A wind power low-output event generated difference training method and prediction method
Patent Information
- Application Number
- CN202510367919.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-26
- Publication Date
- 2026-09-25
- Estimated Expiration
- 2045-03-26
AI Technical Summary
然而，风电的强随机性和波动性带来了显著挑战，尤其是在风电并网比例不断上升的情况下，电力系统的“保供电”能力面临更大考验
[0026]本公开提供了一种风电低出力事件生成式有差训练方法，包括：获取至少一个风电场的运行数据和同一时段至少一个风电场的数值天气预报数据；根据运行数据和数值天气预报数据构建第一训练集；其中，第一训练集是指风电低出力事件相关的真实数据集；根据第一训练集和获取的噪声数据对预先构建的事件样本生成模型进行训练，以根据训练好的事件样本生成模型生成第二训练集；其中，第二训练集是指风电低出力事件相关的生成数据集；基于第一训练集和第二训练集对预先构建的事件预测模型进行训练，得到训练好的事件预测模型；其中，事件预测模型用于预测风电低出力事件。本申请提供的方法，训练得到了能够准确预测未来风电低出力事件的模型，有效提升了提升电网安全稳定性和“保供电”能力。
Smart Images
Figure CN120524318B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power generation technology, and in particular to a differential training method and prediction method for generating low-output wind power events. Background Technology
[0002] As a clean and renewable energy source, wind power will play a vital role in ensuring power supply. However, the inherent randomness and volatility of wind power present significant challenges, especially with the increasing proportion of wind power connected to the grid, putting the power system's ability to guarantee power supply to the greater test. Low wind power output events refer to a sharp decline or shutdown of wind power output within a specific period, severely impacting the grid's supply and demand balance and potentially leading to safety incidents such as grid frequency drops and power outages. Therefore, accurately predicting low wind power output events is crucial for improving grid security, stability, and the ability to guarantee power supply. Summary of the Invention
[0003] To address the aforementioned technical problems, this disclosure provides a generative differential training method and a prediction method for wind power low output events.
[0004] In a first aspect, embodiments of this disclosure provide a generative differential training method for low-output wind power events, including:
[0005] Obtain operational data from at least one wind farm and numerical weather forecast data from at least one wind farm during the same time period;
[0006] The first training set is constructed based on operational data and numerical weather prediction data; the first training set refers to the real dataset related to low wind power output events.
[0007] The pre-built event sample generation model is trained based on the first training set and the acquired noise data, so as to generate a second training set based on the trained event sample generation model; wherein, the second training set refers to the generated dataset related to low wind power output events.
[0008] The pre-built event prediction model is trained based on the first and second training sets to obtain the trained event prediction model; the event prediction model is used to predict low wind power output events.
[0009] Secondly, embodiments of this disclosure provide a method for predicting low-output wind power events, including:
[0010] Obtain weather forecast data for a preset time period;
[0011] Weather forecast data is used as input to the event prediction model to predict low wind power output events within a preset time period, and the event prediction results are obtained. The event prediction model is trained using the first aspect of the wind power low output event generative differential training method.
[0012] Thirdly, embodiments of this disclosure provide a wind power low-output event generation differential training device, comprising:
[0013] The first acquisition unit is used to acquire the operation data of at least one wind farm and the numerical weather forecast data of at least one wind farm in the same period.
[0014] The construction unit is used to construct the first training set based on operational data and numerical weather forecast data; wherein, the first training set refers to the real dataset related to low wind power output events;
[0015] The first training unit is used to train a pre-built event sample generation model based on the first training set and the acquired noise data, so as to generate a second training set based on the trained event sample generation model; wherein, the second training set refers to the generated dataset related to low wind power output events.
[0016] The second training unit is used to train the pre-built event prediction model based on the first and second training sets to obtain the trained event prediction model; wherein the event prediction model is used to predict wind power low output events.
[0017] Fourthly, embodiments of this disclosure provide a device for predicting low-output wind power events, comprising:
[0018] The second acquisition unit is used to acquire weather forecast data for a preset time period;
[0019] The event prediction unit is used to take weather forecast data as input to the event prediction model, predict low wind power output events within a preset time period, and obtain the event prediction results; wherein, the event prediction model is trained by the method in the first aspect.
[0020] Fifthly, embodiments of this disclosure provide an electronic device, including:
[0021] Memory;
[0022] Processor; and
[0023] Computer programs;
[0024] The computer program is stored in memory and configured to be executed by a processor to implement the methods described in the first and second aspects above.
[0025] In a sixth aspect, embodiments of this disclosure provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the methods described in the first and second aspects above.
[0026] This disclosure provides a generative training method for low-output wind power events, comprising: acquiring operational data of at least one wind farm and numerical weather prediction data of at least one wind farm during the same period; constructing a first training set based on the operational data and numerical weather prediction data; wherein the first training set refers to a real dataset related to low-output wind power events; training a pre-constructed event sample generation model based on the first training set and acquired noise data to generate a second training set based on the trained event sample generation model; wherein the second training set refers to a generated dataset related to low-output wind power events; training a pre-constructed event prediction model based on the first and second training sets to obtain a trained event prediction model; wherein the event prediction model is used to predict low-output wind power events. The method provided in this application trains a model capable of accurately predicting future low-output wind power events, effectively improving the safety and stability of the power grid and its ability to ensure power supply. Attached Figure Description
[0027] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.
[0028] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.
[0029] Figure 1 A flowchart illustrating a wind power low-output event generation differential training method provided in this embodiment of the present disclosure;
[0030] Figure 2 This is a schematic diagram of the structure of an event sample generation model provided in an embodiment of the present disclosure;
[0031] Figure 3 This is a schematic diagram of the structure of an event prediction model provided in an embodiment of the present disclosure;
[0032] Figure 4 A flowchart illustrating a method for predicting low wind power output events provided in this embodiment of the disclosure;
[0033] Figure 5 A flowchart illustrating another wind power low-output event generation differential training method provided in this embodiment of the present disclosure;
[0034] Figure 6 This is a schematic diagram of the structure of a wind power low output event generation differential training device provided in an embodiment of the present disclosure;
[0035] Figure 7 A schematic diagram of the structure of a wind power low-output event prediction device provided in an embodiment of this disclosure;
[0036] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation
[0037] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.
[0038] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.
[0039] Currently, most research focuses on predicting future wind power output sequences, with limited studies specifically addressing low-output wind power events. Due to the low probability of these events and their significant small sample size, existing prediction models struggle to be effectively trained using only representative samples, resulting in suboptimal prediction performance. Therefore, there is an urgent need to develop small-sample prediction methods suitable for low-output wind power events to improve the wind power dispatch capability of power systems and ensure reliable grid operation. Solving this problem is crucial for achieving low-carbon and sustainable development of the power system.
[0040] To address the aforementioned technical problems, this disclosure provides a generative differential training method for low-output wind power events. This will be described in detail through one or more of the following embodiments.
[0041] The wind power low-output event generative differential training method provided in this disclosure is applicable to model training scenarios. This method can be executed by a wind power low-output event generative differential training device, which can be implemented in software and / or hardware and can be integrated into an electronic device. The electronic device can include, but is not limited to, mobile terminals such as smartphones, laptops, digital radio receivers, personal digital assistants (PDAs), tablet PCs, PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), wearable devices, etc., as well as fixed terminals such as digital televisions, desktop computers, smart home devices, etc.
[0042] Figure 1 This is a flowchart illustrating a wind power low-output event generative differential training method provided in an embodiment of this disclosure, specifically including as follows: Figure 1 The steps shown are as follows:
[0043] S101. Obtain operational data of at least one wind farm and numerical weather forecast data of at least one wind farm during the same period.
[0044] Understandably, at least one wind farm is selected as the research object from among multiple wind farms within the study area. A wind farm refers to an area with a concentrated installation of multiple wind turbine generators, possessing a certain scale of power generation facilities. Subsequently, actual operational data from at least one wind farm is acquired. This wind farm can form a wind power cluster in the region. Operational data typically includes wind turbine power, output, wind speed, wind direction, temperature, humidity, and other data, directly reflecting the wind farm's power generation performance and output changes. Simultaneously, numerical weather prediction data for at least one wind farm is acquired for the same time period. Numerical weather prediction data is meteorological data calculated and simulated based on atmospheric physics and dynamics equations, including predicted wind speed, air pressure, humidity, and other data. Data acquired at the same time period reflects conditions within the same time frame, ensuring correlation and consistency between data, and aiding in analyzing the predictive impact of weather forecasts on wind farm power generation capacity. Furthermore, since wind speed and direction directly affect wind farm power generation capacity, accurate weather forecast data is crucial for predicting low wind power output events.
[0045] Understandably, after obtaining the operational data, preprocessing such as data cleaning can be performed. Preprocessing includes verification and screening, removal of erroneous data, and reasonable interpolation and correction.
[0046] One possible implementation uses a wind power cluster consisting of multiple wind farms in a province as the research object. The operational data mainly includes power data, while the numerical weather prediction data from multiple wind farm sites mainly includes wind speed data. The power data sampling period is one year, with a time resolution of 15 minutes, meaning each 15-minute interval is considered a sampling point. Power data for each wind farm is collected at each sampling point, resulting in multiple power data points collected at each point. The total power data collected at each sampling point can then be calculated, and further, all total power data collected every 10 days can be considered as a single sample. The numerical weather prediction data also has a one-year sampling period, and both the weather forecast and operational data are collected within the same timeframe, with a time resolution of 15 minutes. The forecast period is 10 days into the future, meaning the weather conditions for the next 10 days are predicted on the 1st. All wind speed data predicted for the next 10 days is considered as a single sample. The sampling interval for a single sample in the operational data is 10 days, and the corresponding predicted wind speed data in the weather forecast data is also for 10 days, meaning there is a one-to-one correspondence between the sampling parameters of the two datasets.
[0047] S102. Construct the first training set based on operational data and numerical weather forecast data.
[0048] The first training set refers to the real dataset related to low wind power output events; the first training set includes the first sample set and the first label set of the first sample set.
[0049] Understandably, based on the above S101, a first training set is constructed using operational data and weather forecast data from at least one wind farm. The first training set includes a first sample set and a first label set. The first sample set includes multiple real samples, each of which is a combination of operational data and weather forecast data. For example, a sample might be "Wind farm output: 50MW, forecast wind speed: 10m / s". The first label set includes labels for each real sample; for example, a label could be the type of power fluctuation corresponding to a low wind power output event.
[0050] Optionally, a first training set can be constructed based on the operational data and numerical weather prediction data, which can be achieved through the following steps:
[0051] Calculate a first sequence of operational data; wherein the first sequence reflects the sum of operational data collected by at least one wind farm within a set time period; perform dimensionality reduction processing on numerical weather forecast data to obtain a second sequence with the same dimension as the first sequence; fuse the first sequence and the second sequence to obtain a first sample set; identify at least one low-output wind power event corresponding to the first sample set, and determine the label of at least one low-output wind power event to construct a first label set.
[0052] Understandably, taking 10 days of power data as a real sample, the total power data collected from multiple wind farms at each sampling point over 10 days is calculated. This total power data from multiple real samples forms the first sequence. The weather forecast data is then dimensionality-reduced to obtain a second sequence with the same dimension as the first sequence; both the first and second sequences are one-dimensional. Specifically, principal component analysis can be used to reduce the dimensionality of the weather forecast data, resulting in a one-dimensional dataset. Subsequently, the first and second sequences corresponding to the same time period are fused, for example, by splicing, to obtain the first sample set. Low-output wind power events are identified from the multiple real samples included in the first sample set, and the label of each sample is determined based on the identified category, resulting in the first label set. Specifically, with a 10-day sampling period, the total power data of multiple wind farms collected every 15 minutes is calculated, meaning one total power data point corresponds to every 15 minutes. For example, if three power data points (denoted as a, b, and c) are collected from three wind farms, the sum of a, b, and c is calculated to obtain the total power data (denoted as d), where d represents the sampling data corresponding to one sampling point. Two total power data points exist after 30 minutes, and a large number of total power data points exist after 10 days. This large number of total power data points are statistically analyzed to determine whether they meet the identification criteria of a duration exceeding one day and an average power less than 10% of the rated power. If these criteria are met, the sample is identified as a genuine sample of a low-output wind power event. The identification method for other low-output wind power events is not limited. Subsequently, a low-output wind power event sample set is constructed based on the first label set and the first sample set.
[0053] Among them, at least one wind power low output event includes small fluctuations, medium fluctuations, and large fluctuations, which are classified by the power fluctuation amplitude.
[0054] Understandably, based on the first sample set, the power fluctuation amplitude of the real samples is used as the screening criterion to classify the identified samples containing low wind power output events into small fluctuations (often occurring in summer), medium fluctuations (often occurring in autumn), and large fluctuations (often occurring in winter or spring). Understandably, different power fluctuation amplitudes usually correspond to different seasons.
[0055] Optionally, identify at least one low-output wind power event corresponding to the first sample set, determine the label of at least one low-output wind power event, and construct a first label set. This can be achieved through the following steps:
[0056] Wind power low output events are identified based on the total power data included in each sample in the first sample set; the wind power low output events corresponding to the first sample whose total power data is less than a first set threshold are identified as small fluctuations; the wind power low output events corresponding to the second sample whose total power data is greater than the first set threshold and less than the second set threshold are identified as medium fluctuations; the wind power low output events corresponding to the third sample whose total power data is greater than the second set threshold are identified as large fluctuations; and the small fluctuations, medium fluctuations, or large fluctuations corresponding to each sample are encoded to construct a first label set.
[0057] Understandably, based on the first sample set, power fluctuation amplitudes are classified according to a first set threshold and a second set threshold. Specifically, the first set threshold is 30%, and the second set threshold is 60%. Power fluctuation amplitudes less than 30% are classified as small fluctuations, power fluctuation amplitudes greater than 30% and less than 60% are classified as medium fluctuations, and power fluctuation amplitudes greater than 60% are classified as large fluctuations. Subsequently, based on the classification results, a set of conditional labels for low wind power output events is constructed using OneHot encoding. For example, small fluctuations are encoded as 001, medium fluctuations as 002, and large fluctuations as 003. The number of set thresholds can be set according to user needs and is not limited here. In addition, after classifying the power fluctuation amplitudes and types according to the set thresholds, the accuracy of the classification of power fluctuation amplitudes for low wind power output events can be verified by referring to the seasons in which each type of power fluctuation amplitude usually occurs.
[0058] S103. Train the pre-built event sample generation model based on the first training set and the acquired noise data, so as to generate a second training set based on the trained event sample generation model.
[0059] The second training set refers to the generated dataset related to low wind power output events.
[0060] Understandably, based on the above S102, a first training set (including a first sample set and a first label set) and noisy data are used to train the event sample generation model. The event sample generation model learns how to generate new training samples based on the first training set, and then generates a second training set containing a large number of samples based on the noisy data and the wind power low-output event label set. The event sample generation model expands the wind power low-output event sample pool, solves the problem of small sample size, and provides crucial input for accurate prediction of wind power low-output events.
[0061] Optionally, the pre-built event sample generation model can be trained using the first training set and the acquired noisy data to generate a second training set. This can be achieved through the following steps:
[0062] An event sample generation model is constructed for generating low-output wind power events. The event sample generation model includes a generator and a discriminator. The acquired noise data and the first label set included in the first training set are used as inputs to the generator to obtain the generated sample set output by the generator. The generated sample set and the first sample set included in the first training set are used as inputs to the discriminator to obtain the true / false judgment result of the sample output by the discriminator. When the generator and the discriminator reach an equilibrium state after adversarial training, the trained event sample generation model is obtained, and a second training set is generated based on the trained event sample generation model.
[0063] Understandably, an event sample generation model for low wind power output events can be constructed based on a conditional generative adversarial network (GAN). A GAN consists of a generator and a discriminator, and its essence is game-like learning between the generator and the discriminator. The training objective of the generator is to make the probability distribution p of the generated samples G(z)... G (z) As closely as possible to the probability distribution p of the real sample x data (z) Same. The training goal of the discriminator is to correctly distinguish whether its input is a real sample or a generated sample as much as possible. After the generator and discriminator are trained adversarially, they can reach a state of equilibrium, so that the distribution of artificially generated data gradually approaches the real data, achieving a level of realism that is indistinguishable from the real data.
[0064] Understandably, noise data can be measurement noise. When sensor measurements are inaccurate, the measured data such as wind speed, temperature, and humidity may contain errors. Other possible noise data are not limited and can be determined according to user needs. Subsequently, the noise data and the first label set (serving as real labels) are used as input to the generator, outputting multiple first generated samples to construct a generated sample set. Then, the generated sample set and the first sample set including real samples are used as input to the discriminator, outputting a judgment result for each generated sample, indicating whether the generated sample is real or fake. After the discriminator and generator reach an equilibrium state through at least one adversarial training, a trained discriminator and a trained generator are obtained, resulting in a trained event sample generation model. Subsequently, the noise data and the first label set are used as input to the trained generator, outputting multiple second generated samples to form a second training set. The noise data can be the same as or different from the noise data used during generator training; this is not limited here.
[0065] In one possible implementation, during the training of the event sample generation model, the learning rate of the generator can be set to 0.0001, the learning rate of the discriminator to 0.0005, the batch size to 16, and a total of 1000 epochs can be trained.
[0066] The generator comprises at least one first multilayer perceptron, a long short-term memory network, and at least one first attention layer. The at least one first multilayer perceptron is sequentially connected, and the at least one first attention layer is connected to each other via residual connections. The long short-term memory network connects the last first multilayer perceptron and the first first attention layer. The first first multilayer perceptron receives noise data and a first set of labels as input, and the last residual connection outputs the generated sample set. The discriminator comprises at least one second multilayer perceptron and at least one second attention layer. The at least one second multilayer perceptron is sequentially connected, and the at least one second attention layer is connected to each other via residual connections. The last second attention layer is connected to the first second multilayer perceptron. The first residual connection receives the first sample set and the generated sample set as input, and the last second multilayer perceptron outputs the true or false judgment result of the sample.
[0067] For example, see Figure 2 , Figure 2 This is a schematic diagram of the structure of an event sample generation model provided in an embodiment of the present disclosure. The generator consists of at least one first multilayer perceptron, a long short-term memory network, and at least one first attention layer (Self-Attention), as shown below. Figure 2 As shown, the generator includes two first multilayer perceptrons, one long short-term memory network (LSTM), and two first attention layers. The two first multilayer perceptrons are sequentially connected. The first first multilayer perceptron receives noisy data and a first label set. The LSTM network connects the second first multilayer perceptron and the first first attention layer. The two first attention layers are connected via residual connections, which can be understood as convolutional neural networks (CNNs), meaning each first attention layer is connected to a CNN. The last CNN outputs the first generated sample. It is understood that the number of first multilayer perceptrons, LSTM networks, and first attention layers is not limited. The discriminator consists of at least one second multilayer perceptron and at least one second attention layer (Self-Attention), as shown... Figure 2 As shown, the discriminator includes two second multilayer perceptrons and two second attention layers. The two second attention layers are connected via residual connections, which can also be viewed as convolutional neural networks (CNNs). That is, the input to each second attention layer is the output of a CNN. The first CNN receives the first sample set and the first generated sample. The two second multilayer perceptrons are connected sequentially, and the last second attention layer is connected to the first second multilayer perceptron. The last second multilayer perceptron outputs the true / false judgment result of the sample. It is understandable that the network structures of the first and second multilayer perceptrons can be the same or different.
[0068] S104. Train the pre-built event prediction model based on the first and second training sets to obtain the trained event prediction model.
[0069] Among them, the event prediction model is used to predict low wind power output events.
[0070] Understandably, based on the above S103, an event prediction model for predicting low wind power output events is established based on a meta-learning architecture. This meta-learning architecture includes two modules: a base learner and a meta-learner. The base learner uses CNN-ProbSparse and Attention-GRU models, while the meta-learner uses an XGBoost model. Subsequently, the event prediction model is trained using a first training set (as real samples) and a second training set (as generated samples). This event prediction model is used to predict whether a low wind power output event will occur in the future and to forecast the power curve based on weather forecast data that predicts future weather conditions.
[0071] Optionally, the pre-built event prediction model can be trained based on the first and second training sets to obtain a trained event prediction model. This can be achieved through the following steps:
[0072] An event prediction model for predicting low wind power output events is constructed; the event prediction model includes a base learner and a meta learner; the base learner is trained based on a second training set to obtain a trained base learner; the first sample set included in the first training set is used as the input of the trained base learner to obtain the event prediction result output by the base learner; the meta learner is trained based on the event prediction result and the first sample set.
[0073] Understandably, the training process of the event prediction model is as follows: First, the base learner is trained based on the second training set (generated samples) to obtain a trained base learner. Then, the first sample set (real samples) is input into the trained base learner to obtain the preliminary event prediction result for each real sample output by the base learner. Finally, the event prediction results and the first sample set are fused to train the meta-learner, resulting in a trained meta-learner.
[0074] Optionally, the base learner can be trained using the second training set to obtain a trained base learner. This can be achieved through the following steps:
[0075] Identify at least one low-output wind power event corresponding to the second sample set included in the second training set, and determine at least one generated sample corresponding to each low-output wind power event; set a corresponding sub-module for each low-output wind power event; wherein, the base learner includes at least one sub-module; train each sub-module according to at least one generated sample corresponding to each sub-module.
[0076] Understandably, average power and event duration are used as identification criteria to identify at least one low-output wind power event corresponding to the generated samples included in the second training set. Subsequently, power fluctuation amplitude is used as a screening condition to determine the second sample set corresponding to each power fluctuation amplitude, that is, the second training set is further divided into multiple second sample sets. The power fluctuation amplitude includes the aforementioned small, medium, and large fluctuation types. The specific identification and screening processes are described in the above embodiments and will not be repeated here. A corresponding sub-module is set up for each power fluctuation amplitude's second sample set. The base learner includes multiple sub-modules, or multiple base learners are directly set up, with one corresponding base learner for each power fluctuation amplitude's second sample set. The following detailed explanation uses setting up one sub-module for each power fluctuation amplitude as an example. For each sub-module, the corresponding second sample set is used as input to train each sub-module.
[0077] Optionally, the submodule includes a convolutional neural network model, an attention layer, multiple recurrent neural network models, and multiple linear layers corresponding to the multiple recurrent neural network models. The attention layer connects the convolutional neural network model and the multiple recurrent neural network models. The convolutional neural network model receives at least one corresponding generated sample, and the multiple linear layers output event prediction results. The event prediction results include occurrence prediction results and power prediction results. The occurrence prediction results are used to record whether a low wind power output event has occurred, and the power prediction results are used to record the power change of the low wind power output event.
[0078] For example, see Figure 3 , Figure 3This is a schematic diagram of the structure of an event prediction model provided in an embodiment of the present disclosure. The event prediction model includes a base model (base learner) and a meta-model (original learner). The base model includes a CNN, Attention, two GRUs, and two Linears. The input of the CNN is a TASK, which refers to a second sample set corresponding to a certain power fluctuation amplitude. For example, TASK1 refers to the second sample set corresponding to small fluctuations, TASK2 refers to the second sample set corresponding to medium fluctuations, and TASK3 refers to the second sample set corresponding to large fluctuations. The output of the CNN is the input of the Attention, and the output of the Attention is the input of the two GRUs. Each GRU has a corresponding Linear, and the output of the GRU is the input of the Linear. A multi-task framework is used to construct an occurrence prediction decoder based on GRU, Linear, and Focal loss and a power prediction decoder based on GRU, Linear, and MSE loss. The occurrence prediction decoder is used to predict whether a low wind power output event has occurred, and the power prediction decoder is used to generate the power curve when a low wind power output event has occurred. After training the base models corresponding to the three power fluctuation amplitudes, real samples are used as input to at least one base model to obtain the occurrence prediction result and power prediction result output by at least one base model. Subsequently, the occurrence prediction result and power prediction result output by each base model can be fused to obtain the fused prediction result. The input of the meta-learner is the fused prediction result and the first sample set as real samples, and the output is the occurrence prediction result used to record whether a low wind power output event has occurred, as well as the power prediction result used to record the power curve, and other event prediction results.
[0079] Optionally, the meta-learner can be trained based on the event prediction results and the first sample set, which can be achieved through the following steps:
[0080] The at least one event prediction result output by at least one sub-module of the base learner is fused to obtain a fused prediction result; the meta-learner is trained based on the fused prediction result and the first sample set.
[0081] Understandably, the event prediction results output by each base model are fused to obtain a fused prediction result. For example, taking three base models, the three occurrence prediction results can be fused / stitched together, the three power prediction results can be stitched together, and finally the two parts can be stitched together to obtain the fused prediction result. Alternatively, the occurrence prediction results and power prediction results output by each base model can be stitched together first, and then the three parts can be stitched together to obtain the fused prediction result. The specific fusion order is not limited and can be determined according to user needs. Subsequently, the fused prediction result is used as the input to the meta-model, which outputs a final occurrence prediction result and a power prediction result.
[0082] In one possible implementation, during the training of the event prediction model, the second training set is conditionally divided into three sample sets: small fluctuations, medium fluctuations, and large fluctuations, as described above based on the power fluctuation amplitude. One base learner is trained for each sample set, with a learning rate of 0.0001, a batch size of 16, and an epoch of 100. The Focalloss loss function uses alpha of 0.7 and beta of 2. The meta-learner model is XGBoost, with a learning rate of 0.001, n_estimators = 200, and max_depth = 6.
[0083] The wind power low-output event generative differential training method provided in this disclosure firstly builds an event sample generation model based on generative adversarial networks, realizing the effective generation of wind power low-output event samples and expanding the wind power low-output event sample, which can cope with the small sample size of low-output events. Then, considering the "quantity-value" difference between generated samples and real samples, a wind power low-output event prediction model is established based on a meta-learning architecture to solve the difference between generated samples and real samples. Through multi-task learning, it achieves joint accurate prediction of whether a wind power low-output event has occurred and the power of the wind power low-output event. On the one hand, it provides early warning for the power system, ensuring the safe and stable operation of the power system, and on the other hand, it also improves the wind power penetration rate. Secondly, by using the pre-trained base model with generated samples and fine-tuning the model with real samples, the generalization ability of the model is improved, and accurate prediction of wind power low-output events is achieved. In addition, the wind power low-output event generative differential training prediction can be used in power system optimal scheduling and is applicable to any time scale and spatial scale.
[0084] Based on the above embodiments, Figure 4 A flowchart illustrating a method for predicting low-output wind power events provided in this embodiment of the disclosure, specifically including as follows: Figure 4 The following steps are shown:
[0085] S401. Obtain weather forecast data for a preset time period.
[0086] Understandably, the preset time period can be the next ten days. That is, after the event prediction model is trained, weather forecast data for future events can be obtained. The length of the preset time period is not limited.
[0087] S402. Use weather forecast data as input to the event prediction model to predict low wind power output events within a preset time period and obtain the event prediction results.
[0088] The event prediction model was trained using the aforementioned wind power low-output event generative differential training method.
[0089] Understandably, based on the above S401, weather forecast data is used as input to the event prediction model trained using the aforementioned wind power low-output event generative difference training method. The event prediction model predicts wind power low-output events for the next ten days, obtaining the event prediction results for the next ten days. These results include whether a wind power low-output event will occur in the next ten days and the power at which such an event occurs. Understandably, the specific prediction process of the event prediction model is described in the above embodiment and will not be repeated here.
[0090] The method for predicting low-output wind power events disclosed herein enables accurate prediction of such events.
[0091] Based on the above embodiments, Figure 5 This is a flowchart illustrating another wind power low-output event generative differential training method provided in this disclosure embodiment, specifically including as follows: Figure 5 The following steps are shown:
[0092] 1) Obtain actual operating data of each wind farm in the region, and clean and preprocess the actual operating data. At the same time, obtain numerical weather forecast data of each wind farm location during the same period; 2) Perform dimensionality reduction processing on the numerical weather forecast data; 3) Identify low wind power output events and construct a sample set of low wind power output events; 4) Construct a set of conditional labels for low wind power output events; 5) Construct a sample generation model for low wind power output events; 6) Implementation of low wind power output event sample generation based on the event sample generation model; 7) Establish a prediction model for low wind power output events based on a meta-learning architecture; 8) Train the prediction model for low wind power output events.
[0093] It is understood that the specific implementation steps of 1)-8) above are as described in the above embodiments, and will not be repeated here.
[0094] Based on the above embodiments, Figure 6This is a schematic diagram of a wind power low-output event generation differential training device provided in an embodiment of this disclosure. The wind power low-output event generation differential training device provided in this embodiment can execute the processing flow provided in the embodiment of the wind power low-output event generation differential training method, such as... Figure 6 As shown, the device 600 includes:
[0095] The first acquisition unit 601 is used to acquire the operation data of at least one wind farm and the numerical weather forecast data of at least one wind farm in the same period.
[0096] Construction unit 602 is used to construct the first training set based on operational data and numerical weather forecast data; wherein, the first training set refers to the real dataset related to low wind power output events;
[0097] The first training unit 603 is used to train a pre-built event sample generation model based on the first training set and the acquired noise data, so as to generate a second training set based on the trained event sample generation model; wherein, the second training set refers to the generation dataset related to low wind power output events.
[0098] The second training unit 604 is used to train the pre-built event prediction model based on the first training set and the second training set to obtain the trained event prediction model; wherein, the event prediction model is used to predict wind power low output events.
[0099] The first training set includes the first sample set and the first label set of the first sample set.
[0100] Optionally, building unit 602 is used for:
[0101] Calculate the first sequence of operational data; wherein the first sequence reflects the sum of operational data collected by at least one wind farm within a set time period;
[0102] Numerical weather forecast data is dimensionality reduced to obtain a second sequence with the same dimensions as the first sequence;
[0103] The first sequence and the second sequence are fused to obtain the first sample set;
[0104] Identify at least one low-output wind power event corresponding to the first sample set, and determine the label of at least one low-output wind power event to construct the first label set.
[0105] Among them, at least one wind power low output event includes small fluctuations, medium fluctuations, and large fluctuations, which are classified by the power fluctuation amplitude.
[0106] Optionally, building unit 602 is used for:
[0107] Wind power low output events are identified based on the total power data included in each sample in the first sample set.
[0108] The wind power low output event corresponding to the first sample whose total power data is less than the first set threshold is determined as a small fluctuation.
[0109] The wind power low output event corresponding to the second sample whose total power data is greater than the first set threshold and less than the second set threshold is determined as a medium amplitude fluctuation.
[0110] The low wind power output event corresponding to the third sample whose total power data is greater than the second set threshold is identified as a large fluctuation;
[0111] Encode the small, medium, or large fluctuations corresponding to each sample to construct the first label set.
[0112] Optionally, the first training unit 603 is used for:
[0113] An event sample generation model is constructed for generating low-output wind power events; the event sample generation model includes a generator and a discriminator.
[0114] The acquired noisy data and the first label set included in the first training set are used as input to the generator to obtain the generated sample set output by the generator;
[0115] The generated sample set and the first sample set included in the first training set are used as inputs to the discriminator to obtain the true or false judgment result of the sample output by the discriminator;
[0116] Once the generator and discriminator reach an equilibrium state through adversarial training, a well-trained event sample generation model is obtained, which is then used to generate a second training set.
[0117] Optionally, the generator in the device 600 includes at least one first multilayer perceptron, a long short-term memory network, and at least one first attention layer. The at least one first multilayer perceptron is connected sequentially, and the at least one first attention layer is connected to each other through residual connections. The long short-term memory network connects the last first multilayer perceptron and the first first attention layer. The first first multilayer perceptron receives noise data and a first label set as input, and the last residual connection outputs to generate a sample set.
[0118] The discriminator includes at least one second multilayer perceptron and at least one second attention layer. The at least one second multilayer perceptron is connected sequentially, and the at least one second attention layer is connected to each other through residual connections. The last second attention layer is connected to the first second multilayer perceptron. The first residual connection receives the first sample set and the generated sample set as input, and the last second multilayer perceptron outputs the true or false judgment result of the sample.
[0119] Optionally, the second training unit 604 is used for:
[0120] Construct an event prediction model for predicting low wind power output events; the event prediction model includes a base learner and a meta-learner;
[0121] The base learner is trained based on the second training set to obtain a trained base learner;
[0122] The first sample set included in the first training set is used as the input of the trained base learner to obtain the event prediction results output by the base learner.
[0123] The meta-learner is trained based on the event prediction results and the first sample set.
[0124] Optionally, the second training unit 604 is used for:
[0125] Identify at least one low-output wind power event corresponding to the second sample set included in the second training set, and determine at least one generated sample corresponding to each low-output wind power event;
[0126] A corresponding submodule is set up for each type of low wind power output event; wherein, the base learner includes at least one submodule;
[0127] Each submodule is trained based on at least one generated sample corresponding to each submodule.
[0128] Optionally, the submodule of device 600 includes a convolutional neural network model, an attention layer, multiple recurrent neural network models, and multiple linear layers corresponding to the multiple recurrent neural network models. The attention layer connects the convolutional neural network model and the multiple recurrent neural network models. The convolutional neural network model receives at least one corresponding generated sample, and the multiple linear layers output event prediction results. The event prediction results include occurrence prediction results and power prediction results. The occurrence prediction results are used to record whether a low wind power output event has occurred, and the power prediction results are used to record the power change of the low wind power output event.
[0129] Optionally, the second training unit 604 is used for:
[0130] The prediction results of at least one event output by at least one submodule of the base learner are fused to obtain a fused prediction result.
[0131] The meta-learner is trained based on the fusion prediction results and the first sample set.
[0132] Figure 6 The model training apparatus of the illustrated embodiment can be used to execute the technical solutions of the above method embodiments. Its implementation principle and technical effect are similar, and will not be repeated here.
[0133] Based on the above embodiments, Figure 7 This is a schematic diagram of a wind power low-output event prediction device provided in an embodiment of this disclosure. The wind power low-output event prediction device provided in this embodiment can execute the processing flow provided in the wind power low-output event prediction method embodiment, such as... Figure 7 As shown, the wind power low-output event prediction device 700 includes:
[0134] The second acquisition unit 701 is used to acquire weather forecast data for a preset time period;
[0135] The event prediction unit 702 is used to take weather forecast data as input to the event prediction model, predict low wind power output events within a preset time period, and obtain the event prediction results; wherein, the event prediction model is trained by the above-mentioned wind power low output event generative differential training method.
[0136] Figure 7 The wind power low output event prediction device of the embodiment shown can be used to execute the technical solution of the above method embodiment. Its implementation principle and technical effect are similar, and will not be described again here.
[0137] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. See below for details. Figure 8 The diagram illustrates a structural schematic suitable for implementing the electronic device 100 in the embodiments of this disclosure. The electronic device 100 in the embodiments of this disclosure may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), wearable electronic devices, etc., as well as fixed terminals such as digital TVs, desktop computers, smart home devices, etc. Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.
[0138] like Figure 8 As shown, the electronic device 100 may include a processing unit 101 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 102 or a program loaded from a storage device 108 into a random access memory (RAM) 103 to implement the wind power low-output event generation differential training method and prediction method as described in the embodiments of this disclosure. Various programs and data required for the operation of the electronic device 100 are also stored in the RAM 103. The processing unit 101, ROM 102, and RAM 103 are interconnected via a bus 104. An input / output (I / O) interface 105 is also connected to the bus 104.
[0139] Typically, the following devices can be connected to I / O interface 105: input devices 106 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 107 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 108 including, for example, magnetic tapes, hard disks, etc.; and communication devices 109. Communication device 109 allows electronic device 100 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 7 An electronic device 100 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.
[0140] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts, thereby implementing the above-described wind power low-output event generative differential training method and prediction method. In such embodiments, the computer program can be downloaded and installed from a network via communication device 109, or installed from storage device 108, or installed from ROM 102. When the computer program is executed by processing device 101, it performs the functions defined in the methods of embodiments of this disclosure.
[0141] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.
[0142] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.
[0143] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.
[0144] Optionally, when one or more of the above-described procedures are executed by the electronic device, the electronic device may also perform other steps described in the above embodiments.
[0145] Computer program code for performing the operations of this disclosure can be written in one or more programming languages or a combination thereof, including but not limited to object-oriented programming languages such as Java, Smalltalk, and C++, as well as conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on the device computer, partially on the device computer, as a standalone software package, partially on the device computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer can be connected to the device computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).
[0146] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0147] The units described in the embodiments of this disclosure can be implemented in software or hardware. The names of the units are not, in some cases, intended to limit the specific unit.
[0148] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.
[0149] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
[0150] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or gateway that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or gateway. Without further limitations, an element defined by the phrase "comprising a model training" does not exclude the presence of other identical elements in the process, method, article, or gateway that includes said element.
[0151] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A differential training method for generating low-output wind power events, characterized in that, include: Obtain operational data from at least one wind farm and numerical weather forecast data from the at least one wind farm during the same time period; A first training set is constructed based on the operational data and the numerical weather forecast data; wherein, the first training set refers to a real dataset related to low wind power output events, and the first training set includes a first sample set and a first label set of the first sample set; The pre-built event sample generation model is trained based on the first training set and the acquired noise data, so as to generate a second training set based on the trained event sample generation model; wherein, the second training set refers to the generated dataset related to low wind power output events. The pre-built event prediction model is trained based on the first training set and the second training set to obtain the trained event prediction model; wherein, the event prediction model is used to predict wind power low output events; The event sample generation model includes a generator and a discriminator. The generator includes at least one first multilayer perceptron, a long short-term memory network, and at least one first attention layer. The at least one first multilayer perceptron is connected sequentially, and the at least one first attention layer is connected to each other via residual connections. The long short-term memory network connects the last first multilayer perceptron and the first first attention layer. The first first multilayer perceptron receives the noise data and the first label set as input, and the last residual connection outputs to generate a sample set. The discriminator includes at least one second multilayer perceptron and at least one second attention layer. The at least one second multilayer perceptron is connected sequentially. The at least one second attention layer is connected to each other through a residual connection. The last second attention layer is connected to the first second multilayer perceptron. The first residual connection receives the first sample set and the generated sample set as input. The last second multilayer perceptron outputs the true or false judgment result of the sample. The step of training a pre-built event prediction model based on the first training set and the second training set to obtain the trained event prediction model includes: An event prediction model for predicting low wind power output events is constructed; wherein, the event prediction model includes a base learner and a meta learner; the base learner is trained according to a second training set to obtain a trained base learner; a first sample set included in the first training set is used as the input of the trained base learner to obtain the event prediction result output by the base learner; the meta learner is trained according to the event prediction result and the first sample set. The step of training the base learner based on the second training set to obtain a trained base learner includes: Identify at least one low-output wind power event corresponding to the second sample set included in the second training set, and determine at least one generated sample corresponding to each low-output wind power event; set a corresponding sub-module for each low-output wind power event; wherein, the base learner includes at least one sub-module; train each sub-module according to at least one generated sample corresponding to each sub-module; The submodule includes a convolutional neural network model, an attention layer, multiple recurrent neural network models, and multiple linear layers corresponding to the multiple recurrent neural network models. The attention layer connects the convolutional neural network model and the multiple recurrent neural network models. The convolutional neural network model receives at least one corresponding generated sample, and the multiple linear layers output the event prediction result. The event prediction result includes an occurrence prediction result and a power prediction result. The occurrence prediction result is used to record whether a low wind power output event has occurred, and the power prediction result is used to record the power change of the low wind power output event.
2. The method according to claim 1, characterized in that, The construction of the first training set based on the operational data and the numerical weather forecast data includes: Calculate a first sequence of the operational data; wherein the first sequence reflects the sum of operational data collected by the at least one wind farm within a set time period; The numerical weather forecast data is subjected to dimensionality reduction processing to obtain a second sequence with the same dimension as the first sequence; The first sequence and the second sequence are fused to obtain the first sample set; Identify at least one low-output wind power event corresponding to the first sample set, determine the label of the at least one low-output wind power event, and construct the first label set.
3. The method according to claim 2, characterized in that, The at least one low-output wind power event includes small fluctuations, medium fluctuations, and large fluctuations categorized by power fluctuation amplitude. The process of identifying at least one low-output wind power event corresponding to the first sample set and determining the label of the at least one low-output wind power event, constructing the first label set, includes: Wind power low output events are identified based on the total power data included in each sample in the first sample set. The wind power low output event corresponding to the first sample whose total power data is less than the first set threshold is determined as the small fluctuation. The wind power low output event corresponding to the second sample whose total power data is greater than the first set threshold and less than the second set threshold is determined as the medium amplitude fluctuation. The wind power low output event corresponding to the third sample whose total power data is greater than the second set threshold is identified as the large fluctuation; The first label set is constructed by encoding the small fluctuation, the medium fluctuation, or the large fluctuation corresponding to each sample.
4. The method according to claim 1, characterized in that, The step of training a pre-built event sample generation model based on the first training set and acquired noise data, and generating a second training set based on the trained event sample generation model, includes: Construct an event sample generation model for generating low-output wind power events; The acquired noise data and the first label set included in the first training set are used as input to the generator to obtain the generated sample set output by the generator; The generated sample set and the first sample set included in the first training set are used as inputs to the discriminator to obtain the true or false judgment result of the sample output by the discriminator; When the generator and the discriminator reach an equilibrium state after adversarial training, the trained event sample generation model is obtained, and a second training set is generated based on the trained event sample generation model.
5. The method according to claim 1, characterized in that, The step of training the meta-learner based on the event prediction result and the first sample set includes: The prediction results of at least one event output by at least one submodule of the base learner are fused to obtain a fused prediction result. The meta-learner is trained based on the fusion prediction results and the first sample set.
6. A method for predicting low wind power output events, characterized in that, include: Obtain weather forecast data for a preset time period; The weather forecast data is used as input to the event prediction model to predict low wind power output events within the preset time period, and the event prediction result is obtained; wherein, the event prediction model is trained by the wind power low output event generative differential training method according to any one of claims 1-5.