperature and humidity. The enhancement module is used to enhance the first power output sequence using the second power output sequence to obtain an enhanced third power output sequence.

[0012] The specific technical solution of the third embodiment of the present invention is as follows: a new energy power data enhancement device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method as described in any one of the first embodiments of this application.

[0013] The specific technical solution of the fourth embodiment of the present invention is as follows: a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor performs the steps of the method as described in any one of the first embodiments of this application.

[0014] Implementing the embodiments of the present invention will have the following beneficial effects: This invention synchronously collects time-series data on snow cover rate and a first power output sequence from renewable energy power plants during blizzards. It then identifies target snow cover rate time-series sequences where the snow cover change rate exceeds a first preset threshold within a first preset unit time, thereby determining power data sparse regions in the first power output sequence. The invention acquires sparse start and end times, snow cover rate change trends, and first meteorological data for these power data sparse regions. These are then input into a preset time-series generative adversarial network to generate a second power output sequence. This second power output sequence supplements the missing and inaccurate data in the first power output sequence. Therefore, the third power output sequence obtained by enhancing the first power output sequence with the second power output sequence possesses more complete and accurate power output data. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the steps of a method for enhancing new energy power data; Figure 2 This is a schematic diagram of the structure of a new energy power data enhancement system; Among them, 201 is the data acquisition module; 202 is the target sequence extraction module; 203 is the scarce region extraction module; 204 is the sequence generation module; and 205 is the enhancement module. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] The terms "first," "second," etc., used in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or modules is not limited to the listed steps or modules, but may optionally include steps or modules not listed, or may optionally include other steps or modules inherent to such processes, methods, products, or apparatus.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] Please see Figure 1 The above is a flowchart of the steps of a new energy power data enhancement method according to the first embodiment of this application, in order to obtain more complete and accurate power output data. The method includes: Step 101: Synchronously collect the time series sequence of snow cover rate and the first power output sequence of the new energy power station during the blizzard weather; Step 102: Obtain the target snow cover rate time series from the snow cover rate time series where the snow cover rate of change exceeds the first preset threshold within a first preset unit time. Step 103: Determine the power data sparse region in the first power output sequence based on the target snow coverage time series; the power data sparse region is the time range in which the number of power output data collection points is less than a second preset threshold within a preset second unit time. Step 104: Obtain the sparse start and end time periods, snow cover rate change trend, and first meteorological data for the power data-scarce region, and input the sparse start and end time periods, snow cover rate change trend, and first meteorological data into a preset time-series generative adversarial network to generate a second power output sequence for the power data-scarce region; the meteorological data includes at least temperature and humidity; Step 105: Enhance the first power output sequence using the second power output sequence to obtain an enhanced third power output sequence.

[0021] Specifically, the snow cover rate time series and the first power output sequence of the station are collected simultaneously during blizzard weather. For example, snow cover rate and power output data are recorded every 10 minutes for n hours. From the snow cover rate time series, target snow cover rate time series with a snow cover change rate exceeding a first preset threshold within a first preset unit time are selected. Analysis revealed that the snow cover change rate exceeded the threshold within three 1-hour time periods. Based on the target snow cover rate time series, power data sparse regions in the first power output sequence are identified. The second unit time is preset to 30 minutes. When the number of power output data collection points within 30 minutes is less than the second preset threshold, it is determined to be a power data sparse region. It was determined that there are two power data sparse regions. The sparse start and end time periods, snow cover rate change trends, and first meteorological data (temperature, humidity) of these two power data sparse regions are obtained. The above information is input into a preset time series generative adversarial network, which is trained on a large amount of historical data and can generate reasonable power output sequences based on the input. After the temporal generative adversarial network is run, a second power output sequence is generated for regions with scarce power data. Finally, the corresponding regions in the first power output sequence are enhanced using the second power output sequence to obtain an enhanced third power output sequence. This makes the power output data more complete and accurate, providing a reliable basis for the stable operation analysis of new energy power plants.

[0022] The method in this embodiment synchronously collects the snow cover rate time series and the first power output sequence of new energy power stations during blizzard weather. It then locates the target snow cover rate time series where the snow cover change rate exceeds a first preset threshold within a first preset unit time, thereby identifying power data-scarce regions in the first power output sequence. The sparse start and end time periods, snow cover rate change trends, and first meteorological data of these power data-scarce regions are acquired and input into a preset time series generative adversarial network to generate a second power output sequence. This second power output sequence supplements the missing and inaccurate data in the first power output sequence. Therefore, the third power output sequence obtained by enhancing the first power output sequence with the second power output sequence possesses more complete and accurate power output data.

[0023] In a specific embodiment, enhancing the first power output sequence with the second power output sequence to obtain an enhanced third power output sequence includes: inserting the second power output sequence into the sparse start and end time periods of the first power output sequence to obtain the third power output sequence. By inserting the second power output sequence, data gaps in the first power output sequence are filled, making the power output data more continuous and complete in the time dimension, providing a more comprehensive data foundation for subsequent analysis.

[0024] In a specific embodiment, determining the power data sparse region in the first power output sequence based on the target snow cover time series includes: determining the duration corresponding to the target snow cover time series; and obtaining the power data sparse region within the duration in the first power output sequence. By first determining the duration of the target snow cover time series and then searching for the power data sparse region in the corresponding first power output sequence, the missing data segments that may affect the power output due to snow cover changes caused by blizzards can be located more accurately, avoiding blind searching and improving the efficiency and accuracy of data processing.

[0025] In a specific embodiment, the step of inputting the sparse start and end time periods, the snow cover change trend, and the first meteorological data into a preset temporal generative adversarial network to generate a second power output sequence for the power data-scarce region includes: extracting the power output range corresponding to the target time period as a boundary condition in the first power output sequence; the target time period is the time period before and after the sparse start and end time periods after being extended by a preset length; constructing input data based on the boundary conditions, the snow cover change trend, and the first meteorological data; and concatenating the input data with a preset random noise vector and inputting it into the preset temporal generative adversarial network to generate a second power output sequence for the power data-scarce region.

[0026] Specifically, if a region with scarce power data has a sparse start-end time period from hour 20 to hour 21:30, this time period is extended by one hour before and after, resulting in a target time period from hour 19 to hour 22:30. In the first power output sequence, the power output range corresponding to this target time period is extracted, with a minimum power of 100kW and a maximum power of 800kW, serving as boundary conditions. Simultaneously, the snow cover change trend (e.g., rapidly increasing from 30% to 60%) and first meteorological data (e.g., temperature -5℃, humidity 80%) for this power-scarce region are acquired. Input data is constructed based on the boundary conditions, snow cover change trend, and first meteorological data. This input data is concatenated with a preset random noise vector and then fed into a preset temporal generative adversarial network (GAN). This network, trained on a large amount of historical data, can generate a reasonable power output sequence based on the input. After network operation, a second power output sequence for this power-scarce region is generated. Finally, the second power output sequence is inserted into the sparse start-end time period of the first power output sequence to obtain an enhanced third power output sequence. By extracting the power output range corresponding to the target time period as boundary conditions, a clear constraint range is provided for the time-series generative adversarial network. This makes the generated second power output sequence more consistent with the actual output characteristics of renewable energy power plants within that time period, avoiding the generation of unreasonable large fluctuations in data and improving the rationality of the generated data.

[0027] In a specific embodiment, after obtaining the enhanced third power output sequence, the method further includes: acquiring second meteorological data of the new energy power station during a blizzard period; fusing the snow cover time series sequence, the third power output sequence, the second meteorological data, and a preset weighting value corresponding to each second meteorological data to obtain a multi-dimensional enhanced output sequence; extracting the fourth power output sequence corresponding to the sparse start and end time periods from the enhanced output sequence, and extracting the fifth power output sequence corresponding to the sparse start and end time periods from the first output sequence; calculating the similarity between the fourth power output sequence and the fifth power output sequence; if the similarity is less than a third preset threshold, then the fourth power output sequence is taken as the power data sparse region, and the method of acquiring the sparse start and end time periods, snow cover change trend, and first meteorological data of the power data sparse region is returned, and the sparse start and end time periods, snow cover change trend, and first meteorological data are input into a preset time series generative adversarial network to generate the second power output sequence of the power data sparse region, until the similarity is not less than the third preset threshold, and the last obtained enhanced third power output sequence is the fourth power output sequence.

[0028] Specifically, sparse start and end time periods, snow cover change trends, and first meteorological data (temperature, humidity) are acquired from regions with scarce power data. These are input into a preset time-series generative adversarial network (GAN) to generate corresponding second power output sequences. These sequences are then inserted into the first power output sequence to obtain an enhanced third power output sequence. Second meteorological data, including wind speed, wind direction, and air pressure, is acquired from the new energy power plants during blizzard periods. The snow cover time-series sequence, the third power output sequence, the second meteorological data, and the preset weighted values ​​corresponding to each second meteorological data point are fused together, and a weighted average is used to calculate a multi-dimensional enhanced output sequence. A fourth power output sequence corresponding to the sparse start and end time periods is extracted from the enhanced output sequence, and a fifth power output sequence corresponding to the sparse start and end time periods is extracted from the first output sequence. The cosine similarity algorithm is used to calculate the similarity between the fourth and fifth power output sequences. Assuming the third preset threshold is 0.8, if the calculated similarity is 0.7, which is less than the third preset threshold, the fourth power output sequence is used as correction data for the sparse power data region. The sparse start and end time periods, snow cover change trends, and first meteorological data for that region are then re-acquired and input into the preset time-series generative adversarial network to generate the second power output sequence, updating the enhanced third power output sequence. This process is repeated until the calculated similarity is not less than 0.8. The last enhanced third power output sequence obtained is the fourth power output sequence. By calculating the similarity between the fourth and fifth power output sequences and performing iterative corrections, deviations between the generated second power output sequence and the actual situation can be detected in a timely manner. When the similarity is less than the third preset threshold, data is regenerated until the requirement is met, thereby effectively improving the accuracy of the generated data and making the enhanced power output sequence closer to the true value.

[0029] In a specific embodiment, after obtaining the fourth power output sequence, the method further includes: obtaining a first dominant frequency and a first dominant frequency amplitude of the fourth power output sequence; the first dominant frequency is the frequency with the largest amplitude in the fourth power output sequence; the first dominant frequency and the first dominant frequency amplitude constitute a first feature vector; obtaining a second dominant frequency and a second dominant frequency amplitude of the first power output sequence; the second dominant frequency is the frequency with the largest amplitude in the first power output sequence; the second dominant frequency and the second dominant frequency amplitude constitute a second feature vector; calculating the Euclidean distance between the first feature vector and the second feature vector; using the minimum Euclidean distance as the objective, solving the linear transformation matrix formed by the first feature vector and the second feature vector using the least squares method to obtain the feature adjustment parameters of the fourth power output sequence; the feature adjustment parameters are the dominant frequency adjustment parameters and the dominant frequency amplitude adjustment parameters; optimizing the fourth power output sequence using the feature adjustment parameters to obtain a fifth power output sequence. For each frequency component in the fourth power output sequence, corresponding adjustments are made according to the dominant frequency adjustment parameters and the dominant frequency amplitude adjustment parameters to obtain a perfect fifth power output sequence.

[0030] In a specific embodiment, the method further includes: obtaining a power decline output sequence in the first power output sequence; then obtaining the second dominant frequency and the second dominant frequency amplitude of the first power output sequence includes: obtaining the second dominant frequency and the second dominant frequency amplitude of the power decline output sequence; the second dominant frequency is the frequency with the largest amplitude in the power decline output sequence. Obtaining the power decline output sequence in the first power output sequence and obtaining the second dominant frequency and the second dominant frequency amplitude for these sequences can focus on the key stages of power changes at new energy power plants under blizzard conditions. Power decline is often closely related to factors such as equipment operating status and meteorological conditions. Characteristic analysis of these stages can provide a deeper understanding of the performance of the power plant under adverse weather conditions. Compared with characteristic analysis of the entire first power output sequence, analysis of the power decline output sequence is more targeted. During the power decline stage, the change pattern and influencing factors of power output may differ from those of the normal stage. By analyzing the characteristics of these stages separately, the frequency and amplitude characteristics related to power decline can be captured more accurately, providing a more precise basis for subsequent optimization.

[0031] In a specific embodiment, a complete embodiment of the new energy power data enhancement method is shown in S1-S7.

[0032] S1. Collect time-series sequences of snow cover rate and corresponding power output from new energy power plants during blizzard weather, extract segments of drastic changes in snow cover rate and power decline segments, and generate a preliminary monitoring dataset: Obtain snowfall intensity data during blizzard weather using meteorological radar, calculate the hourly snow cover rate increment based on the correlation between snowfall intensity and snow accumulation rate on photovoltaic panel surfaces, and activate a high-frequency monitoring mode when the snow cover rate increment exceeds a first preset threshold, recording snow cover rate values ​​and corresponding power output values ​​at preset time intervals to obtain time-series monitoring data. Use a sliding window method to segment the time-series monitoring data, setting the window duration and sliding step size. For each data point within a window, fit a curve of snow cover rate change over time using the least squares method, and obtain the slope of the fitted curve as the rate of change of coverage rate. If the rate of change of coverage rate exceeds a second preset threshold and the power output decline during the same period exceeds a third preset threshold, then mark the time window as a segment of drastic change. Based on the timestamp information of the drastic change segments, temperature, humidity and wind speed monitoring data for the corresponding time periods are extracted. The Pearson correlation coefficient between each meteorological factor and the power decrease is calculated. Dramatic change segments with correlation coefficients exceeding the fourth preset threshold are selected. The selected drastic change segments and their corresponding snow cover rate sequences and power output sequences are combined to construct a preliminary monitoring dataset.

[0033] Specifically, in one implementation, a weather radar detects snowfall intensity by emitting electromagnetic waves and receiving reflected signals, the snowfall intensity being quantified in millimeters per hour.

[0034] Specifically, based on the relationship between radar echo intensity and snow particle density, a mapping table is established between snowfall intensity and snow accumulation rate on the photovoltaic panel surface. This mapping table takes into account the influence of ambient temperature on snowflake density. When the temperature is close to zero degrees, the accumulation rate of wet snow is relatively fast.

[0035] It should be noted that the calculation of the snow cover increment is based on the rate of change of the effective light-receiving area of ​​the photovoltaic panel. The snow distribution on the surface of the photovoltaic panel is monitored in real time using image recognition technology, and the ratio of the snow-covered area to the total area is used as the coverage rate. When the detected snow cover increment exceeds a first preset threshold, the monitoring system automatically switches from the regular hourly sampling to a high-frequency monitoring mode. The preset time interval can be dynamically adjusted according to the intensity of the blizzard, and is usually set to 5 to 15 minutes.

[0036] Preferably, in the implementation of the sliding window method, the choice of window duration needs to balance data accuracy and computational efficiency.

[0037] For example, in a typical blizzard event, a window duration of 30 minutes and a sliding step of 10 minutes can be set to capture the rapid changing trend of snow cover. The specific implementation of the least squares fitting method is to construct a linear regression model of time and snow cover, minimizing the sum of squared errors between the observed and fitted values. The slope of the resulting fitted curve directly reflects the rate of change of snow cover.

[0038] In one possible implementation, the Pearson correlation coefficient is calculated as the ratio of the product of the covariance and the standard deviation of the meteorological factor series and the power decline magnitude series.

[0039] For example, the correlation coefficient between temperature series and power decline reflects the characteristic that snow is not easy to melt under low temperature conditions, resulting in persistently low power, while the correlation coefficient of wind speed series reflects the law that strong winds may blow away some snow, thus affecting power recovery.

[0040] Understandably, the construction of the preliminary monitoring dataset not only includes the time series data itself, but also integrates meteorological background information, forming a multi-dimensional data structure. Each data sample includes a timestamp, snow cover sequence, power output sequence, and corresponding meteorological parameters such as temperature, humidity, and wind speed.

[0041] S2. Analyze the preliminary monitoring dataset to determine the intervals of drastic changes in snow cover rate. Based on these intervals, determine the location and range of power data sparse areas and generate sparse area identifiers. Extract time-series snow cover rate data from the preliminary monitoring dataset, calculate the difference in coverage rate values ​​between adjacent moments, divide by the corresponding time interval to obtain the coverage rate change rate at each moment. When the coverage rate change rate at multiple consecutive moments exceeds a preset rate threshold, record the start and end times of this consecutive time period to determine the intervals of drastic changes in snow cover rate. For each interval of drastic changes in snow cover rate, count the number of sampling points for power output data from new energy power plants within the interval, calculate the sampling density per unit time. If the sampling density is lower than a preset density threshold or there are consecutive data gaps exceeding a preset duration, determine that the interval is a power data sparse area, and record the start and end time coordinates and data missing characteristics of the sparse area. Based on the start and end time coordinates and data missing characteristics, identification information including region number, start timestamp, end timestamp, and missing rate value is constructed. A unique identifier code for each scarce region is generated by string concatenation and numerical encoding. The identifier codes of all scarce regions and their corresponding coverage change information are summarized to obtain the scarce region identifier.

[0042] Specifically, the calculation of the snow cover rate change rate is achieved using the difference method. The process involves sorting the snow cover rate time series data according to the timestamp, subtracting the coverage rate values ​​of two adjacent sampling points to obtain the coverage rate increment, and then dividing by the time interval between the two sampling points to obtain the average rate of change within that period.

[0043] For example, during a blizzard, if the snow cover is 25% at 10:00 AM and 38% at 10:05 AM, the rate of change within those 5 minutes is 2.6% per minute. In one implementation, a sliding detection window mechanism is used to determine whether the rate of change at multiple consecutive moments exceeds a threshold. The window size is set to 3 to 5 consecutive sampling points. When the rate of change calculated from all sampling points within the window is greater than a preset rate threshold, the time corresponding to the starting position of the window is recorded as the start time of the drastic change interval. The window continues to slide backward until a sampling point with a rate lower than the threshold appears, and the time of the previous sampling point is recorded as the end time.

[0044] Specifically, determining areas of scarce power data involves an assessment across two dimensions. The first dimension is sampling density assessment, quantified by the ratio of the actual number of sampling points per unit time to the standard sampling frequency. The standard sampling frequency is determined according to the monitoring specifications for new energy power plants, typically one sampling point every 5 minutes. The second dimension is data continuity assessment, detecting the existence of continuous data gaps, which refer to periods where multiple expected sampling times are not actually recorded. If either dimension fails to meet the requirements, the area is determined to be a scarce region.

[0045] Preferably, the construction process of scarce region identifiers includes two stages: information organization and encoding. In the information organization stage, the region serial number, start timestamp, end timestamp, and missing rate value are arranged in a fixed format to form structured identifier information. In the encoding stage, a string concatenation method is used to convert the region serial number into a two-digit string, the timestamp into a standard format string, and the missing rate to two decimal places and converted to a percentage. The parts are connected by underscores to form a unique identifier code such as "01_202501120800_202501121000_75".

[0046] In one possible implementation, the record of coverage change information includes the maximum rate of change, the average rate of change, and the cumulative change within the interval. This information is stored in association with an identifier code, providing important feature references for subsequent data augmentation and model training.

[0047] The power drop segment itself is not directly used to calculate the sampling density, but it helps identify critical time intervals (i.e., intervals of drastic changes in snow cover). Within these intervals, the system counts the number of sampling points for power data: for example, if the time period corresponding to the power drop segment is [t1, t2], then the system counts the number of sampling points for all power output data within that interval. The sampling density per unit time (e.g., the number of sampling points per minute or every 5 minutes) is then calculated.

[0048] Data missing characteristics include sampling density values ​​(e.g., number of sampling points per unit time), missing rate values ​​(e.g., percentage), start and end times of consecutive data gaps (if any), and other characteristics such as the duration of the data gap segment. Statistical sampling density: Within intervals of drastic changes in snow cover, the actual number of sampling points for power output data is counted, and the sampling density per unit time is calculated (e.g., number of sampling points divided by interval duration). Preset density thresholds are set based on monitoring specifications for new energy power plants (e.g., a standard sampling frequency of one point every 5 minutes; if the actual density is lower than this frequency, it is considered low density). Detecting consecutive data gaps: Checking for consecutive data gap segments exceeding a preset duration within the interval (e.g., 15 consecutive minutes without any power data records). This is achieved through continuity analysis of the timestamp sequence. Calculating the missing rate: The missing rate can be calculated by comparing the actual number of sampling points with the expected number of sampling points (based on the standard sampling frequency). For example, if the interval duration is 60 minutes and the standard sampling frequency is one point every 5 minutes, the expected number of sampling points is 12; if there are only 6 actual points, the missing rate is 50%.

[0049] S3. Process the scarce region identifier and the power output sequence using a temporal generative adversarial network (GAN) to generate a supplementary power data sequence, and merge them to generate an expanded power dataset: Parse the start and end positions of the missing time period based on the scarce region identifier. Extract power data of preset lengths before and after the missing time period from the existing power output sequence of new energy power plants as boundary conditions. Construct conditional input data including timestamps, snow cover change trends, and meteorological parameters. Concatenate the conditional input data with a preset random noise vector and input it into the temporal generative adversarial network. Process the data through multiple layers of neurons in the generative network to obtain an initial synthetic power sequence. Input the initial synthetic power sequence and the actual power output sequence into a discriminant network, respectively. Calculate the probability value of the discriminant network output. If the probability value is lower than a preset threshold, calculate the mean square error between the generative network output and the actual data as a loss value. Update the generative network weights using gradient descent based on the loss value to regenerate the power sequence until the probability value of the discriminant network output reaches the preset threshold, thus obtaining the supplementary power sequence. The supplementary power sequence is subjected to a continuity test, calculating whether the rate of change of power between adjacent time points exceeds a preset range. If abrupt changes exceeding the range are found, local sequence segments before and after the abrupt change are extracted, and cubic spline interpolation is used to smooth the abrupt change, resulting in a smoothed power sequence. The smoothed power sequence is then inserted into the corresponding missing positions of the existing renewable energy power plant power output sequence according to timestamp order. After merging, the total number of data points and time span are statistically analyzed. If overlapping data points are found, the original data is retained and the generated data is deleted, resulting in an expanded renewable energy power plant output power dataset.

[0050] The specific steps for generating a supplementary power sequence based on network weights are as follows: Step 1: Concatenate conditional data (such as boundary power, timestamps, and meteorological parameters obtained from sparse region identifiers) with a random noise vector to form the input data. This input data is fed into the generator network. Each layer and each neuron of the generator network (usually a recurrent neural network such as LSTM or GRU) performs a linear weighted summation of the input data according to its current weights and biases, and then performs a non-linear transformation through an activation function. Starting from the first time step, the network processes the sequence step by step, outputting a power value at each time step. The outputs of all time steps are concatenated to form the initial synthesized power sequence. At this point, the quality of this sequence depends entirely on the quality of the current weights of the generator network.

[0051] Step 2: Input the generated initial sequence and the true power sequence (extracted from existing data) into the discriminant network (usually a one-dimensional CNN or RNN). The discriminant network uses its current weights to extract features from the sequence and finally outputs a "probability of authenticity" between 0 and 1 through a sigmoid function. The closer the probability is to 1, the more authentic the discriminant network considers the sequence to be.

[0052] Step 3: This is the most crucial step: updating the generator network weights based on probability values. The loss function quantifies the "failure" of the generator network. For example, if the generated sequence is scored 0.1 by the discrimination network (i.e., judged as false), while the goal is 1 (hoping the discrimination network judges it as true), then a large loss value will be generated. A commonly used loss function is binary cross-entropy. Calculate the gradient of the loss function with respect to all weights of the generator network. The gradient indicates the direction and magnitude by which each weight should be adjusted to reduce the loss (i.e., make the generated sequence more realistic). Using an optimizer (such as Adam), update the generator network weights according to the calculated gradient and the preset learning rate. New weight = old weight - learning rate × gradient. With the generator network using the updated weights, return to step 1 and regenerate a new power sequence. Since the weights have been fine-tuned, the generated sequence this time should theoretically be closer to the real data.

[0053] Step 4: The above process is repeated (usually thousands or even millions of times) until a stopping condition is met (such as the discrimination network being unable to distinguish between true and false signals, i.e., the probability stabilizes at around 0.5; or the preset number of iterations is reached). The supplementary power sequence is then subjected to a continuity test, smoothing out abrupt changes exceeding a preset range to generate a smoothed power sequence. This smoothed power sequence is then inserted into the missing positions of the power output sequence, and the results are merged to generate the expanded power dataset.

[0054] Specifically, the start and end positions parsed from the scarce region identifiers are used to locate the missing time period. Power data for 30 minutes before and after this time period are extracted as boundary conditions. These boundary data contain trend information on power changes. The snow cover change trend is obtained by calculating the slope of the snow cover before the missing time period. Meteorological parameters include time series of temperature, humidity, and wind speed. These multidimensional information are aligned by time and concatenated into a condition vector with a dimension of 64. This vector is then merged with a 128-dimensional random noise vector to form a 192-dimensional network input.

[0055] Preferably, the forward propagation process of the generator network employs a gated recurrent unit (ROU) mechanism, controlling information flow through update and reset gates. When conditional input data enters the network, the update gate determines how much historical information to retain based on the current input and the hidden state of the previous time step, while the reset gate determines how much past state to ignore. This gating mechanism allows the network to selectively memorize important power change patterns, especially during critical periods when rapidly increasing snow cover leads to a sharp drop in power. Through multi-layered recursive computation, the network progressively transforms random noise into a power sequence that conforms to the true distribution, with the output at each time step relying on the accumulated information from all previous time steps.

[0056] For example, the probability calculation for authenticity in the discrimination network employs a binary classification output mechanism. After receiving the power sequence, the network extracts local features through three layers of one-dimensional convolutions with kernel sizes of 3, 5, and 7 to capture patterns at different time scales. The extracted features are then processed by global average pooling and input into a fully connected layer. Finally, a sigmoid activation function outputs a probability value between 0 and 1, where a value close to 1 indicates that the data is real, and a value close to 0 indicates that the data is generated.

[0057] Specifically, the loss value is calculated using a weighted combination of adversarial loss and reconstruction loss. The adversarial loss uses binary cross-entropy to measure the classification accuracy of the discriminant network, while the reconstruction loss uses mean squared error to measure the continuity between the generated sequence and the boundary conditions. During gradient descent updates, the discriminant network parameters are first fixed, the gradient of the generator network is calculated and updated; then the generator network parameters are fixed again, and the discriminant network is updated. This alternating training method continues until the discriminant network can no longer distinguish between real and generated data, at which point the generated power sequence already possesses the statistical characteristics of real data.

[0058] In one possible implementation, continuity testing is performed by calculating the standard deviation of the power change rate. For the generated power sequence, the ratio of the power difference between any two adjacent moments to the time interval is calculated to obtain the instantaneous change rate sequence. If the change rate at a certain moment exceeds three times the standard deviation of the mean change rate in the historical data, it is marked as a breakpoint.

[0059] It should be noted that the application of cubic spline interpolation ensures the smoothness of the power curve. During interpolation, five data points before and after the abrupt change point are used as control points to construct a cubic polynomial function. The function's values ​​at the control points must be consistent with the original data, and its first and second derivatives must be continuous. The polynomial coefficients are obtained by solving a system of linear equations, and the interpolated function is used to replace the data in the abrupt change interval, achieving a smooth transition. This approach maintains the overall trend of power variation while eliminating unreasonable jumps.

[0060] For example, during data supplementation for a blizzard event, the original power sequence was missing between 10:30 and 11:30. The generating network generated 60 data points based on the power decline trend from 10:00 to 10:30 and the low power maintenance state after 11:30. The generated sequence showed that the power gradually decreased from 500 kW at 10:30 to 50 kW at 11:00, and then remained at a low level, which is consistent with the physical process of rapid snow accumulation leading to the complete coverage of photovoltaic panels.

[0061] Understandably, the data merging process employs a timestamp indexing mechanism to precisely insert the generated power sequence into the missing positions of the original sequence. When timestamp overlap is detected, the original, authentic data is prioritized and preserved by comparing data source markers, ensuring the authenticity of the dataset is not compromised. The resulting extended dataset contains the complete time series.

[0062] S4. Extract power statistical features from the expanded power dataset, determine the uniformity of the distribution of the power statistical features, and adjust the uneven power statistical features according to historical blizzard weather power change records to generate an adjusted power statistical feature set: calculate the mean, variance, skewness, and kurtosis statistics of the power sequence from the expanded new energy power station output power dataset, and simultaneously calculate the Pearson correlation coefficient between the snow cover rate sequence and the power decrease rate sequence. The difference between the maximum and minimum values ​​of the statistical power in different coverage intervals is used as the fluctuation range. Combine the statistics into a probability statistical feature vector to obtain an initial feature set. Use the Gaussian kernel density estimation method to calculate the probability density distribution of each feature in the initial feature set within its value range. Divide the value range into multiple intervals and calculate the density value in each interval. Determine the distribution uniformity by calculating the standard deviation of the density values ​​of all intervals. If the standard deviation exceeds a preset threshold, retrieve power change data similar to the current meteorological conditions from historical blizzard weather records and extract the mean and variance parameters of its feature distribution. A normal distribution is constructed based on the mean and variance parameters as the target distribution. The cumulative distribution function of the initial feature set is calculated. The initial feature values ​​are mapped to the corresponding feature values ​​of the target distribution through the cumulative distribution function matching method. Feature values ​​that exceed the physical reasonable range after mapping are truncated so that the power value is kept between zero and the rated power, thus obtaining the adjusted power statistical feature set.

[0063] Specifically, in one implementation, the extraction process of power statistical features involves the calculation of statistics across multiple dimensions. The mean reflects the average level of power output, the variance measures the dispersion of power fluctuations, the skewness describes the asymmetry of the power distribution, and the kurtosis characterizes the sharpness of the distribution.

[0064] Specifically, when blizzards cause snow to accumulate rapidly, the power sequence will exhibit a negatively skewed distribution, meaning that most values ​​are concentrated in the low-power region, with negative skewness values ​​and large absolute values.

[0065] It should be noted that the Pearson correlation coefficient reveals the strength of the linear relationship between snow cover and power decline. In actual calculations, after standardizing the snow cover and power series, the covariance of the two is calculated and divided by the product of their respective standard deviations. When the correlation coefficient is close to -1, it indicates a strong negative correlation between increased snow cover and decreased power, a relationship that is particularly pronounced during blizzards.

[0066] Preferably, the Gaussian kernel density estimation method estimates the probability density by placing a Gaussian kernel function at each data point and then superimposing all kernel functions. The bandwidth parameter of the kernel function determines the smoothness of the estimation; too small a bandwidth will result in an overly coarse estimation, while too large a bandwidth will lead to over-smoothing and loss of detail. When determining the uniformity of the distribution, the eigenvalue domain is divided into 10 equally wide intervals, and the density integral value for each interval is calculated. The standard deviation of these density values ​​reflects the degree of non-uniformity of the distribution.

[0067] For example, when uneven distribution is detected, data with similar conditions are retrieved from historical blizzard weather records. Similarity is determined by comparing the Euclidean distance of meteorological parameters, including dimensions such as snowfall intensity, ambient temperature, and duration. The retrieved historical data provides a reference template for the target distribution, with its mean and variance parameters used to construct an ideal normal distribution.

[0068] In one possible implementation, the cumulative distribution function matching process first calculates the empirical cumulative distribution function of the initial feature set, then finds the probability value of each feature value in the cumulative distribution, and finally maps it to the new feature value through the inverse cumulative distribution function of the target normal distribution. This mapping preserves the relative order of the data while adjusting the distribution shape.

[0069] Specifically, truncation ensures that the mapped power values ​​conform to physical constraints. When the mapped value is less than zero, it is set to zero; when the mapped value exceeds the station's rated power, it is limited to the rated power value. This approach ensures both the physical validity of the data and the effectiveness of the statistical characteristics.

[0070] S5. The adjusted power statistical feature set and the preliminary monitoring dataset are fused to generate a comprehensive enhanced data sequence. The continuity of power changes in the comprehensive enhanced data sequence is analyzed to generate an optimized enhanced dataset: The adjusted power statistical feature set is aligned with the preliminary monitoring dataset according to the timestamp index. The fusion weight is determined based on the correlation coefficient between each feature and the power output. A weighted average method is used to embed the statistical feature values ​​into the monitoring data at the corresponding time, forming a multi-dimensional data structure containing the original monitoring values ​​and statistical features, resulting in a preliminary fusion sequence. Continuity detection is performed on the preliminary fusion sequence. The square root of the sum of the squares of the differences between adjacent data points in the power dimension and the coverage dimension is calculated as a distance metric. If the distance exceeds a preset threshold, it is marked as a potential breakpoint. The slope of the change of each preset number of data points before and after the breakpoint is extracted. Transitional data is generated at the breakpoint using cubic spline interpolation to obtain a continuous data sequence. The continuous data sequence is smoothed using Kalman filtering. The state transition matrix is ​​set according to the first-order difference relationship of power changes, and the observation matrix is ​​constructed based on the linear relationship between snow cover and power. Through recursive calculations in two stages—prediction and update—the comprehensive enhanced data sequence is obtained. The variance of each data point in the comprehensive enhanced data sequence within a preset window is calculated as the local variance, and the variance of the entire sequence is calculated as the global variance. The ratio of the local variance to the global variance is used as the quality score. If the score is lower than a preset quality threshold, the data point is marked as an anomaly. High-quality data points are retained and anomalies are removed to obtain the optimized and enhanced dataset.

[0071] It should be noted that the determination of the fusion weights is based on the correlation strength between each statistical feature and the power output. In the actual implementation, the Pearson correlation coefficient between each statistical feature and the power change in historical data is calculated, and the absolute value of the correlation coefficient is normalized so that the sum of all weights is 1.

[0072] For example, statistics related to snow cover typically have high weights because they have a direct causal relationship with power decline; while meteorological parameters such as temperature and humidity have relatively low weights, but still play a supporting role under certain conditions. In one implementation, the weighted average fusion process involves using the original monitoring value at each time point as a baseline, multiplying the corresponding statistical feature values ​​at that time point by their respective weights, and then summing these values ​​to the baseline value. This fusion method preserves the authenticity of the original data while enhancing its statistical characteristics. The resulting multidimensional data structure includes multiple dimensions such as timestamps, original power values, weighted statistical feature values, and snow cover, each providing information support from different perspectives for subsequent processing.

[0073] Specifically, the distance metric used in continuity detection employs a modified form of Euclidean distance. Considering the dimensional differences between power and coverage, the data for both dimensions are first standardized, mapping them to the same numerical range. Then, the sum of the squared differences in the standardized power and coverage dimensions at adjacent time points is calculated, and the square root is taken to obtain the overall distance. When this distance suddenly increases, exceeding three standard deviations of the historical distance mean, it is identified as a potential breakpoint.

[0074] Preferably, the application of cubic spline interpolation ensures a smooth data transition. At the breakpoint, five data points before and after the breakpoint are extracted as control points, and a cubic polynomial function is constructed. This function satisfies the constraint that the function value at each control point is consistent with the original data, and that the first and second derivatives are continuous. By solving the system of linear equations formed by these constraints, the polynomial coefficients are obtained, and then the interpolated data at the breakpoint is calculated. The transition data generated by this method conforms to the local variation trend while maintaining overall continuity.

[0075] For example, the state transition matrix of the Kalman filter is constructed based on the physical laws governing power change. During blizzards, power change is primarily influenced by the rate of snow accumulation, exhibiting an approximately linear decreasing trend. The state vector contains two components: the power value and the rate of power change. The state transition matrix describes the state evolution from the current moment to the next. The diagonal elements of the matrix are close to 1, indicating the persistence of the state; the off-diagonal elements reflect the influence of the rate of change on the power value. The observation matrix uses snow cover as the observation, converting the observations into state estimates through a pre-established cover-power mapping relationship. The filtering process includes two stages: prediction and update. In the prediction stage, the state estimate for the next moment is calculated based on the state transition matrix; in the update stage, the predicted value is adjusted using the Kalman gain, combined with the actual observations, to obtain the final filtered output. This recursive calculation method effectively suppresses measurement noise while preserving the true power change trend.

[0076] In one possible implementation, local variance is calculated using a sliding window method, with the window size set to include 10 data points before and after the data point. The variance within the window is calculated and used as the local variance value at the center point. The global variance is the population variance calculated over the entire data sequence. The ratio of the two reflects the degree of local data fluctuation relative to the overall anomaly.

[0077] It should be noted that the quality score threshold is set based on statistical analysis principles. The distribution of quality scores is calculated using historical normal data, and the 95th percentile is used as the threshold. When the quality score of a data point falls below this threshold, it indicates that the local fluctuation at that point is abnormally drastic, possibly due to sensor malfunction or other interference factors causing the outlier.

[0078] Specifically, outlier removal employs a mark-and-replace strategy. Data points marked as outliers are not simply deleted, but replaced with linear interpolations of the preceding and following normal data points. This approach removes outlier interference while maintaining the integrity of the time series. The resulting optimized and augmented dataset preserves the true power variation characteristics while eliminating the influence of noise and outliers.

[0079] S6. The optimized and augmented dataset is validated based on the scarce region identifier, and the similarity between the comprehensive augmented data sequence and the power decline segment is analyzed to determine the validated augmented data version: Data segments corresponding to the specified time period are extracted from the optimized and augmented dataset based on the scarce region identifier. Simultaneously, power decline segments of the same time period are obtained from the original data as reference benchmarks. After aligning the two data segments by time, the sum of squared differences in power values ​​at each time point is calculated to obtain the data segment to be validated. The dynamic time warping algorithm is used to calculate the similarity between the data segment to be validated and the power decline segment. By constructing a distance matrix between time series, dynamic programming is used to find the alignment path with the smallest cumulative distance. The cumulative distance on the path is normalized and used as the similarity value. If the similarity value is lower than a preset similarity threshold, the data segment is marked as a region requiring optimization. For the region requiring optimization, the parameters of the temporal generative adversarial network are adjusted. The generator's learning rate is multiplied by a preset factor to increase its adaptation speed, and the discriminator's discrimination threshold is reduced by a preset ratio to enhance its discrimination strictness. The generator weights are updated using gradient descent. After each iteration, the data for the region requiring optimization is regenerated, and a new similarity value is calculated. When the new similarity value reaches the preset similarity threshold or the number of iterations reaches the upper limit, the update stops. The current generator weight and discriminator threshold are saved as optimization parameters. The data of all scarce regions are regenerated using the optimization parameters and merged with the original data to obtain a verified enhanced data version.

[0080] Specifically, scarce region identifiers play a positioning role in the verification process. By parsing the timestamp information in the identifiers, the specific time period to be verified in the optimized and enhanced dataset can be accurately located. In practice, the data segments extracted from the optimized and enhanced dataset contain power value sequences that have undergone fusion and optimization, while the power decrease segments obtained from the original data retain the actual power change characteristics under blizzard conditions. Time alignment between the two is achieved through timestamp matching, ensuring that each data point at any given moment corresponds one-to-one. In one implementation, the calculation of the sum of squared differences involves point-by-point comparison. For the two aligned sequences, the difference between the enhanced data power value and the original power value is calculated at each moment, and the squared differences are accumulated. This accumulated value reflects the degree of deviation between the two sequences as a whole. When the deviation is large, it indicates that there is a significant difference between the enhanced data and the real data, requiring further optimization.

[0081] Specifically, the Dynamic Time Warping algorithm is the core technology of the verification process. This algorithm measures the similarity between two time series by constructing a distance matrix. Each element of the distance matrix represents the distance between a certain time in one series and a certain time in another series. The key to the algorithm is to find an optimal path from the top left corner to the bottom right corner of the matrix, which represents the best alignment of the two series. The path selection follows three constraints: a boundary condition ensures that the path starts at the top left corner and ends at the bottom right corner; a continuity condition ensures that the points on the path are continuous in time; and a monotonicity condition ensures that the path does not backtrack in time. Using dynamic programming, the algorithm calculates the cumulative distance at each position, which equals the distance at the current position plus the minimum cumulative distance from the previous position. Finally, the cumulative distance on the path is normalized to the path length to obtain a similarity value between 0 and 1, with the value closer to 1 indicating greater similarity between the two series.

[0082] Preferably, when the similarity value is lower than a preset threshold, the system identifies a data quality problem and initiates a parameter optimization process. The preset threshold is typically set to 0.85, a value determined based on extensive historical validation experience, ensuring data quality without being overly stringent and causing frequent retraining. Data segments marked as requiring optimization have their specific time range and similarity deviation recorded; this information guides subsequent parameter adjustments.

[0083] For example, the parameter tuning of a temporal generative adversarial network employs an adaptive strategy. The generator's learning rate is dynamically increased based on the degree of similarity deviation; the greater the deviation, the higher the learning rate increase, typically between 1.5 and 3 times. This adjustment allows the generator to quickly adapt to the characteristics of the target data distribution. The discriminator threshold is also reduced by a percentage related to the degree of deviation, ranging from 20% to 50%, with stricter discrimination criteria forcing the generator to produce data closer to the true distribution.

[0084] In one possible implementation, the gradient descent method updates the generator weights using a mini-batch gradient descent strategy. Each iteration randomly samples a batch of data from the region to be optimized, calculates the gradient of the loss function between the generated data and the real data, and updates the weight matrix based on the gradient direction and learning rate. The loss function is a combination of mean squared error and adversarial loss, considering both the numerical similarity of the data and the similarity of their distributions. After each weight update, the data for that region is immediately regenerated using the new weights, and a new similarity value is calculated, forming a closed-loop optimization process.

[0085] It's important to note that the iteration stopping condition balances optimization effectiveness and computational efficiency. Reaching the similarity threshold is an ideal stopping condition, indicating that the generated data is sufficiently close to the real data. The maximum number of iterations is typically set to 100 to prevent infinite loops in situations where optimization is difficult. When any stopping condition is met, the system saves the current network parameter state.

[0086] Specifically, the application of optimization parameters ensures that all scarce regions receive high-quality data supplementation. The saved generator weights and discriminator thresholds are used as a set of optimization parameters to regenerate data for all scarce regions. This batch processing method guarantees data consistency and avoids data distribution discrepancies caused by using different parameters for different regions.

[0087] For example, during a certain validation process, the system detected similarities of 0.72, 0.78, and 0.81 for three scarce regions, all below the threshold of 0.85. After parameter optimization and 50 iterations, the similarities of the three regions improved to 0.88, 0.91, and 0.89, all meeting the quality requirements. The augmented data version not only filled in the missing data but also maintained a distribution characteristic highly consistent with the real data.

[0088] S7. Extract feature distribution from the validated enhanced data version, and adjust the feature distribution to match the power decline segment in the preliminary monitoring dataset to obtain complete enhanced data reflecting extreme weather power changes: From the validated enhanced data version, use Fast Fourier Transform to convert the power time-series data to the frequency domain, extract the frequency component with the largest amplitude in the spectrum as the dominant frequency, and record its frequency and amplitude. Simultaneously, calculate the change in snow cover rate between adjacent time points using difference operations, and take the maximum value as the change amplitude. Combine the dominant frequency, dominant frequency amplitude, and snow cover rate change amplitude to form a feature vector. Extract feature vectors from the power decline segment of the preliminary monitoring dataset using the same method as the target distribution, calculate the Euclidean distance between the feature vector and the target distribution in the corresponding dimension, and use the least squares method to solve the linear transformation matrix to minimize the distance between the transformed feature vector and the target distribution, thus obtaining the feature adjustment parameters. The enhanced data is reconstructed based on the feature adjustment parameters. The adjusted frequency and amplitude are converted back to the time domain through inverse Fourier transform to obtain a new power sequence. The snow cover value at each time point is recalculated according to the adjusted change amplitude. The reconstructed power sequence and the snow cover sequence are aligned and merged according to the timestamp to obtain complete enhanced data reflecting the power changes in extreme weather.

[0089] Specifically, in one implementation, the Fast Fourier Transform (FFT) is used to transform power time-series data from the time domain to the frequency domain, enabling spectral analysis of power fluctuation characteristics. The process involves performing a Discrete Fourier Transform on the power sequence to obtain complex spectral data, calculating the amplitude spectrum of each frequency component (where the amplitude equals the modulus of the complex number), and searching for the frequency point with the largest amplitude in the spectrum. This frequency is the dominant frequency, reflecting the main periodic characteristics of the power fluctuation.

[0090] For example, when the main frequency is 0.2 Hz, it indicates that the main period of power fluctuation is about 5 minutes, which is consistent with the intermittent characteristics of snow accumulation during blizzard weather.

[0091] It should be noted that the calculation of the snow cover change rate uses the first-order difference method. The difference between the values ​​of two adjacent moments in the snow cover time series is taken to obtain the instantaneous rate of change sequence of snow cover. The entire rate of change sequence is then iterated through to find the rate of change with the largest absolute value as the magnitude of the change. This feature reflects the severity of snow cover during blizzards.

[0092] Preferably, the feature vector construction combines frequency domain features and time domain features. The dominant frequency value, dominant frequency amplitude, and maximum coverage variation range constitute a three-dimensional feature vector, and each dimension is normalized to ensure the comparability of features with different dimensions. The feature vector of the target distribution is extracted from the power descent segment using the same method and serves as a reference benchmark for feature adjustment.

[0093] Specifically, the process of solving the linear transformation matrix using the least squares method involves constructing an overdetermined system of equations. Using the source eigenvectors as input and the target eigenvectors as the desired output, the elements of the transformation matrix are determined by minimizing the sum of squared prediction errors. The transformation matrix is ​​a 3×3 square matrix containing parameters for scaling and linear combination.

[0094] In one possible implementation, the inverse Fourier transform reconstruction process maintains signal integrity. Based on the adjusted dominant frequency and amplitude, the complex values ​​of the corresponding frequency components in the spectrum are modified, while other frequency components remain unchanged. The modified spectrum is then converted back to the time domain using an inverse discrete Fourier transform, yielding the adjusted power sequence. This sequence retains the basic trend of the original data while incorporating typical fluctuation characteristics under extreme weather conditions. The final step of data merging ensures time consistency. Both the reconstructed power sequence and the coverage sequence calculated based on the new amplitude of change retain their original timestamps. The two sequences are aligned and combined using timestamp indexing to form complete enhanced data containing both power and coverage dimensions, accurately reflecting the power variation patterns of renewable energy power plants under blizzard conditions.

[0095] By collecting time-series data on snow cover and power output sequences, this invention identifies periods of drastic changes in snow cover and segments of power decline, constructing a preliminary monitoring dataset and locating data-scarce areas based on a rate-of-change threshold. A temporal generative adversarial network (GAN) is used to generate supplementary power data. After expanding the dataset, statistical features are extracted, uneven distribution is addressed, and a comprehensive enhanced data sequence is generated. Similarity verification and iterative optimization of network parameters ensure data continuity and accuracy. Finally, the feature distribution is adjusted to match the actual power decline segments, forming complete enhanced data. The core innovation of this invention lies in accurately filling data gaps through the combination of adversarial networks and feature adjustment, improving power prediction accuracy under extreme weather conditions, and providing reliable support for the optimization of new energy power plant operations.

[0096] In a specific embodiment, please refer to Figure 2 This is a schematic diagram of a new energy power data enhancement system according to a second embodiment of this application. The system includes: a data acquisition module 201, a target sequence extraction module 202, a scarce region extraction module 203, a sequence generation module 204, and an enhancement module 205. The data acquisition module 201 is used to simultaneously acquire the snow cover rate time series sequence and the first power output sequence of the new energy power station during blizzard weather. The target sequence extraction module 202 is used to obtain the target snow cover rate time series sequence in the snow cover rate time series sequence, where the snow cover change rate exceeds a first preset threshold. The scarce region extraction module 203 is used to determine the first power output sequence based on the target snow cover rate time series sequence. The power data is scarce in the region; the power data scarce region is a time range in which the number of power output data collection points within a preset second unit time is less than a second preset threshold; the sequence generation module 204 is used to acquire the sparse start and end time periods, snow cover rate change trend and first meteorological data of the power data scarce region, and input the sparse start and end time periods, snow cover rate change trend and first meteorological data into a preset temporal generative adversarial network to generate a second power output sequence of the power data scarce region; the meteorological data includes at least temperature and humidity; the enhancement module 205 is used to enhance the first power output sequence using the second power output sequence to obtain an enhanced third power output sequence.

[0097] In a specific embodiment, the third embodiment of this application provides a new energy power data enhancement device, including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method as described in any one of the first embodiments of this application.

[0098] In a specific embodiment, the fourth embodiment of this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method as described in any one of the first embodiments of this application.

[0099] The above embodiments merely illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for enhancing new energy power data, characterized in that, The method includes: Simultaneously collect the time series of snow cover rate and the first power output sequence of the new energy power station during blizzard weather; In the snow cover time series, obtain the target snow cover time series in which the snow cover change rate exceeds the first preset threshold within a first preset unit time. The power data sparse region in the first power output sequence is determined based on the target snow coverage time series; the power data sparse region is the time range in which the number of power output data collection points is less than a second preset threshold within a preset second unit time. The sparse start and end time periods, snow cover change trends, and first meteorological data of the power data-scarce region are obtained, and the sparse start and end time periods, snow cover change trends, and first meteorological data are input into a preset temporal generative adversarial network to generate a second power output sequence of the power data-scarce region; the meteorological data includes at least temperature and humidity. The first power output sequence is enhanced using the second power output sequence to obtain an enhanced third power output sequence.

2. The new energy power data enhancement method as described in claim 1, characterized in that, The step of enhancing the first power output sequence using the second power output sequence to obtain an enhanced third power output sequence includes: The second power output sequence is inserted into the sparse start and end time period of the first power output sequence to obtain the third power output sequence.

3. The new energy power data enhancement method as described in claim 1, characterized in that, The step of determining the power data sparse region in the first power output sequence based on the target snow cover time series includes: Determine the duration corresponding to the time series of the target snow cover rate; In the first power output sequence, obtain the power data sparse region within the duration.

4. The new energy power data enhancement method as described in claim 1, characterized in that, The step of inputting the sparse start and end time periods, the snow cover rate change trend, and the first meteorological data into a preset time-series generative adversarial network to generate a second power output sequence for the power data-sparse region includes: In the first power output sequence, the power output range corresponding to the target time period is extracted as a boundary condition; the target time period is the time period before and after the sparse start and end time periods after each extended preset length. Input data is constructed based on the boundary conditions, the snow cover rate change trend, and the first meteorological data. The input data is concatenated with a preset random noise vector and then input into the preset temporal generative adversarial network to generate a second power output sequence for the power data-scarce region.

5. The new energy power data enhancement method as described in claim 1, characterized in that, After obtaining the enhanced third power output sequence, the method further includes: Acquire the second meteorological data of the new energy power station during the blizzard period; The snow cover time series, the third power output series, the second meteorological data, and the preset weighting value corresponding to each second meteorological data are fused to obtain a multi-dimensional enhanced output series; Extract the fourth power output sequence corresponding to the sparse start and end time period from the enhanced output sequence, and extract the fifth power output sequence corresponding to the sparse start and end time period from the first output sequence; Calculate the similarity between the fourth power output sequence and the fifth power output sequence; If the similarity is less than the third preset threshold, the fourth power output sequence is taken as the power data scarce region. The steps of obtaining the sparse start and end time period, snow cover change trend and first meteorological data of the power data scarce region are returned, and the sparse start and end time period, snow cover change trend and first meteorological data are input into the preset temporal generative adversarial network to generate the second power output sequence of the power data scarce region. This process continues until the similarity is not less than the third preset threshold. The last obtained enhanced third power output sequence is the fourth power output sequence.

6. The new energy power data enhancement method as described in claim 5, characterized in that, After obtaining the fourth power output sequence, the method further includes: Obtain the first dominant frequency and the first dominant frequency amplitude of the fourth power output sequence; the first dominant frequency is the frequency with the largest amplitude in the fourth power output sequence; the first dominant frequency and the first dominant frequency amplitude constitute a first feature vector; Obtain the second dominant frequency and the second dominant frequency amplitude of the first power output sequence; the second dominant frequency is the frequency with the largest amplitude in the first power output sequence; the second dominant frequency and the second dominant frequency amplitude constitute a second feature vector; Calculate the Euclidean distance between the first feature vector and the second feature vector; With the goal of minimizing the Euclidean distance, the linear transformation matrix formed by the first eigenvector and the second eigenvector is solved using the least squares method to obtain the feature adjustment parameters of the fourth power output sequence; the feature adjustment parameters are the main frequency adjustment parameters and the main frequency amplitude adjustment parameters. The fourth power output sequence is optimized using the aforementioned feature adjustment parameters to obtain the fifth power output sequence.

7. The new energy power data enhancement method as described in claim 6, characterized in that, The method further includes: Obtain the power decrease output sequence from the first power output sequence; The step of obtaining the second main frequency and the second main frequency amplitude of the first power output sequence includes: Obtain the second main frequency and the second main frequency amplitude of the power decrease output sequence; the second main frequency is the frequency with the largest amplitude in the power decrease output sequence.

8. A new energy power data enhancement system, characterized in that, The system includes: a data acquisition module, a target sequence extraction module, a scarce region extraction module, a sequence generation module, and an enhancement module; The data acquisition module is used to simultaneously acquire the time series sequence of snow cover rate and the first power output sequence of the new energy power station during blizzard weather; The target sequence extraction module is used to obtain a target snow coverage time series sequence in the snow coverage time series where the snow coverage change rate exceeds a first preset threshold within a first preset unit time. The scarce region extraction module is used to determine the scarce power data region in the first power output sequence based on the target snow coverage time series; the scarce power data region is the time range in which the number of power output data collection points is less than a second preset threshold within a preset second unit time period. The sequence generation module is used to acquire the sparse start and end time periods, snow cover change trends, and first meteorological data of the power data-scarce region, and input the sparse start and end time periods, snow cover change trends, and first meteorological data into a preset temporal generative adversarial network to generate a second power output sequence of the power data-scarce region; the meteorological data includes at least temperature and humidity. The enhancement module is used to enhance the first power output sequence using the second power output sequence to obtain an enhanced third power output sequence.

9. A new energy power data enhancement device, comprising a memory and a processor, characterized in that, The memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method as described in any one of claims 1 to 7.

Images