Training method and device of ultra-short-term power prediction model, equipment and medium

Abstract

The present disclosure relates to a training method and device of an ultra-short-term power prediction model, equipment and a medium. The method comprises: obtaining an incoming flow numerical weather prediction wind speed sequence, an incoming flow historical measured power sequence, a historical numerical weather prediction wind speed sequence, a historical database measured power sequence and a historical incoming flow predicted power; based on the incoming flow numerical weather prediction wind speed sequence, the incoming flow historical measured power sequence and the historical numerical weather prediction wind speed sequence, selecting sample historical measured power for training from the historical database measured power sequence; training an initial model using the sample historical measured power and the historical incoming flow predicted power until the current model under the current training times satisfies the training termination condition, obtaining an ultra-short-term power prediction model of a target wind farm. In this way, the training samples are selected from long-term data in combination with short-term data, and the ultra-short-term power prediction model is trained based on the samples, which improves the ultra-short-term power prediction accuracy and solves the concept drift problem.

Description

Technical Field

[0001] This disclosure relates to the field of renewable energy power generation technology, and in particular to a training method, apparatus, equipment and medium for an ultra-short-term power prediction model. Background Technology

[0002] Ultra-short-term wind power forecasting enables power prediction for short-term future periods by mining the time-series dependencies in the power sequence. It can provide future output information for real-time scheduling, frequency regulation, and intraday market trading.

[0003] Currently, due to changes in local weather systems, aging of generator units, and changes in control strategies, the temporal dependencies between power sequences may change, a phenomenon known as concept drift. This causes the prediction accuracy of power prediction models trained on historical data to decrease over time.

[0004] Therefore, in the face of ultra-short-term power prediction for wind farms, there is an urgent need to provide a training method for ultra-short-term power prediction models to improve the accuracy of ultra-short-term power prediction for wind farms. Summary of the Invention

[0005] To address the aforementioned technical problems, this disclosure provides a training method, apparatus, device, and medium for an ultra-short-term power prediction model.

[0006] Firstly, this disclosure provides a training method for an ultra-short-term power prediction model, including:

[0007] The system acquires the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, and the historical incoming flow predicted power. The incoming flow numerical weather forecast wind speed sequence and the incoming flow historical measured power sequence are historical forecast data of the target wind farm within a first historical time period before the historical moment. The historical incoming flow predicted power is the predicted power data of the target wind farm within a first historical time period after the historical moment. The historical numerical weather forecast wind speed sequence and the historical database measured power sequence are historical forecast data of the target wind farm within a second historical time period before the historical moment. Furthermore, the first historical time period is shorter than the second historical time period, and the first historical time period is shorter than a preset time threshold.

[0008] Based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence, sample historical measured power is selected from the historical database measured power sequence for training.

[0009] The initial model is trained using the historical measured power of the sample and the historical predicted power of the incoming flow until the current model at the current training iteration meets the training termination condition, thus obtaining the ultra-short-term power prediction model of the target wind farm.

[0010] Secondly, this disclosure provides a training device for an ultra-short-term power prediction model, comprising:

[0011] The acquisition module is used to acquire the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, and the historical incoming flow predicted power. The incoming flow numerical weather forecast wind speed sequence and the incoming flow historical measured power sequence are historical forecast data of the target wind farm within a first historical time period before the historical moment. The historical incoming flow predicted power is the predicted power data of the target wind farm within a first historical time period after the historical moment. The historical numerical weather forecast wind speed sequence and the historical database measured power sequence are historical forecast data of the target wind farm within a second historical time period before the historical moment. Furthermore, the first historical time period is shorter than the second historical time period, and the first historical time period is shorter than a preset time threshold.

[0012] The determination module is used to select sample historical measured power from the historical database measured power sequence for training based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence.

[0013] The model training module is used to train an initial model using the historical measured power of the sample and the historical predicted power of the incoming flow until the current model at the current training iteration meets the training termination condition, thereby obtaining the ultra-short-term power prediction model of the target wind farm.

[0014] Thirdly, embodiments of this disclosure also provide an electronic device, including:

[0015] processor;

[0016] Memory, used to store executable instructions;

[0017] The processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the method described in the first aspect.

[0018] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method provided in the first aspect.

[0019] The technical solution provided in this disclosure has the following advantages compared with the prior art:

[0020] This disclosure discloses a training method, apparatus, device, and medium for an ultra-short-term power prediction model, comprising: acquiring incoming flow numerical weather forecast wind speed sequences, historical measured power sequences of incoming flows, historical numerical weather forecast wind speed sequences, historical measured power sequences from a historical database, and historical incoming flow predicted power. The incoming flow numerical weather forecast wind speed sequences and historical measured power sequences are historical forecast data of the target wind farm within a first historical time period prior to the historical moment, while the historical incoming flow predicted power is the predicted power data of the target wind farm within a first historical time period after the historical moment. The historical numerical weather forecast wind speed sequences and historical data... The measured power sequence in the database is the historical forecast data of the target wind farm within the second historical time period before the historical moment. The first historical time period is shorter than the second historical time period and the first historical time period is shorter than a preset time threshold. Based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence, sample historical measured power is selected from the measured power sequence in the historical database for training. The initial model is trained using the sample historical measured power and the historical incoming flow predicted power until the current model meets the training termination condition at the current training iteration, thus obtaining the ultra-short-term power prediction model of the target wind farm. Since the data volume of short-term incoming flow numerical weather forecast wind speed sequences and historical measured power sequences before a historical time is relatively small, but the data volume of long-term historical numerical weather forecast wind speed sequences and historical database measured power sequences before a historical time is relatively large, it is possible to combine short-term historical data with long-term historical data to select training samples. Specifically, this involves combining incoming flow numerical weather forecast wind speed sequences, historical measured power sequences, and historical numerical weather forecast wind speed sequences, and selecting historical measured power samples from the historical database measured power sequences with a large data volume. This enables the training of an ultra-short-term power prediction model based on the historical measured power samples and historical incoming flow predicted power, thereby improving the accuracy of ultra-short-term power prediction and solving the concept drift problem. Attached Figure Description

[0021] 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.

[0022] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a training method for an ultra-short-term power prediction model provided in this embodiment of the disclosure;

[0024] Figure 2 A schematic flowchart of S120 provided in an embodiment of this disclosure;

[0025] Figure 3 A schematic diagram of the structure of a training device for an ultra-short-term power prediction model provided in an embodiment of this disclosure;

[0026] Figure 4 This is a schematic diagram of the structure of a training system for an ultra-short-term power prediction model provided in an embodiment of this disclosure. Detailed Implementation

[0027] 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.

[0028] 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.

[0029] In related technologies, wind speed and power data from a relatively long historical period are typically used to train ultra-short-term power prediction models. However, due to changes in local weather systems, generator aging, and changes in control strategies, the temporal dependencies between power sequences change, resulting in concept drift. Consequently, the prediction accuracy of offline models decreases over time.

[0030] To address the aforementioned issues, this disclosure provides a training method, apparatus, device, and medium for an ultra-short-term power prediction model.

[0031] The following is combined with Figure 1 The training method for the ultra-short-term power prediction model provided in this disclosure is described. In this disclosure, the training method for the ultra-short-term power prediction model can be executed by an electronic device or a server. The electronic device may include devices with communication capabilities such as tablet computers, desktop computers, and laptop computers, or devices simulated by virtual machines or simulators. The server may include server clusters and cloud servers. Below, the training method for the ultra-short-term power prediction model will be explained in detail, using an electronic device as the execution subject.

[0032] Figure 1 A flowchart illustrating a training method for an ultra-short-term power prediction model provided in an embodiment of this disclosure is shown.

[0033] like Figure 1As shown, the training method for this ultra-short-term power prediction model may include the following steps.

[0034] S110. Obtain the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, and the historical incoming flow predicted power. Among them, the incoming flow numerical weather forecast wind speed sequence and the incoming flow historical measured power sequence are the historical forecast data of the target wind farm within the first historical time period before the historical moment. The historical incoming flow predicted power is the predicted power data of the target wind farm within the first historical time period after the historical moment. The historical numerical weather forecast wind speed sequence and the historical database measured power sequence are the historical forecast data of the target wind farm within the second historical time period before the historical moment. Furthermore, the first historical time period is shorter than the second historical time period, and the first historical time period is shorter than a preset time threshold.

[0035] In this embodiment, the electronic device acquires wind speed and power data over a relatively long period of time, namely, historical numerical weather forecast wind speed sequences and historical measured power sequences from the database. At the same time, it acquires wind speed and power data over a short period of time, namely, incoming flow numerical weather forecast wind speed sequences, historical measured power sequences of incoming flows, and historical predicted power of incoming flows. This allows the electronic device to combine historical short-term data with historical long-term data to select training samples.

[0036] Understandably, since the first historical time period is shorter than the second historical time period, and the first historical time period is shorter than a preset time threshold, the incoming flow numerical weather forecast wind speed sequence represents the past short-term incoming flow forecast wind speed, the historical incoming flow measured power sequence represents the past short-term measured power, the historical incoming flow predicted power represents the past short-term predicted power, the historical numerical weather forecast wind speed sequence represents the past long-term forecast wind speed, and the historical database measured power sequence represents the past long-term measured power. A historical moment refers to any moment before the current moment. Electronic devices train ultra-short-term power prediction models by collecting short- and long-term data from historical moments.

[0037] For example, the historical moment when training the ultra-short-term power prediction model is April 15, 2024, and the preset time threshold is 15 days. The first historical time period can be the three days or the week before April 15, 2024, and the second historical time period can be the month or the three months before April 15, 2024.

[0038] Optionally, the preset time threshold can be a time parameter in hours. The first historical time period can be understood as the forecast period, specifically a time period of 1 hour, 4 hours, 6 hours, etc., before the historical moment. The second historical time period can be understood as the review window, specifically a time period of 7 days, 15 days, 30 days, 90 days, etc., before the historical moment.

[0039] The target wind farm refers to any wind farm that requires ultra-short-term power prediction.

[0040] S120. Based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence, select sample historical measured power from the historical database measured power sequence for training.

[0041] In this embodiment, the electronic device combines the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence. It selects historical measured power samples for training from the large historical database of measured power sequences, thereby mining a large number of training samples similar to past short-term data from long-term data to train the ultra-short-term power prediction model. This improves the timeliness of ultra-short-term power prediction and solves the concept drift problem.

[0042] Specifically, the electronic device performs feature matching on the historical numerical weather forecast wind speed sequence and the incoming flow numerical weather forecast wind speed sequence, and performs feature matching on the historical database measured power sequence and the historical incoming flow measured power sequence. Based on the matching degree, it selects sample historical measured power from the historical database measured power sequence for training.

[0043] S130. Train the initial model using the historical measured power and historical incoming flow predicted power of the sample until the current model meets the training termination condition under the current training number, and obtain the ultra-short-term power prediction model of the target wind farm.

[0044] In this embodiment, the electronic device performs online training based on the historical measured power and historical predicted power of the incoming flow. During the training process, iteratively determines whether the current model under the current training number meets the training termination condition. If the training termination condition is met, the current model under the current training number is used as the ultra-short-term power forecast model of the target wind farm, thereby better dealing with various scenarios such as repeated drift and continuous drift.

[0045] The training termination conditions include one or more preset conditions corresponding to the initial model's loss value, the change in loss value, and the learning rate. Accordingly, the specific implementation methods of S130 include, but are not limited to, the following:

[0046] The initial model is used to process the historical measured power of the sample to obtain the model predicted power; the current learning rate of the initial model is obtained, and the second loss data of the initial model is calculated based on the historical incoming flow predicted power and the model predicted power; the initial model is trained online based on the current learning rate and the second loss data until at least one of the second loss data, the change in the second loss data, and the current learning rate satisfies the corresponding preset condition under the current training number. Then, the current model under the current training number is determined to meet the training termination condition, and the ultra-short-term power prediction model of the target wind farm is obtained.

[0047] To balance online training time and effectiveness, an "early termination mechanism" is employed to control the number of training iterations. This mechanism involves real-time monitoring of model performance during online training, such as calculating loss data, changes in loss data, and the learning rate. If the loss data, changes in loss data, and the learning rate are all within threshold ranges, it indicates that the model meets preset conditions for the current training iterations, triggering the "early termination mechanism" to stop further training. This effectively avoids overfitting or underfitting and addresses the issue of the inability to adjust the number of online modeling iterations.

[0048] The learning rate is a crucial parameter affecting the model training process. Setting it too high can cause the loss data to exceed the global optimum; setting it too low will result in slow changes in the loss data, increasing the model's convergence complexity and making it prone to getting stuck in local minima or saddle points. Specifically, this embodiment introduces a callback function during training to monitor the second loss data in real time. As the second loss data continuously changes, the learning rate is adjusted adaptively.

[0049] In this way, by monitoring at least one of the second loss data, changes in the second loss data, and the current learning rate, the model training process can be controlled to improve the training accuracy of the ultra-short-term power prediction model.

[0050] This disclosure provides a training method for an ultra-short-term power prediction model, which acquires incoming flow numerical weather forecast wind speed sequences, historical measured power sequences, historical numerical weather forecast wind speed sequences, historical database measured power sequences, and historical incoming flow predicted power. The incoming flow numerical weather forecast wind speed sequences and historical measured power sequences are historical forecast data for the target wind farm within a first historical time period before the historical moment. The historical incoming flow predicted power is the predicted power data for the target wind farm within a first historical time period after the historical moment. The historical numerical weather forecast wind speed sequences and historical database measured power sequences... The sequence consists of historical forecast data for the target wind farm within the second historical time period prior to the historical moment, where the first historical time period is shorter than the second historical time period and the first historical time period is shorter than a preset time threshold. Based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence, sample historical measured power is selected from the historical database measured power sequence for training. The initial model is trained using the sample historical measured power and the historical incoming flow predicted power until the current model at the current training iteration meets the training termination condition, thus obtaining the ultra-short-term power prediction model for the target wind farm. Since the data volume of short-term incoming flow numerical weather forecast wind speed sequences and historical measured power sequences before a historical time is relatively small, but the data volume of long-term historical numerical weather forecast wind speed sequences and historical database measured power sequences before a historical time is relatively large, it is possible to combine short-term historical data with long-term historical data to select training samples. Specifically, this involves combining incoming flow numerical weather forecast wind speed sequences, historical measured power sequences, and historical numerical weather forecast wind speed sequences, and selecting historical measured power samples from the historical database measured power sequences with a large data volume. This enables the training of an ultra-short-term power prediction model based on the historical measured power samples and historical incoming flow predicted power, thereby improving the accuracy of ultra-short-term power prediction and solving the concept drift problem.

[0051] Furthermore, after the ultra-short-term power prediction model is trained online, the electronic device can use the ultra-short-term power prediction model to determine the ultra-short-term predicted power. The method further includes: using the ultra-short-term power prediction model of the target wind farm, performing power prediction processing on the measured incoming power within the first historical time period before the current moment, and determining the incoming predicted power of the target wind farm within the first future time period corresponding to the first historical time period after the current moment as the ultra-short-term predicted power.

[0052] For example, if the historical time when training the ultra-short-term power prediction model is April 15, 2024, and the current time when using the ultra-short-term power prediction model is May 15, 2024, and the first historical time period is the three days and one week before April 15, 2024, then the first future time period is the three days and one week after May 15, 2024.

[0053] Since the first historical time period is a short-term period before the current moment, the first future time period is a short-term period after the current moment. Thus, by using a highly accurate ultra-short-term power prediction model, we can accurately predict future ultra-short-term power data, which can then be used as the ultra-short-term predicted power within the first future time period, thereby effectively improving the accuracy of ultra-short-term power prediction.

[0054] In another embodiment of this disclosure, S120 will be explained in detail.

[0055] Figure 2 A flowchart of S120 provided in an embodiment of this disclosure is shown.

[0056] like Figure 2 As shown, S120 may specifically include the following steps.

[0057] S210. Perform deep fluctuation feature extraction on the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence, respectively, to determine the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the incoming flow historical measured power sequence.

[0058] To fully depict the fluctuation characteristics within the sequence, the electronic device can encode the wind speed sequence and power sequence into a two-dimensional image, and then extract more comprehensive fluctuation characteristics from the two-dimensional image, while also preserving temporal features. Accordingly, the specific implementation methods of S210 include, but are not limited to, the following:

[0059] S1. Gram angle field coding is performed on the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence to generate coding features corresponding to the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence.

[0060] S2. Deep fluctuation feature extraction is performed on the coding features corresponding to the historical numerical weather forecast wind speed sequence, the coding features corresponding to the historical database measured power sequence, the coding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the coding features corresponding to the historical measured power sequence of the incoming flow, respectively, to obtain the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the historical measured power sequence of the incoming flow.

[0061] To eliminate noise in the wind speed and power sequences and improve the efficiency of fluctuation feature extraction, before executing S1, the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence can be cleaned separately. Then, according to the first historical time period, the cleaned incoming flow numerical weather forecast wind speed sequence and the incoming flow historical measured power sequence are divided into multiple sequence segments, and according to the second historical time period, the cleaned historical numerical weather forecast wind speed sequence and the historical database measured power sequence are divided into multiple sequence segments. Then, Gram angle field coding is performed on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, the historical numerical weather forecast wind speed sequence, and the historical database measured power sequence containing multiple sequence segments, generating the coding features corresponding to the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence.

[0062] The specific implementation method of S2 includes, but is not limited to, the following: using a portion of the coding features corresponding to the historical numerical weather forecast wind speed sequence, the coding features corresponding to the historical database measured power sequence, the coding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the coding features corresponding to the historical measured power sequence of the incoming flow, an unsupervised training is performed on the initial convolutional neural network to obtain the target convolutional neural network; based on the convolutional module in the target convolutional neural network, deep fluctuation feature extraction is performed on the coding features corresponding to the historical numerical weather forecast wind speed sequence, the coding features corresponding to the historical database measured power sequence, the coding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the coding features corresponding to the historical measured power sequence of the incoming flow, respectively, to obtain the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the historical measured power sequence of the incoming flow.

[0063] The initial convolutional neural network (CNN) includes a convolutional module, a clustering module, a fully connected network, and a classification network. Specifically, the method for determining the target CNN includes: extracting features from a subset of features using the convolutional module in the initial CNN to obtain feature vectors corresponding to these features; clustering these feature vectors using the clustering module in the initial CNN to obtain first-class labels; classifying these feature vectors using the fully connected network and the classification network in the initial CNN to obtain second-class labels; and updating the parameters of the initial CNN using the first-class and second-class labels of the features until the first loss data determined by the updated first-class and second-class labels of the features is less than a preset loss threshold, thus obtaining the target CNN.

[0064] Specifically, the clustering module in the initial convolutional neural network can use clustering methods such as K-means to determine the first-class labels corresponding to a subset of features. The first-class labels refer to the predicted labels of a subset of features, while the second-class labels refer to the original labels of a subset of features. The first loss data can be calculated using methods such as the cross-entropy loss function. Then, based on the first loss data, the initial convolutional neural network is trained in reverse to update the parameters of the convolutional modules in the initial convolutional neural network until the first loss data is less than a preset loss threshold, thus obtaining the target convolutional neural network.

[0065] Furthermore, by utilizing the convolutional module in the target convolutional neural network, deep fluctuation feature extraction is performed on the coding features corresponding to the historical numerical weather forecast wind speed sequence, the coding features corresponding to the historical database measured power sequence, the coding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the coding features corresponding to the historical measured power sequence of the incoming flow, respectively, to obtain the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the historical measured power sequence of the incoming flow.

[0066] In this way, the Gram angle field coding and unsupervised image feature extraction method can extract more comprehensive power fluctuation features and wind speed fluctuation features, while also preserving temporal features.

[0067] S220. Based on the fluctuation characteristics of the historical numerical weather forecast wind speed sequence, the amplitude characteristics of the historical database measured power sequence, the fluctuation characteristics of the incoming flow numerical weather forecast wind speed sequence, and the amplitude characteristics of the incoming flow historical measured power sequence, select sample historical measured power from the historical database measured power sequence.

[0068] In this embodiment, the specific implementation method of S220 includes, but is not limited to, the following: constructing a first feature matrix based on the fluctuation characteristics corresponding to the historical numerical weather forecast wind speed sequence and the amplitude characteristics corresponding to the historical measured power sequence in the database, and constructing a second feature matrix based on the fluctuation characteristics corresponding to the incoming flow numerical weather forecast wind speed sequence and the amplitude characteristics corresponding to the historical measured power sequence in the incoming flow; matching the first feature matrix and the second feature matrix to determine the distance between matching point pairs in the first feature matrix and the second feature matrix; selecting a target amplitude feature with a distance less than a preset value from the fluctuation characteristics corresponding to the historical numerical weather forecast wind speed sequence and the amplitude characteristics corresponding to the historical measured power sequence in the database contained in the first feature matrix; obtaining the historical measured power sequence in the database corresponding to the target amplitude feature as the sample historical measured power.

[0069] Specifically, firstly, the fluctuation characteristics corresponding to the historical numerical weather forecast wind speed sequence and the amplitude characteristics corresponding to the historical measured power sequence are concatenated by rows or columns to construct a first feature matrix. Similarly, the fluctuation characteristics corresponding to the incoming flow numerical weather forecast wind speed sequence and the amplitude characteristics corresponding to the incoming flow historical measured power sequence are concatenated by rows or columns to construct a second feature matrix. Then, the first and second feature matrices are matched, and the distance between matching point pairs in the first and second feature matrices is calculated using methods such as Euclidean distance. Next, the distances between all matching point pairs are sorted, and the target amplitude characteristics with distances less than a preset value are selected. Finally, the historical measured power sequence corresponding to the target amplitude characteristics is used as the sample historical measured power.

[0070] In this way, since two-dimensional images contain more comprehensive feature information, by performing targeted matching of features, sample historical measured power with similar features to the historical measured power sequence of incoming flow can be selected from the historical measured power sequence of the large historical database to increase the representativeness of the prediction scenario. Therefore, based on the sample historical measured power and the historical predicted power of incoming flow, the ultra-short-term power prediction model can be trained accurately.

[0071] This disclosure also provides a training apparatus for an ultra-short-term power prediction model to implement the above-described training method for the ultra-short-term power prediction model. The following is in conjunction with... Figure 3 The following explanation is provided. In this embodiment, the training device for the ultra-short-term power prediction model is configured on an electronic device or a server. The electronic device may include devices with communication capabilities such as tablets, desktop computers, and laptops, or devices simulated by virtual machines or simulators. The server may include server clusters and cloud servers.

[0072] Figure 3A schematic diagram of the structure of a training device for an ultra-short-term power prediction model provided in an embodiment of this disclosure is shown.

[0073] like Figure 3 As shown, the training device 300 for the ultra-short-term power prediction model may include:

[0074] The acquisition module 310 is used to acquire the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, and the historical incoming flow predicted power. The incoming flow numerical weather forecast wind speed sequence and the incoming flow historical measured power sequence are historical forecast data of the target wind farm within a first historical time period before the historical moment. The historical incoming flow predicted power is the predicted power data of the target wind farm within a first historical time period after the historical moment. The historical numerical weather forecast wind speed sequence and the historical database measured power sequence are historical forecast data of the target wind farm within a second historical time period before the historical moment. Furthermore, the first historical time period is shorter than the second historical time period, and the first historical time period is shorter than a preset time threshold.

[0075] The determination module 320 is used to select sample historical measured power from the historical database measured power sequence for training based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence and the historical numerical weather forecast wind speed sequence;

[0076] The model training module 330 is used to train an initial model using the historical measured power of the sample and the historical predicted power of the incoming flow until the current model at the current training number meets the training termination condition, thereby obtaining the ultra-short-term power prediction model of the target wind farm.

[0077] This disclosure discloses a training device for an ultra-short-term power prediction model, which acquires incoming flow numerical weather forecast wind speed sequences, historical measured power sequences, historical numerical weather forecast wind speed sequences, historical database measured power sequences, and historical incoming flow predicted power. The incoming flow numerical weather forecast wind speed sequences and historical measured power sequences are historical forecast data for the target wind farm within a first historical time period before the historical moment. The historical incoming flow predicted power is the predicted power data for the target wind farm within a first historical time period after the historical moment. The historical numerical weather forecast wind speed sequences and historical database measured power sequences... The system uses historical forecast data for the target wind farm within the second historical time period prior to the historical moment, where the first historical time period is shorter than the second historical time period and the first historical time period is shorter than a preset time threshold. Based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence, the system selects sample historical measured power from the historical database measured power sequence for training. The system uses the sample historical measured power and the historical incoming flow predicted power to train the initial model until the current model meets the training termination condition at the current training iteration, thus obtaining the ultra-short-term power prediction model for the target wind farm. Since the data volume of short-term incoming flow numerical weather forecast wind speed sequences and historical measured power sequences before a historical time is relatively small, but the data volume of long-term historical numerical weather forecast wind speed sequences and historical database measured power sequences before a historical time is relatively large, it is possible to combine short-term historical data with long-term historical data to select training samples. Specifically, this involves combining incoming flow numerical weather forecast wind speed sequences, historical measured power sequences, and historical numerical weather forecast wind speed sequences, and selecting historical measured power samples from the historical database measured power sequences with a large data volume. This enables the training of an ultra-short-term power prediction model based on the historical measured power samples and historical incoming flow predicted power, thereby improving the accuracy of ultra-short-term power prediction and solving the concept drift problem.

[0078] In some embodiments of this disclosure, the determining module 320 includes:

[0079] The feature extraction unit is used to perform deep fluctuation feature extraction on the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence, respectively, to determine the fluctuation feature corresponding to the historical numerical weather forecast wind speed sequence, the amplitude feature corresponding to the historical database measured power sequence, the fluctuation feature corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude feature corresponding to the incoming flow historical measured power sequence;

[0080] The sample selection unit is used to select the sample historical measured power from the historical measured power sequence based on the fluctuation characteristics corresponding to the historical numerical weather forecast wind speed sequence, the amplitude characteristics corresponding to the historical measured power sequence, the fluctuation characteristics corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude characteristics corresponding to the historical measured power sequence.

[0081] In some embodiments of this disclosure, the feature extraction unit includes:

[0082] The coding subunit is used to perform Gram angle field coding on the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence, respectively, to generate coding features corresponding to the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence;

[0083] The feature extraction subunit is used to perform deep fluctuation feature extraction on the encoded features corresponding to the historical numerical weather forecast wind speed sequence, the encoded features corresponding to the historical database measured power sequence, the encoded features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the encoded features corresponding to the historical measured power sequence of the incoming flow, respectively, to obtain the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the historical measured power sequence of the incoming flow.

[0084] In some embodiments of this disclosure, the feature extraction subunit is used for:

[0085] Using a subset of the coding features corresponding to the historical numerical weather forecast wind speed sequence, the coding features corresponding to the historical database measured power sequence, the coding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the coding features corresponding to the historical measured power sequence of the incoming flow, an unsupervised training is performed on the initial convolutional neural network to obtain the target convolutional neural network.

[0086] Based on the convolutional module in the target convolutional neural network, deep fluctuation feature extraction is performed on the encoding features corresponding to the historical numerical weather forecast wind speed sequence, the encoding features corresponding to the historical database measured power sequence, the encoding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the encoding features corresponding to the historical measured power sequence of the incoming flow, respectively, to obtain the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the historical measured power sequence of the incoming flow.

[0087] In some embodiments of this disclosure, the feature extraction subunit is specifically used for:

[0088] Based on the convolutional module in the initial convolutional neural network, feature extraction is performed on the partial features to obtain the feature vector corresponding to the partial features;

[0089] Based on the clustering module in the initial convolutional neural network, the feature vectors corresponding to a subset of features are clustered to obtain the first class label corresponding to a subset of features.

[0090] Based on the fully connected network and classification network in the initial convolutional neural network, the feature vectors corresponding to the partial features are classified to obtain the second category label corresponding to the partial features.

[0091] Using the first type of label corresponding to the partial features and the second type of label corresponding to the partial features, the parameters of the initial convolutional neural network are updated until the first loss data determined based on the first type of label corresponding to the partial features and the second type of label corresponding to the partial features after the update is less than a preset loss threshold, thus obtaining the target convolutional neural network.

[0092] In some embodiments of this disclosure, the sample selection unit includes:

[0093] A sub-unit is constructed to construct a first feature matrix based on the fluctuation characteristics corresponding to the historical numerical weather forecast wind speed sequence and the amplitude characteristics corresponding to the historical measured power sequence in the database, and to construct a second feature matrix based on the fluctuation characteristics corresponding to the incoming flow numerical weather forecast wind speed sequence and the amplitude characteristics corresponding to the historical measured power sequence in the incoming flow.

[0094] The distance calculation subunit is used to match the first feature matrix and the second feature matrix to determine the distance between matching point pairs in the first feature matrix and the second feature matrix;

[0095] The feature selection subunit is used to select the target amplitude feature whose distance is less than a preset value from the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence contained in the first feature matrix and the amplitude features corresponding to the historical database measured power sequence.

[0096] The sample sub-selection unit is used to obtain the historical measured power sequence corresponding to the target amplitude feature, and use it as the historical measured power of the sample.

[0097] In some embodiments of this disclosure, the model training module 330 includes:

[0098] The first processing unit is used to process the historical measured power of the sample using the initial model to obtain the model predicted power;

[0099] The loss calculation unit is used to obtain the current learning rate of the initial model and calculate the second loss data of the initial model based on the historical incoming flow prediction power and the model prediction power.

[0100] An iterative training unit is used to train the initial model online based on the current learning rate and the second loss data until at least one of the second loss data, the change in the second loss data, and the current learning rate satisfies the corresponding preset condition under the current training number. Then, it is determined that the current model under the current training number satisfies the training termination condition, and the ultra-short-term power prediction model of the target wind farm is obtained.

[0101] In some embodiments of this disclosure, the device further includes:

[0102] The short-term predicted power module is used to perform power prediction processing on the measured incoming power of the target wind farm in the first historical time period before the current moment using the ultra-short-term power prediction model of the target wind farm, and to determine the incoming predicted power of the target wind farm in the first future time period corresponding to the first historical time period after the current moment as the ultra-short-term predicted power.

[0103] It should be noted that, Figure 3 The training device 300 for the ultra-short-term power prediction model shown can perform... Figures 1-2 The various steps in the method embodiment shown are implemented. Figures 1-2 The processes and effects in the method embodiments shown are not described in detail here.

[0104] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure is shown.

[0105] like Figure 4 As shown, the electronic device may include a processor 401 and a memory 402 storing computer program instructions.

[0106] Specifically, the processor 401 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0107] Memory 402 may include a large-capacity storage for information or instructions. For example, and not limitingly, memory 402 may include a hard disk drive (HDD), a floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 402 may include removable or non-removable (or fixed) media. Where appropriate, memory 402 may be internal or external to the integrated gateway device. In a particular embodiment, memory 402 is a non-volatile solid-state memory. In a particular embodiment, memory 402 includes read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (Electrically Programmable ROM, EPROM), an electrically erasable programmable PROM (EEPROM), an electrically alterable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0108] The processor 401 reads and executes computer program instructions stored in the memory 402 to perform the steps of the training method for the ultra-short-term power prediction model provided in this embodiment of the disclosure.

[0109] In one example, the electronic device may also include a transceiver 403 and a bus 404. Wherein, as... Figure 4 As shown, the processor 401, memory 402 and transceiver 403 are connected via bus 404 and communicate with each other.

[0110] Bus 404 includes hardware, software, or both. For example, and not limitingly, a bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industrial Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a MicroChannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 404 may include one or more buses. Although specific buses are described and illustrated in the embodiments of this application, this application considers any suitable bus or interconnection.

[0111] Of course, the computer-executable instructions provided in the embodiments of this disclosure are not limited to the above-described method operations, but can also perform related operations of the training method of the ultra-short-term power prediction model provided in any embodiment of this disclosure.

[0112] Based on the above description of the implementation methods, those skilled in the art can clearly understand that this disclosure can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer cloud platform (which may be a personal computer, server, or network cloud platform, etc.) to execute the training method of the ultra-short-term power prediction model provided in the various embodiments of this disclosure.

[0113] Note that the above description is merely a preferred embodiment and the technical principles employed in this disclosure. Those skilled in the art will understand that this disclosure is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of this disclosure. Therefore, although this disclosure has been described in detail through the above embodiments, it is not limited to the above embodiments. Many other equivalent embodiments may be included without departing from the concept of this disclosure, and the scope of this disclosure is determined by the scope of the appended claims.

Claims

1. A training method for an ultra-short-term power prediction model, characterized in that, include: The system acquires the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, and the historical incoming flow predicted power. The incoming flow numerical weather forecast wind speed sequence and the incoming flow historical measured power sequence are historical forecast data of the target wind farm within a first historical time period before the historical moment. The historical incoming flow predicted power is the predicted power data of the target wind farm within a first historical time period after the historical moment. The historical numerical weather forecast wind speed sequence and the historical database measured power sequence are historical forecast data of the target wind farm within a second historical time period before the historical moment. Furthermore, the first historical time period is shorter than the second historical time period, and the first historical time period is shorter than a preset time threshold. Based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence, sample historical measured power is selected from the historical database measured power sequence for training. The initial model is trained using the historical measured power of the sample and the historical predicted power of the incoming flow until the current model at the current training iteration meets the training termination condition, thus obtaining the ultra-short-term power prediction model of the target wind farm. The step of selecting sample historical measured power from the historical database measured power sequence for training, based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence, includes: Deep fluctuation feature extraction is performed on the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence to determine the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the incoming flow historical measured power sequence. Based on the fluctuation characteristics of the historical numerical weather forecast wind speed sequence, the amplitude characteristics of the historical database measured power sequence, the fluctuation characteristics of the incoming flow numerical weather forecast wind speed sequence, and the amplitude characteristics of the incoming flow historical measured power sequence, the sample historical measured power is selected from the historical database measured power sequence.

2. The method according to claim 1, characterized in that, The process of extracting deep fluctuation features from the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence to determine the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the incoming flow historical measured power sequence includes: Gram angle field coding is performed on the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence to generate coding features corresponding to the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence. Deep fluctuation feature extraction is performed on the coding features corresponding to the historical numerical weather forecast wind speed sequence, the coding features corresponding to the historical database measured power sequence, the coding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the coding features corresponding to the historical measured power sequence of the incoming flow, respectively, to obtain the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the historical measured power sequence of the incoming flow.

3. The method according to claim 2, characterized in that, The deep fluctuation feature extraction is performed on the coding features corresponding to the historical numerical weather forecast wind speed sequence, the coding features corresponding to the historical database measured power sequence, the coding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the coding features corresponding to the historical measured power sequence of the incoming flow, respectively, to obtain the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the historical measured power sequence of the incoming flow, including: Using a subset of the coding features corresponding to the historical numerical weather forecast wind speed sequence, the coding features corresponding to the historical database measured power sequence, the coding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the coding features corresponding to the historical measured power sequence of the incoming flow, an unsupervised training is performed on the initial convolutional neural network to obtain the target convolutional neural network. Based on the convolutional module in the target convolutional neural network, deep fluctuation feature extraction is performed on the encoding features corresponding to the historical numerical weather forecast wind speed sequence, the encoding features corresponding to the historical database measured power sequence, the encoding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the encoding features corresponding to the historical measured power sequence of the incoming flow, respectively, to obtain the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the historical measured power sequence of the incoming flow.

4. The method according to claim 3, characterized in that, The method of using a subset of the coding features corresponding to the historical numerical weather forecast wind speed sequence, the coding features corresponding to the historical measured power sequence, the coding features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the coding features corresponding to the historical measured power sequence of the incoming flow to perform unsupervised training on the initial convolutional neural network to obtain the target convolutional neural network includes: Based on the convolutional module in the initial convolutional neural network, feature extraction is performed on the partial features to obtain the feature vector corresponding to the partial features; Based on the clustering module in the initial convolutional neural network, the feature vectors corresponding to a subset of features are clustered to obtain the first class label corresponding to a subset of features. Based on the fully connected network and classification network in the initial convolutional neural network, the feature vectors corresponding to the partial features are classified to obtain the second category label corresponding to the partial features. Using the first type of label corresponding to the partial features and the second type of label corresponding to the partial features, the parameters of the initial convolutional neural network are updated until the first loss data determined based on the first type of label corresponding to the partial features and the second type of label corresponding to the partial features after the update is less than a preset loss threshold, thus obtaining the target convolutional neural network.

5. The method according to claim 1, characterized in that, The step of selecting the sample historical measured power from the historical measured power sequence based on the fluctuation characteristics corresponding to the historical numerical weather forecast wind speed sequence, the amplitude characteristics corresponding to the historical measured power sequence, the fluctuation characteristics corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude characteristics corresponding to the historical measured power sequence includes: A first feature matrix is ​​constructed based on the fluctuation characteristics of the historical numerical weather forecast wind speed sequence and the amplitude characteristics of the historical measured power sequence in the database; and a second feature matrix is ​​constructed based on the fluctuation characteristics of the incoming flow numerical weather forecast wind speed sequence and the amplitude characteristics of the incoming flow historical measured power sequence. Match the first feature matrix and the second feature matrix to determine the distance between matching point pairs in the first feature matrix and the second feature matrix; From the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence and the amplitude features corresponding to the historical database measured power sequence contained in the first feature matrix, select the target amplitude feature whose distance is less than a preset value. Obtain the historical measured power sequence corresponding to the target amplitude feature from the database, and use it as the sample historical measured power.

6. The method according to claim 1, characterized in that, The process of training an initial model using the historical measured power of the sample and the historical predicted power of the incoming flow until the current model at the current training iteration meets the training termination condition, thereby obtaining the ultra-short-term power prediction model for the target wind farm, includes: The historical measured power of the sample is processed using the initial model to obtain the model-predicted power; Obtain the current learning rate of the initial model, and calculate the second loss data of the initial model based on the historical incoming flow prediction power and the model prediction power; Based on the current learning rate and the second loss data, the initial model is trained online until at least one of the second loss data, the change in the second loss data, and the current learning rate satisfies the corresponding preset condition under the current training number. Then, it is determined that the current model under the current training number satisfies the training termination condition, and the ultra-short-term power prediction model of the target wind farm is obtained.

7. The method according to claim 1, characterized in that, Also includes: Using the ultra-short-term power prediction model of the target wind farm, the measured incoming power in the first historical time period before the current moment is processed for power prediction, and the incoming power prediction of the target wind farm in the first future time period corresponding to the first historical time period after the current moment is determined as the ultra-short-term predicted power.

8. A training device for an ultra-short-term power prediction model, characterized in that, include: The acquisition module is used to acquire the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, and the historical incoming flow predicted power. The incoming flow numerical weather forecast wind speed sequence and the incoming flow historical measured power sequence are historical forecast data of the target wind farm within a first historical time period before the historical moment. The historical incoming flow predicted power is the predicted power data of the target wind farm within a first historical time period after the historical moment. The historical numerical weather forecast wind speed sequence and the historical database measured power sequence are historical forecast data of the target wind farm within a second historical time period before the historical moment. Furthermore, the first historical time period is shorter than the second historical time period, and the first historical time period is shorter than a preset time threshold. The determination module is used to select sample historical measured power from the historical database measured power sequence for training based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence. The model training module is used to train an initial model using the historical measured power of the sample and the historical predicted power of the incoming flow until the current model at the current training number meets the training termination condition, thereby obtaining the ultra-short-term power prediction model of the target wind farm. The step of selecting sample historical measured power from the historical database measured power sequence for training, based on the incoming flow numerical weather forecast wind speed sequence, the incoming flow historical measured power sequence, and the historical numerical weather forecast wind speed sequence, includes: Deep fluctuation feature extraction is performed on the historical numerical weather forecast wind speed sequence, the historical database measured power sequence, the incoming flow numerical weather forecast wind speed sequence, and the incoming flow historical measured power sequence to determine the fluctuation features corresponding to the historical numerical weather forecast wind speed sequence, the amplitude features corresponding to the historical database measured power sequence, the fluctuation features corresponding to the incoming flow numerical weather forecast wind speed sequence, and the amplitude features corresponding to the incoming flow historical measured power sequence. Based on the fluctuation characteristics of the historical numerical weather forecast wind speed sequence, the amplitude characteristics of the historical database measured power sequence, the fluctuation characteristics of the incoming flow numerical weather forecast wind speed sequence, and the amplitude characteristics of the incoming flow historical measured power sequence, the sample historical measured power is selected from the historical database measured power sequence.

9. An electronic device, characterized in that, include: processor; Memory, used to store executable instructions; The processor is configured to read the executable instructions from the memory and execute the executable instructions to implement the method of any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The storage medium stores a computer program that, when executed by a processor, causes the processor to implement the method described in any one of claims 1-7.

Images