Weather classification method, system and equipment in new energy high-influence process and medium

Abstract

The invention discloses a new energy high-influence process weather classification method, system and device and a medium, and relates to the technical field of wind power. The method comprises the following steps: screening sensitive weather factors of a target electric field, and obtaining historical data of the target electric field; the historical data comprises historical power data and NWP data; defining a new energy high-influence weather process based on the sensitive weather factors and the historical data, determining a meteorological factor threshold of the new energy high-influence weather process, and performing data generation by using a WGAN generative adversarial network to obtain a sample set; constructing a step-by-step hierarchical extraction classification model, and performing iterative training by using the sample set; and the trained step-by-step layered extraction classification model is used to identify weather at a to-be-detected time point, and a final weather classification result is obtained. According to the method, the weather classification accuracy is improved by considering the influence of the extreme weather on the new energy output and the internal relation of various meteorological factors among different extreme weathers.

Description

Technical Field

[0001] The present invention relates to the field of wind power technology, and in particular to a new energy high-impact process weather classification method, system, equipment and medium. Background Art

[0002] With the increasing prominence of renewable energy power generation, China has made significant progress in installed capacity, power generation, and energy consumption. Meanwhile, global average temperatures have continued to rise in recent years, and extreme weather events have become frequent. Power systems are vulnerable to weather conditions, exposing them to potential threats and risks to their operation and stability when faced with various weather conditions, especially extreme weather events. However, current classification of extreme weather events lacks a fully unified standard, and its impact on renewable energy output is often not considered, resulting in a lack of targeted analysis and classification. To minimize or avoid the significant impact or even damage of extreme weather on new power systems, it is crucial to define, classify, and identify high-impact extreme weather standards.

[0003] At present, the implementation of extreme weather mainly focuses on the impact on the overall new power system. Renewable energy, such as wind energy and solar energy, as an important part of building a new power system, also has a significant impact on the weather vulnerability of the power system. Therefore, it is crucial to explore the impact of extreme weather on renewable energy. However, the methods used in the current implementation mostly model different extreme weather scenarios separately and then output classification prediction results separately, which lacks consideration of the intrinsic connection between various meteorological factors in different extreme weather. At the same time, the use of multiple independent classification models for separate output will also have problems such as excessive model training time cost and low computational efficiency.

[0004] While multi-task learning (MTL) can effectively mitigate these issues, the MTL models currently used in various implementations suffer from inherent negative transfer and seesaw problems. This occurs when the correlation between tasks is weak, resulting in poor joint training results or the performance of one task being sacrificed to improve the performance of another. Furthermore, the scarcity of extreme weather data samples poses significant challenges to the modeling and training of machine learning models. Therefore, a new method for weather classification and prediction is needed. Summary of the Invention

[0005] The purpose of the present invention is to provide a new energy high-impact process weather classification method, system, equipment and medium, aiming to solve or improve at least one of the above-mentioned technical problems.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A classification method for new energy high-impact process weather, including:

[0008] Screening sensitive weather factors of the target electric field and obtaining historical data of the target electric field; the historical data includes historical power data and NWP data; the target electric field includes a wind farm and a photovoltaic farm;

[0009] Based on the sensitive weather factors and the historical data, define the weather process with high impact on new energy, determine the meteorological factor threshold of the weather process with high impact on new energy, and use the WGAN generative adversarial network to generate data to obtain a sample set;

[0010] Constructing a hierarchical extraction and classification model, and performing iterative training using the sample set; the hierarchical extraction and classification model adopts a PLE network, which is composed of a specific expert network and a shared expert network;

[0011] The trained hierarchical extraction and classification model is used to identify the weather at the time point to be tested, and the final weather classification result is obtained.

[0012] Optionally, the sensitive weather factors of the target electric field specifically include: the sensitive weather factors of the wind field include wind speed and temperature; the sensitive weather factors of the photovoltaic field include solar radiation intensity.

[0013] Optionally, the step of defining a weather process with a high impact on new energy based on the sensitive weather factors and the historical data, determining a threshold of meteorological factors for the weather process with a high impact on new energy, and generating data using a WGAN generative adversarial network to obtain a sample set specifically includes:

[0014] The abnormal weather time points of the historical data are defined as weather processes with high impact on new energy, and an initial data set is constructed based on the sensitive weather factor data corresponding to the weather processes with high impact on new energy; the abnormal weather time points are weather time points with abnormal power forecast results and abnormal output conditions;

[0015] The initial data set is expanded using a WGAN generative adversarial network to obtain a sample set; the WGAN generative adversarial network restricts the strategy set of the discriminator to Lipschitz continuity on the basis of a basic generative adversarial network.

[0016] Optionally, defining the abnormal weather time points of the historical data as new energy high-impact weather processes, and constructing an initial data set based on the sensitive weather factor data corresponding to the new energy high-impact weather processes, specifically includes:

[0017] Based on the historical data, determining a time point when the power prediction error is greater than a first set threshold as a weather time point when an abnormal power prediction result occurs, and determining a time point when the actual power and the predicted power drop by more than a second set threshold as a weather time point when an abnormal output condition occurs;

[0018] The average value and standard deviation of the sensitive weather factor data corresponding to the abnormal weather time point are calculated, and the data points that deviate from the average value by 1-2 standard deviations are taken as the meteorological factor threshold for the new energy high-impact weather process;

[0019] The time points corresponding to meteorological data above the meteorological factor threshold are labeled "1", and the time points corresponding to meteorological data below or equal to the meteorological factor threshold are labeled "0", and the initial data set is generated based on historical data and corresponding data labels.

[0020] Optionally, the process of constructing the hierarchical extraction classification model includes:

[0021] Build a pre-trained network based on the PLE network;

[0022] The sample set is input into the pre-trained network, and the cross entropy loss is selected as the loss function of the model training. When the pre-trained network meets the set number of iterations or the set evaluation index, it is determined to be a trained hierarchical extraction and classification model.

[0023] The present invention also provides a new energy high-impact process weather classification system, comprising:

[0024] A data acquisition unit, configured to screen sensitive weather factors of a target electric field and obtain historical data of the target electric field; the historical data includes historical power data and NWP data; the target electric field includes a wind farm and a photovoltaic farm;

[0025] a sample generation unit, configured to define a weather process with a high impact on new energy based on the sensitive weather factors and the historical data, determine a threshold value of meteorological factors for the weather process with a high impact on new energy, and generate data using a WGAN generative adversarial network to obtain a sample set;

[0026] A model training unit is used to construct a hierarchical extraction and classification model and perform iterative training using the sample set; the hierarchical extraction and classification model adopts a PLE network, which is composed of a specific expert network and a shared expert network;

[0027] The weather classification unit is used to identify the weather at the time point to be tested using the trained hierarchical extraction classification model to obtain the final weather classification result.

[0028] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned new energy high-impact process weather classification method.

[0029] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned new energy high-impact process weather classification method.

[0030] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0031] The present invention discloses a method, system, device, and medium for classifying weather processes with high impact from new energy sources. The method comprises screening sensitive weather factors of a target electric field and obtaining historical data of the target electric field; the historical data comprises historical power data and NWP data; defining weather processes with high impact from new energy sources based on the sensitive weather factors and the historical data, determining meteorological factor thresholds for weather processes with high impact from new energy sources, and generating data using a WGAN generative adversarial network to obtain a sample set; constructing a hierarchical extraction classification model and iteratively training it using the sample set; and using the trained hierarchical extraction classification model to identify the weather at the time point to be tested to obtain a final weather classification result. The present invention improves weather classification accuracy by considering the impact of extreme weather on new energy output and the intrinsic connection between various meteorological factors between different extreme weather conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 Schematic diagram of the overall process framework in this embodiment;

[0034] Figure 2 is a structural diagram of the PLE in this embodiment;

[0035] FIG3( a ) is a normal distribution diagram of temperature data of a wind farm at a high error point in this embodiment;

[0036] FIG3( b ) is a normal distribution diagram of wind speed data at a high error point in the wind farm in this embodiment;

[0037] FIG4( a ) is a normal distribution diagram of temperature data at the time of abnormal output of the photovoltaic field in this embodiment;

[0038] FIG4( b ) is a normal distribution diagram of the total precipitation data at the time of abnormal photovoltaic power output in this embodiment;

[0039] FIG5( a ) is a schematic diagram of the classification results of the wind farm PLE model in this embodiment;

[0040] FIG5( b ) is a schematic diagram of the classification results of the photovoltaic electric field PLE model in this embodiment;

[0041] FIG6( a ) is a schematic diagram showing comparison results of the wind farm weather classification model in this embodiment;

[0042] FIG6( b ) is a schematic diagram showing the comparison results of the photovoltaic field weather classification model in this embodiment. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] The purpose of the present invention is to provide a new energy high-impact process weather classification method, system, equipment and medium, aiming to solve or improve at least one of the above-mentioned technical problems.

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1-Figure 2 As shown, the present invention provides a new energy high-impact process weather classification method, comprising:

[0047] Step 1: Screen the sensitive weather factors of the target electric field and obtain historical data of the target electric field; the historical data includes historical power data and NWP data; the target electric field includes a wind farm and a photovoltaic farm.

[0048] Step 2: Based on the sensitive weather factors and the historical data, define the weather process with high impact on new energy, determine the meteorological factor threshold of the weather process with high impact on new energy, and use the WGAN generative adversarial network to generate data to obtain a sample set.

[0049] Step 3: Construct a hierarchical extraction and classification model, and perform iterative training using the sample set; the hierarchical extraction and classification model adopts a PLE network, which is composed of a specific expert network and a shared expert network.

[0050] Step 4: Use the trained hierarchical extraction and classification model to identify the weather at the time point to be tested and obtain the final weather classification result.

[0051] As a specific implementation method, the processing process of each of the above steps is described in detail.

[0052] 1) Analysis of sensitive weather factors

[0053] For wind farms, the weather factors that can seriously affect the output value are mainly wind speed and temperature. Wind speed is the most critical factor affecting the power output of wind farms. The power output of wind turbines is proportional to the cube of wind speed:

[0054]

[0055] Where P is power, ρ is air density, A is the swept area of ​​the wind wheel, v is wind speed, C p is the power coefficient of the wind turbine (usually between 0.35 and 0.45).

[0056] Temperature primarily affects wind turbine power output by affecting air density (ρ). Air density is inversely proportional to temperature: higher temperatures lead to lower air density and lower power output.

[0057]

[0058] Among them, P atm is the atmospheric pressure, R is the gas constant, and T is the absolute temperature.

[0059] For photovoltaic electric fields, the most critical influencing factor is the intensity of solar radiation. That is, the higher the intensity of solar radiation per unit time, the greater the output power of the photovoltaic panels.

[0060] P 光伏 =I×A 光伏 ×η (3)

[0061] Where I is the solar radiation intensity, A is the area of ​​the photovoltaic panel, and η is the efficiency of the photovoltaic panel.

[0062] Under clear weather conditions, the intensity of solar radiation is high, and the power generation capacity of photovoltaic power plants will also increase accordingly. Therefore, the output of photovoltaic power plants will also be affected by weather conditions that can directly affect the intensity of solar radiation, such as rain and snowfall (collectively referred to as precipitation).

[0063] In addition, drastic temperature changes will also affect the output of the photovoltaic field by affecting the photovoltaic field efficiency η. In practical applications, the effects of dust and pollution, as well as system losses, may also be considered. The actual photovoltaic power can be calculated using this correction formula:

[0064] P true =P×K T ×K dirt ×K loss (4)

[0065] Among them, P trueis the actual photovoltaic power, K T is the temperature correction factor, K dirt is the dust pollution correction factor, K loss is the system loss coefficient.

[0066] 2) Definition of high-impact weather processes involving new energy

[0067] The definition method of high-impact weather processes for new energy can be considered from two aspects: one is the abnormal power prediction results of new energy electric fields, and the other is the abnormal output of new energy electric fields.

[0068] Without considering the impact of the power prediction model's inherent flaws, abnormal prediction results are often caused by external meteorological factors, such as ice covering wind turbine blades in wind farms during winter cold snaps. This often causes power prediction results to deviate significantly from actual power, resulting in large prediction errors. Therefore, we can extract time points in the dataset where power prediction errors are large. Assuming that the sensitive weather factor data corresponding to these time points follow a normal distribution, we calculate their mean and standard deviation. We then take data points that deviate from the mean by 1-2 standard deviations and define them as the meteorological factor threshold for a high-impact weather process involving renewable energy. Time points corresponding to meteorological data above this threshold are defined as the time points when a high-impact weather process involving renewable energy occurs, and are labeled "1" accordingly; otherwise, they are labeled "0."

[0069] d avg -(1~2)d sd ≤d≤d avg +(1~2)d sd (5)

[0070] Where d is the data value, d avg is the mean value of the data, d sd is the standard deviation of the data.

[0071] However, because the impact of certain high-impact weather processes is specifically reflected in NWP data, in this case, although the power forecast error does not change much, both actual and predicted power experience a significant decline. For example, for photovoltaic power plants, heavy rainfall events are often accompanied by low irradiance, causing power forecasts, which primarily use irradiance as model input, to also decline. Coupled with the significant daily periodicity of photovoltaic power plants, the changes reflected in the power forecast error are very subtle. Therefore, another definition method directly considers the situation where new energy power plants experience abnormal output, including these time points in the potential for high-impact weather processes. Assuming that the sensitive weather factor data corresponding to these time points satisfy a normal distribution, their mean and standard deviation are calculated. Data points that deviate from the mean by 1-2 standard deviations are then defined as the meteorological factor threshold for high-impact weather processes involving new energy. Time points corresponding to meteorological data above this threshold are defined as the time points when high-impact weather processes involving new energy occur, and are labeled "1" accordingly; otherwise, they are labeled "0."

[0072] 3) WGAN data generation

[0073] In machine learning, data generation is an important technology, especially when real data is scarce or privacy is restricted. Generative adversarial networks (GANs) learn complex data distributions to generate high-quality synthetic data by constructing an adversarial network consisting of a generator G and a discriminator D to play against each other. On this basis, WGAN restricts the discriminator's strategy set to Lipschitz continuity. The Lipschitz continuous function limits the speed at which the function changes. The slope of the function that meets the Lipschitz condition must be less than a real number called the Lipschitz constant. This limits the slope change of the discriminator, inhibits the learning speed of the discriminator, and improves the problem of the generator's gradient disappearance to a certain extent. Its valuation function is as follows. Given the generator G, the optimization goal of the discriminator is to maximize the valuation function V WGAN (G,D).

[0074]

[0075] Among them, G is the generator, D is the discriminator, E is the expected probability, z is the input, x is G(z), z~p z (z) is the noise distribution of random initialization, x~p data (x) is the data distribution to be learned.

[0076] By taking the data corresponding to the sampling points labeled "1" in step 2) as input, we can obtain expanded data with similar features.

[0077] 4) Construct a hierarchical extraction and classification model

[0078] PLE is a type of multi-task learning (MTL), which allows multiple tasks to be learned in parallel at the same time, and the results influence each other. Custom gate control (CGC) is the basic network of PLE. It divides the expert network into two types based on MTL. One is a specific expert network related to a specific task, and the other is a shared expert network shared by all tasks. The specific expert network specific to each task only accepts the tower gradient update parameters of the corresponding task, while the shared expert network is updated with multi-task results. This allows different types of expert networks to focus on learning different knowledge more efficiently and avoid unnecessary interactions, reducing the interference of parameter sharing between expert networks, and effectively alleviating negative transfer and seesaw phenomena. In addition, thanks to the dynamic fusion of inputs by the gating network, CGC can more flexibly find a balance between different subtasks and better handle conflicts between tasks and sample correlation problems. The gating network output g of the kth task k (x) is:

[0079] g k (x) = w k (x)S k (x) (7)

[0080]

[0081] Among them, w k (x) is the softmax function, W g k is the weight matrix, S k (x) is a selection matrix that connects the specific expert network and the shared expert network for the kth task, E T (i,m) is a specific expert network, E T (s,m) It is a shared expert network.

[0082] The final output y of the kth subtask k (x) is:

[0083] y k (x) = t k [g k (x)] (10)

[0084] PLE can be considered as a connection of multiple layers of CGCs, each layer of CGC is called an extraction network, and the gating network is further divided into two types: task-specific gates and shared gates. When PLE is used as a classification task model, the output of the gating network in the jth extraction network of the kth classification subtask is:

[0085] g k,j (x) = w k,j [g k,j-1 (x)]Sk (x) (11)

[0086] After calculating all the gating networks and expert networks, the final output of the k-th classification subtask of PLE is:

[0087] y k (x) = t k [g k,N (x)] (12)

[0088] At this time, y k (x) represents the model's predicted probability for the kth category. However, since the output of the classification task is generally an integer label, it is usually necessary to add a softmax function to the output layer. This embodiment solves this problem by selecting cross entropy loss as the loss function of the model. The cross entropy loss function is a common method to measure the difference between the probability distribution of the model output and the probability distribution of the true label. Then the negative log-likelihood loss is calculated, and its formula is as follows:

[0089]

[0090] Among them, T i Represents the i-th element in the true label vector, y i k (x) represents the i-th element in the model output vector after being processed by the softmax function.

[0091] 5) Simulation calculation

[0092] Simulation input: NWP data (including wind speed, wind direction, temperature, irradiance, precipitation, etc.) of sufficient length for wind farm clusters and photovoltaic farm clusters; data sampling interval is 1 hour; according to steps 1) to 4), obtain extreme weather classification results for wind farm clusters and photovoltaic farm clusters;

[0093] 6) Classification evaluation index analysis

[0094] In order to quantitatively evaluate the classification effect of the proposed model, four commonly used classification evaluation indicators are selected. Accuracy: The proportion of samples correctly classified by the model to the total number of samples. Usually used to measure the overall accuracy of the model and is applicable to data sets with balanced categories. Precision: The proportion of samples predicted as positive examples by the classifier that are actually positive examples. Usually used to measure the accuracy of the model's predictions as positive examples. Recall: The proportion of samples that are correctly predicted as positive examples among all samples that are actually positive examples. Usually used to measure the model's ability to capture positive examples. F1 score (F1-Score): The harmonic mean of precision and recall, used to comprehensively evaluate the performance of the model. It is an indicator used to balance precision and recall, and is applicable to situations where both are important. The calculation formulas for the four evaluation indicators are as follows:

[0095]

[0096] Here, true positives (TP) represent the number of samples correctly predicted by the model as positive. True negatives (TN) represent the number of samples correctly predicted by the model as negative. False positives (FP) represent the number of samples incorrectly predicted by the model as positive from negative classes. This is also known as a "Type I error." If a model frequently predicts non-extreme weather as extreme weather, it may lead to unnecessary resource investment. False negatives (FN) represent the number of samples incorrectly predicted by the model as negative from positive classes. This is also known as a "Type II error." If a model frequently predicts extreme weather as non-extreme weather, non-extreme weather events may be overlooked, resulting in inability to respond in a timely manner, potentially leading to serious damage to power generation equipment. For renewable energy power plants, the consequences of Type II errors are often more severe than Type I errors. Therefore, for extreme weather identification and classification models for renewable energy power plants, recall is slightly more important than precision.

[0097] According to step 5), input the simulated input and its corresponding weather label, and compare the classification result label obtained by the model with the actual label according to the accuracy (14), precision (15), recall (16) and F1 score (17) in step 6) to evaluate the classification results.

[0098] Based on the above technical solution, the following embodiments are provided.

[0099] The actual data set used in this embodiment is the actual data set of a wind farm cluster and a photovoltaic farm cluster in Jiangxi Province, China. The time span is from 0:00 on January 2, 2021 to 0:00 on January 1, 2024, with a time sampling interval of 1 hour. There are 26,256 sets of data in total. There are 67 wind farms and 57 photovoltaic farms. The data set includes historical power data of wind farms and photovoltaic farms, as well as NWP data such as wind speed, wind direction, temperature, surface shortwave (solar) radiation, surface longwave (thermal) radiation, and dew point temperature.

[0100] Step 1: For wind farms: First, select the time points with the highest 15% error in the annual power forecast from 2021 to 2023. These time points are defined as those most significantly affected by weather. Assuming that their temperature and wind speed data follow a normal distribution, the normal distribution plots are shown in Figures 3(a) and 3(b). The mean temperature is 17.93°C with a standard deviation of 9.49°C; the mean wind speed is 4.31 m / s with a standard deviation of 1.31 m / s. To ensure a sufficient number of sampling points, for low-temperature time points, extreme low-temperature weather was defined as temperature data 1.5 standard deviations below the mean, with a total of 435 data points and a temperature threshold of 3.70°C. For high-temperature time points, extreme high-temperature weather was defined as data 1.5 standard deviations above the mean, with a total of 192 data points and a temperature threshold of 32.16°C. For high-wind speed time points, high-wind weather was defined as wind speed data 2 standard deviations above the mean, with a total of 129 data points and a wind speed threshold of 6.94 m / s. For photovoltaic power plants, due to their significant daily periodicity and seasonality, each month was analyzed separately. Days with extreme weather were labeled as extreme weather days. Furthermore, any time during extreme weather days when power was non-zero was recorded as an extreme weather point. This eliminated the influence of nighttime inactivity on labeling and subsequent classification model predictions. First, for each month from 2021 to 2023, we took the daily power peaks and calculated their averages. We then extracted dates where the daily power peaks were below this average. Finally, we extracted the time points within these dates where the power was not zero, defining these as the time points most significantly affected by weather. Assuming that the temperature and precipitation data at these time sampling points follow a normal distribution, the normal distribution diagrams are shown in Figures 4(a) and 4(b). The mean temperature is 21.31°C with a standard deviation of 8.25°C; the mean precipitation is 0.615 mm with a standard deviation of 0.581 mm. For low temperatures, data with temperatures 2 standard deviations below the mean were defined as extreme low temperature weather data. There were 1,375 data points in total, and the temperature threshold was 4.80°C. For high temperatures, due to the high average temperature and large standard deviation, to ensure sufficient number of sampling points, data with temperatures 1.5 standard deviations above the mean were defined as extreme high temperature weather data. There were 1,039 data points in total, and the temperature threshold was 33.69°C. For heavy precipitation, data with total precipitation 2 standard deviations above the mean were defined as heavy precipitation weather data. There were 409 data points in total, and the total precipitation threshold was 1.777 mm.

[0101] Step 2: Divide the dataset as shown in Table 1 and Table 2.

[0102] Table 1 Wind farm dataset division

[0103]

[0104]

[0105] Table 2 Division of photovoltaic electric field dataset

[0106]

[0107] It can be seen that the extreme weather data in the original data accounts for a very small proportion, which is extremely unfavorable for the effective training of the classification and recognition model. Therefore, it is necessary to generate extreme weather data to expand the training set of the model, as shown in Tables 3 and 4.

[0108] Table 3. Wind farm dataset division after data generation

[0109]

[0110] Table 4 Division of photovoltaic electric field dataset after data generation

[0111]

[0112]

[0113] Step 3: Based on the temperature thresholds for high and low temperatures for wind farm clusters and photovoltaic farm clusters obtained in step 1, the wind speed thresholds for strong winds, and the total precipitation thresholds for heavy precipitation, each time sampling point is marked with a corresponding weather label. A label of "1" indicates that the corresponding extreme weather occurred at the time sampling point, and a label of "0" if it did not occur. The NWP data and high-impact weather labels for 2021 and 2022 are used as training sets to train the PLE classification model, and the NWP data and high-impact weather labels for 2023 and the data generated by WGAN are used as test sets. The following methods are used for comparison:

[0114] Method I: No data generation;

[0115] Method II: Generate three types of extreme weather data separately;

[0116] Method III: Generate according to characteristic classification (temperature, wind speed, total precipitation);

[0117] Method IV: Generate three types of extreme weather data as a whole.

[0118] The classification results are shown in Tables 5 to 10.

[0119] Table 5 Classification results of high temperature weather in wind farms

[0120]

[0121] Table 6 Classification results of low temperature weather in wind farms

[0122]

[0123] Table 7 Classification results of strong wind weather in wind farms

[0124]

[0125] Table 8 Classification results of high temperature weather in photovoltaic field

[0126]

[0127] Table 9 Classification results of photovoltaic field low temperature weather

[0128]

[0129] Table 10 Classification results of heavy precipitation weather in photovoltaic field

[0130]

[0131] As shown in the table, the data generated using Method IV is of higher quality, resulting in a slight improvement in the model's classification accuracy. For wind farms, the accuracy of the three extreme weather classifications increased by 1.35%, 1.44%, and 2.28%, respectively. This also significantly improved recall, with increases of 21.24%, 11.82%, and 16.85%, respectively. For photovoltaic farms, the accuracy of the three extreme weather classifications increased by 1.18%, 0.45%, and 1.37%, respectively. This also significantly improved other evaluation metrics. Furthermore, better data quality also resulted in better model convergence. The resulting model classification results are shown in Figures 5(a) and 5(b).

[0132] Step 4: Then, a variety of other classification models are used for comparison. The classification models used include long short-term memory network (LSTM), gated recurrent unit (GRU), convolutional neural network (CNN) and support vector classifier (SVC). The classification results of each model are shown in Figure 6(a) and Figure 6(b), and the evaluation indicators are shown in Tables 11 to 16.

[0133] Table 11 Evaluation indicators of high temperature weather classification model for wind farms

[0134]

[0135] Table 12 Evaluation indicators of low temperature weather classification model for wind farms

[0136]

[0137] Table 13 Evaluation indicators of wind farm strong wind weather classification model

[0138]

[0139]

[0140] Table 14 Evaluation indicators of photovoltaic field high temperature weather classification model

[0141]

[0142] Table 15 Evaluation indicators of photovoltaic field low temperature weather classification model

[0143]

[0144] Table 16 Evaluation indicators of photovoltaic field heavy precipitation weather classification model

[0145]

[0146] A comprehensive analysis of extreme weather classification for the aforementioned wind farm clusters and photovoltaic farm clusters shows that, compared to traditional classification models, PLE's advantage lies in maintaining a high level of prediction results for each task. It is rare for a single task metric to be significantly worse than that of other models. Furthermore, the stronger the correlation between tasks, the better the prediction results. Furthermore, for renewable energy power plants, a greater focus is on the classification model's ability to identify abnormal high-impact weather events—that is, the model's ability to capture positive examples. Therefore, accuracy and recall are more important. Traditional models often suffer from low recall rates in their classification predictions, but PLE can effectively alleviate this problem, making it more versatile in renewable energy power generation.

[0147] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0148] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A classification method for new energy high-impact process weather, characterized in that: include: Screening sensitive weather factors of the target electric field and obtaining historical data of the target electric field; the historical data includes historical power data and NWP data; The target electric field includes a wind farm and a photovoltaic farm; Based on the sensitive weather factors and the historical data, define the weather process with high impact on new energy, determine the meteorological factor threshold of the weather process with high impact on new energy, and use the WGAN generative adversarial network to generate data to obtain a sample set; Constructing a hierarchical extraction and classification model, and performing iterative training using the sample set; the hierarchical extraction and classification model adopts a PLE network, which is composed of a specific expert network and a shared expert network; The trained hierarchical extraction and classification model is used to identify the weather at the time point to be tested, and the final weather classification result is obtained.

2. The new energy high-impact process weather classification method according to claim 1 is characterized in that: The sensitive weather factors of the target electric field specifically include: the sensitive weather factors of the wind field include wind speed and temperature; the sensitive weather factors of the photovoltaic field include solar radiation intensity.

3. The new energy high-impact process weather classification method according to claim 1 is characterized in that: The method of defining a weather process with a high impact on new energy based on the sensitive weather factors and the historical data, determining a threshold of meteorological factors for the weather process with a high impact on new energy, and generating data using a WGAN generative adversarial network to obtain a sample set specifically includes: The abnormal weather time points of the historical data are defined as weather processes with high impact on new energy, and an initial data set is constructed based on the sensitive weather factor data corresponding to the weather processes with high impact on new energy; the abnormal weather time points are weather time points with abnormal power forecast results and abnormal output conditions; The initial data set is expanded using a WGAN generative adversarial network to obtain a sample set; the WGAN generative adversarial network restricts the strategy set of the discriminator to Lipschitz continuity on the basis of a basic generative adversarial network.

4. The new energy high-impact process weather classification method according to claim 3 is characterized in that: The abnormal weather time points of the historical data are defined as new energy high-impact weather processes, and an initial data set is constructed based on the sensitive weather factor data corresponding to the new energy high-impact weather processes, specifically including: Based on the historical data, determining a time point when the power prediction error is greater than a first set threshold as a weather time point when an abnormal power prediction result occurs, and determining a time point when the actual power and the predicted power drop by more than a second set threshold as a weather time point when an abnormal output condition occurs; The average value and standard deviation of the sensitive weather factor data corresponding to the abnormal weather time point are calculated, and the data points that deviate from the average value by 1-2 standard deviations are taken as the meteorological factor threshold for the new energy high-impact weather process; The time points corresponding to meteorological data above the meteorological factor threshold are labeled "1", and the time points corresponding to meteorological data below or equal to the meteorological factor threshold are labeled "0", and the initial data set is generated based on historical data and corresponding data labels.

5. The new energy high-impact process weather classification method according to claim 1 is characterized in that: The construction process of the step-by-step hierarchical extraction classification model includes: Build a pre-trained network based on the PLE network; The sample set is input into the pre-trained network, and the cross entropy loss is selected as the loss function of the model training. When the pre-trained network meets the set number of iterations or the set evaluation index, it is determined to be a trained hierarchical extraction and classification model.

6. A new energy high-impact process weather classification system, characterized in that: include: A data acquisition unit, configured to screen sensitive weather factors of a target electric field and obtain historical data of the target electric field; the historical data includes historical power data and NWP data; the target electric field includes a wind farm and a photovoltaic farm; a sample generation unit, configured to define a weather process with a high impact on new energy based on the sensitive weather factors and the historical data, determine a threshold value of meteorological factors for the weather process with a high impact on new energy, and generate data using a WGAN generative adversarial network to obtain a sample set; A model training unit is used to construct a hierarchical extraction and classification model and perform iterative training using the sample set; the hierarchical extraction and classification model adopts a PLE network, which is composed of a specific expert network and a shared expert network; The weather classification unit is used to identify the weather at the time point to be tested using the trained hierarchical extraction classification model to obtain the final weather classification result.

7. An electronic device, characterized in that: The electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the new energy high-impact process weather classification method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that It stores a computer program, which, when executed by a processor, implements the new energy high-impact process weather classification method according to any one of claims 1 to 5.

Images