Wind power prediction method, device, equipment and storage medium

Abstract

The present application relates to a kind of wind power prediction method, device and storage medium based on time series large model.The wind power prediction method includes: obtaining time series large model and the first training data set of multiple source wind farms;The first training data set is used to train time series large model, and the basic large model for wind power prediction is obtained;The historical wind power array and future weather forecast array of target wind farm are obtained, and multiple source wind farms do not include target wind farm;The historical wind power array and future weather forecast array are used as the input of target large model, and the future wind power array of target wind farm is predicted.This application reduces the dependence of model training on training data amount to some extent, effectively solves the difference in the number of model parameters, further improves the application flexibility and effectiveness of model in variable scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer technology, and in particular to a wind power prediction method and device based on a time series large model, equipment and a storage medium. BACKGROUND

[0002] Nowadays, the wind power prediction method mainly adopts a data-driven technical route, and a high-precision wind power prediction model relies on a large amount of training data. However, for a newly-built, expanded or reconstructed wind farm, the running time is short, and it is unable to provide a sufficient amount of historical data or even no historical data for the wind power prediction model training.

[0003] At present, for small sample wind power prediction, the most commonly used method is the "pre-training-fine-tuning" method of transfer learning, that is, first, a wind power prediction model is trained through sufficient training samples of a source wind farm, and then the power prediction knowledge learned by the model is transferred to a target wind farm. However, there are certain limitations in solving this problem by using only the method of transfer learning: it requires strong "consistency" between wind farms, and needs "one target farm one pre-training", on the one hand, transfer learning is severely dependent on the sample quality and correlation of the source wind farm and the target wind farm, and sometimes it is difficult to obtain source wind farm data with high similarity due to data privacy and other problems; on the other hand, only ordinary transfer learning cannot solve the problem of inconsistent model input length, prediction scale, time resolution and prediction variable quantity between wind farms, limiting the application flexibility and effectiveness in various scenarios. SUMMARY

[0004] In order to solve the above technical problems, the present application provides a wind power prediction method and device based on a time series large model, equipment and a storage medium.

[0005] In a first aspect, the present application provides a wind power prediction method based on a time series large model, comprising:

[0006] obtaining a time series large model and a first training data set of a plurality of source wind farms;

[0007] pre-training the time series large model using the first training data set to obtain a basic large model for wind power prediction;

[0008] obtaining a historical wind power array and a future weather forecast array of a target wind farm, wherein the plurality of source wind farms do not include the target wind farm;

[0009] using the historical wind power array and the future weather forecast array as the input of a target large model to predict a future wind power array of the target wind farm, wherein the target large model is the basic large model or is obtained by fine-tuning the basic large model using a second training data set of the target wind farm.

[0010] In a second aspect, the embodiments of the present disclosure provide a wind power prediction device based on a time series large model, comprising:

[0011] a first obtaining unit, configured to obtain a time series large model and a first training data set of a plurality of source wind farms;

[0012] a training unit, configured to pre-train the time series large model by using the first training data set to obtain a basic large model for wind power prediction;

[0013] a second obtaining unit, configured to obtain a historical wind power array and a future weather forecast array of a target wind farm, wherein the plurality of source wind farms do not include the target wind farm;

[0014] a prediction unit, configured to take the historical wind power array and the future weather forecast array as inputs of a target large model to predict a future wind power array of the target wind farm, wherein the target large model is the basic large model or is obtained by fine-tuning the basic large model by using a second training data set of the target wind farm.

[0015] In a third aspect, the embodiments of the present disclosure provide an electronic device, comprising:

[0016] a memory;

[0017] a processor; and

[0018] a computer program;

[0019] wherein the computer program is stored in the memory and is configured to be executed by the processor to implement the method of the first aspect as described above.

[0020] In a fourth aspect, the embodiments of the present disclosure provide a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the method of the first aspect as described above.

[0021] The wind power prediction method provided by the disclosure comprises: acquiring a time series large model and a first training data set of a plurality of source wind farms; training the time series large model by using the first training data set to obtain a basic large model for wind power prediction; acquiring a historical wind power array and a future weather forecast array of a target wind farm, the plurality of source wind farms not including the target wind farm; taking the historical wind power array and the future weather forecast array as inputs of a target large model to predict a future wind power array of the target wind farm, wherein the target large model is the basic large model or is obtained by fine-tuning the basic large model by using a second training data set of the target wind farm. The application reduces the dependence of model training on the amount of training data to a certain extent, effectively solves the problem of differences in the number of model parameters, and further improves the application flexibility and effectiveness of the model in variable scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0022] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the disclosure and serve to explain the principles of the disclosure together with the specification.

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0024] Figure 1 A flowchart of a wind power prediction method based on a time series large model provided by an embodiment of the disclosure;

[0025] Figure 2 A structural diagram of a time series large model provided by an embodiment of the disclosure;

[0026] Figure 3 A structural diagram of a basic large model provided by an embodiment of the disclosure;

[0027] Figure 4 A flowchart of a wind power prediction method based on a target large model provided by an embodiment of the disclosure;

[0028] Figure 5 A flowchart of a wind power prediction method based on a time series large model provided by an embodiment of the disclosure;

[0029] Figure 6 A two-dimensional structural diagram of a historical wind power training sample provided by an embodiment of the disclosure;

[0030] Figure 7A two-dimensional structural diagram of a weather forecast training sample provided by an embodiment of the present disclosure;

[0031] Figure 8 A flowchart of a wind power prediction method based on a time series large model provided by an embodiment of the present disclosure;

[0032] Figure 9 A flowchart of a wind power prediction method based on a time series large model provided by an embodiment of the present disclosure;

[0033] Figure 10 A flowchart of a wind power prediction method based on a time series large model provided by an embodiment of the present disclosure;

[0034] Figure 11 A structural diagram of a wind power prediction device based on a time series large model provided by an embodiment of the present disclosure;

[0035] Figure 12 A structural diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0036] In order to enable a more clear understanding of the above-mentioned purposes, features and advantages of the present disclosure, the schemes of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0037] In the following description, many specific details are set forth in order to provide a thorough understanding of the present disclosure, but the present disclosure can also be implemented in other ways different from those described herein; obviously, the embodiments in the description are only some of the embodiments of the present disclosure, not all the embodiments.

[0038] In view of the above technical problems, the embodiments of the present disclosure provide a wind power prediction method based on a time series large model, a general wind power prediction model (the above-mentioned basic large model) is constructed, and zero sample / small sample wind power prediction is realized. When there is a prediction demand for a newly-built, expanded or reconstructed target wind farm, the wind power prediction model no longer relies on sufficient historical data of the target wind farm and a large amount of training time, but directly or after fine-tuning, the wind power prediction model can realize low-time-consumption, low-cost, high-adaptation and high-precision wind power prediction by virtue of the experience of the past training data of the wind power prediction model, and effectively solve the difference problems in the number of model input length, prediction scale, time resolution and variables between wind farms. One or more of the following embodiments are used for detailed description.

[0039] Before the method involved in the present application is described in detail, the proper nouns involved are explained first, including:

[0040] Wind farm refers to a power plant that converts wind energy into mechanical energy and then converts the mechanical energy into electrical energy. The wind farm provided by the application includes but is not limited to a wind measurement tower, a wind turbine generator, a power collection line, a booster station, etc. The wind turbine generator receives wind energy and converts it into electrical energy, which is then output through the output end of the booster station. The wind power of the wind farm is the total power output from the output end of the booster station.

[0041] Wind power prediction refers to a technology for predicting the power generation of a wind farm in a future period of time, providing a reference basis for the operation control of the wind farm, the dispatching operation of the power system, and the power market transaction.

[0042] Moirai refers to a deep learning large model for time series prediction, which is trained by the LOSTA dataset (more than 27.6 billion data, covering data in multiple fields such as transportation, energy, power, finance, and medical care), has 14 million parameters, has strong modeling capability, and can process data with different resolutions, different variable quantities, and different distribution characteristics.

[0043] Patch refers to a technology for dividing original data (such as images, time series, or text) into multiple smaller fixed-size fragments when processing data.

[0044] Transfer learning refers to a machine learning method that aims to apply the knowledge of a model trained in one field to related fields to improve training efficiency and prediction performance.

[0045] Source wind farm refers to a wind farm that has been running for a long time and has sufficient training data distributed throughout the country.

[0046] Target wind farm refers to a newly built, expanded, or reconstructed wind farm that has no or only a small amount of training data.

[0047] Pre-training refers to the process of training a wind power prediction model using the training data of a source wind farm.

[0048] Wind power prediction large model base refers to a large model adapted to wind power prediction obtained by pre-training the time series large model Moirai.

[0049] Fine-tuning refers to the process of updating the model parameters in the wind power prediction large model base using the training data of the target wind farm.

[0050] Exclusive wind power prediction large model refers to a high-precision wind power prediction large model applied to a target wind farm.

[0051] Numerical weather forecast data refers to meteorological forecast data that reflects the state of atmospheric movement in a certain period of time in the future, obtained by solving the set of fluid mechanics and thermodynamics equations that describe the weather evolution process based on the actual atmospheric conditions under certain initial and boundary conditions through numerical calculations.

[0052] The wind power prediction method based on a large time series model provided in the embodiments of the present disclosure is applicable to wind power prediction scenarios based on large time series models. The method can be performed by a wind power prediction device based on a large time series model. The device can be implemented in software and / or hardware, and the device can be integrated into an electronic device. The electronic device can include, but is not limited to, mobile terminals such as smartphones, laptops, digital broadcast receivers, personal digital assistants (PDAs), tablet personal computers (Tablet PCs), PMPs (portable multimedia players), in-vehicle terminals (such as in-vehicle navigation terminals), wearable devices, and fixed terminals such as digital televisions, desktop computers, smart home devices, and the like.

[0053] Figure 1 A flow chart of a wind power prediction method based on a time series large model provided by an embodiment of the present disclosure, specifically including the following steps: Figure 1 The following steps are shown:

[0054] S110 : Acquire a large time series model and a first training data set of multiple source wind farms.

[0055] It is understandable that the time series pre-training large model is a large-scale neural network model that has been pre-trained with massive historical data. The model can automatically learn complex time series patterns, and can then quickly adapt to the sample characteristics of different fields through simple adjustments. Due to the structures such as the projection layer with multiple patch sizes and the arbitrary variable attention mechanism within the model, as well as the random sampling strategy of the model input length and prediction scale set during pre-training, the time series pre-training large model can adapt to different prediction spans, model input and output structures and characteristics. The multiple source wind farms can specifically be multiple source wind farms across the country. The first training data set includes sufficient and massive training data from multiple source wind farms across the country under all weather conditions, so that the trained basic large model can be applicable to wind power prediction under all weather conditions and has universal applicability.

[0056] Among them, the time series large model includes the input feature extraction layer, the time series modeling layer and the prediction output mapping layer.

[0057] For example, see Figure 2 , Figure 2A structural schematic diagram of a time series large model is provided for an embodiment of the present disclosure. The time series large model comprises an input feature extraction layer, a time series modeling layer, and a prediction output mapping layer. The time series large model first preprocesses historical data through the input feature extraction layer, extracts cross-frequency and cross-variable key features by using multiple patch sizes and variable encoding techniques; then, the time series modeling layer uses an improved Transformer (attention mechanism model) encoder and an arbitrary variable attention mechanism to model the extracted key features in a deep time sequence dependency; finally, the internal representation is decoded into a mixed distribution parameter through the prediction output mapping layer to generate a prediction result of a future time step. Due to the general time sequence evolution law learned in the large-scale time sequence data training process, the time series large model can be adjusted in feature law for specific field time sequence prediction tasks, such as wind power prediction tasks, to achieve field adaptive prediction accuracy and robustness.

[0058] S120, pre-training the time series large model using the first training data set to obtain a basic large model for wind power prediction.

[0059] Understandably, on the basis of the above S110, the time series large model is sequentially pre-trained using sufficient training samples of multiple source wind farms in the country to obtain a basic large model, that is, the time series large model is pre-trained as a wind power prediction model suitable for wind power prediction. The basic large model can be understood as the above wind power prediction large model base. Sequentially training the time series large model means that the current source wind farm is trained on the basis of the model parameters obtained by the last source wind farm. Specifically, one scenario is to train with the training data of one source wind farm each time, that is, the time series large model is pre-trained through the training data of source wind farm a, and then the pre-training is continued through the training data of source wind farm b on the basis of the pre-training (including model parameters and network structure, etc.). A certain number of source wind farms can also be selected from multiple source wind farms as the current source wind farm for batch training, and other possible training processes are not limited. The basic large model can adapt to different prediction spans, model input and output structures, and features, etc., improving the application flexibility and effectiveness in variable scenarios, and thus realizing zero-sample or small-sample wind power prediction of the target wind farm.

[0060] For example, referring to Figure 3 , Figure 3A structural schematic diagram of a basic large model provided by an embodiment of the present disclosure is obtained by pre-training based on a time series large model. Accordingly, the basic large model comprises a feature extraction layer, a wind power time series modeling layer, and a wind power mapping layer. The feature extraction layer is configured to extract features of historical wind power and numerical weather forecast data. The wind power time series modeling layer is configured to perform deep fusion on the extracted features. The wind power mapping layer is configured to map the fused features to a prediction distribution parameter space to generate a final wind power prediction sequence.

[0061] In S130, a historical wind power array and a future weather forecast array of the target wind farm are obtained.

[0062] The plurality of source wind farms do not include the target wind farm.

[0063] It can be understood that, based on the above S120, the target wind farm refers to a wind farm that is newly built, expanded, or reconstructed, and does not have training data or only has a small amount of training data. The source wind farm refers to a wind farm that has been established for a long time and has massive training data under all weather conditions. The plurality of source wind farms do not include the target wind farm. The historical wind power two-dimensional array and the future weather forecast two-dimensional array of the target wind farm are obtained.

[0064] Optionally, before obtaining the historical wind power array and the future weather forecast array of the target wind farm, the method further comprises:

[0065] A second training data set of the target wind farm is obtained, wherein the sample data amount and the sampling period of the second training data set are less than those of the first training data set. In a case where the second training data set comprises at least one historical wind power training array, the parameters of the basic large model are fine-tuned using the second training data set to obtain a target large model, wherein the target large model is dedicated to wind power prediction of the target wind farm. In a case where the second training data set does not comprise a historical wind power training array, the basic large model is directly used as the target large model.

[0066] It can be understood that the second training data set of the target wind farm has a sample size and a sampling period much smaller than those of the first training data set. The second training data set includes limited training data, which can be understood as a zero sample set or a small sample set. The first training sample set includes all training data under sufficient weather conditions, which can be understood as a large sample set or a massive sample set. If the second training sample set does not include the historical wind power training array, the massive wind power data general time sequence features learned by the basic large model in the training data pre-training of multiple source wind farms are directly used to perform zero sample prediction on the future wind power of the target wind farm. For the scenario where the second training sample set does not include the historical wind power training array but includes the weather forecast training array, the basic large model is directly used for zero sample prediction.

[0067] It can be understood that if the second training data set includes the historical wind power training array, that is, the target wind farm has limited historical power data, the basic large model is fine-tuned using the small sample data of the target wind farm. In this way, the massive time sequence data can reflect the general time sequence features in pre-training, and the differences in specific geographical environment, meteorological conditions, and unit characteristics can be considered. That is, the fine-tuned basic large model has both general time sequence features and differentiated characteristics of the target wind farm. The second training data set includes input data and corresponding output data, wherein the input data is the historical wind power training array and the weather forecast training array, and the output data is the actual wind power training array. One possible application scenario is that the second training data set includes both the historical wind power training array and the weather forecast training array. Another possible application scenario is that the second training data set only includes the historical wind power training array. In the above two cases, the basic large model can be fine-tuned according to the second training data set.

[0068] Optionally, the historical wind power array and the future weather forecast array of the target wind farm are obtained, which can be achieved by the following steps:

[0069] The historical wind power array of the target wind farm in the historical set period and the future weather forecast array in the future set period are obtained.

[0070] It can be understood that the historical setting period refers to the collection period of the historical wind power array, for example, the current time is B moment, the historical setting period refers to the T period before B moment, which can be T hours in particular. The future setting period refers to the prediction period of the wind power, for example, the future setting period refers to the N period after B moment, which can be N hours in particular. The first training data set includes the historical wind power training array of T hours and the weather forecast training array of N hours. Wherein, when the time series large model is pre-trained by using the sufficient training samples (i.e. the first training data set) of the plurality of source wind farms, the above-mentioned T and N do not need to be artificially set, and the length of T and N can be randomly sampled inside the large model. In the process of fine-tuning the basic large model based on the small sample training sample (i.e. the second training data set) of the target wind farm, the length of T and N can be set according to the actual demand.

[0071] S140, the historical wind power array and the future weather forecast array are taken as the input of the target large model to predict the future wind power array of the target wind farm.

[0072] Wherein, the target large model is the basic large model or is obtained by fine-tuning the basic large model by using the second training data set of the target wind farm.

[0073] Optionally, the historical wind power array and the future weather forecast array are taken as the input of the target large model to predict the future wind power array of the target wind farm, which can be realized by the following steps:

[0074] The historical wind power array and the future weather forecast array are taken as the input of the target large model to predict the future wind power array of the target wind farm in the future setting period.

[0075] It can be understood that, on the basis of S130, the historical wind power array includes the wind power values recorded by the target wind farm in a past period of time (such as the past 3 hours, 8 hours, or a week, etc.), that is, the historical power in the past short time. The future weather forecast array includes meteorological forecast data in a specific future time period, such as wind speed, wind direction, temperature, humidity, and other key factors affecting wind power output, that is, the weather situation in the future short time. The historical wind power array and the future weather forecast array are input into the pre-trained target large model, and the target large model is used to infer the future wind power array according to the input data, that is, to predict the wind power output of the target wind farm in a specific future time period, that is, the wind power in the future short time. The prediction period of the weather forecast and the prediction period of the wind power are the same, for example, the past 8 hours of historical power and the future 6 hours of weather forecast are collected, and the future 6 hours of wind power is predicted by the target large model. This kind of way can effectively combine historical wind power data and future weather forecast data, improve the accuracy and robustness of future wind power prediction. Especially suitable for small sample wind farms with data scarcity problem, by introducing weather forecast data as an external variable, not only broadens the information source of the large model, but also improves its generalization ability.

[0076] For example, see Figure 4 , Figure 4 The flowchart of a wind power prediction method based on a target large model provided by the embodiments of the present disclosure is shown. The past T hours of historical wind power data are used as the main input, and the future N hours of numerical weather forecast data are used as the dynamic covariate input, which is the secondary input and plays a secondary auxiliary role. The target large model (wind power prediction model in the figure) is used to calculate the future N hours of wind power prediction data.

[0077] The wind power prediction method based on the time sequence large model provided by the embodiments of the present disclosure fully utilizes the general time sequence pattern representation ability learned in the large-scale time sequence data pre-training based on the open source time sequence large model; meanwhile, the "pre-training-fine-tuning" strategy of transfer learning is introduced, the time sequence large model is pre-trained by using the massive and diversified historical wind power data collected by multiple source wind farms all over the country, so that the universal characteristics and differentiated rules presented by different geographical environments, meteorological conditions and wind turbine configurations can be fully learned and refined, and the generalization ability and robustness of the model can be significantly improved; then, zero-shot prediction can be directly performed for the zero-shot wind farm, and for the small-shot wind farm, only a small amount of training data of the target wind farm is needed to make the pre-trained basic large model find the parameter update direction that best reflects the characteristics of the target wind farm, and then realize low-time-consumption, low-cost, high-precision and high-adaptation small-shot wind power prediction. Therefore, while fully utilizing the existing massive training data of multiple source wind farms, the present disclosure greatly reduces the dependence on long-time and large-scale training data of the target wind farm, and provides a more flexible and feasible solution for actual deployment and application.

[0078] On the basis of the above-mentioned embodiments, Figure 5 The flowchart of the wind power prediction method based on the time sequence large model provided by the embodiments of the present disclosure is shown in the figure, and optionally, the time sequence large model is pre-trained by using the first training data set to obtain a basic large model for wind power prediction, which specifically includes the following steps as shown in the figure: Figure 5

[0079] S501, input the training data set corresponding to multiple wind farms in the first training data set into the time sequence large model in turn, for the current wind farm in the multiple wind farms.

[0080] The historical wind power training sample, the weather forecast training sample and the actual wind power training sample; the current wind farm contains sufficient training samples under all weather conditions.

[0081] As can be understood, the first training data set includes a large number of historical wind power training samples, weather forecast training samples and corresponding actual wind power training samples of each source wind farm. The historical wind power training sample can be understood as a historical power two-dimensional array, and the weather forecast training sample can be understood as a numerical weather forecast two-dimensional array. The historical wind power training sample and the weather forecast training sample are input data, and the actual wind power training sample is output data.

[0082] ​The historical wind power training sample and the weather forecast training sample are two-dimensional arrays, the historical wind power training sample includes a first time dimension and a power category dimension, the first time dimension includes a plurality of first time instants for power sampling, the plurality of first time instants are different by a first set time period, and the power category dimension includes first time feature encoding, historical power data, and wind farm number, and the first time feature encoding corresponds to the first time dimension.

[0083] It can be understood that the historical wind power training sample can be understood as a historical power two-dimensional array, and the historical power two-dimensional array includes a first time dimension and a power category dimension, as shown in Figure 6 Figure 6 A two-dimensional structure diagram of a historical wind power training sample provided by the embodiment of the present disclosure. Figure 6 The horizontal direction is the first time dimension, wherein 1, 2, 3, …, T (T hours in the past) represent the 1st time instant, the 2nd time instant, the 3rd time instant, …, and the Mth time instant, respectively, and adjacent two time instants are different by 15 minutes, which is the first set time period; the vertical direction is the power category dimension, including first time feature encoding, historical power data, and wind farm number, etc. The historical power data can be a specific power value, and in order to distinguish the source domain of multiple source wind farms all over the country, different numbers can be set for each source wind farm, and the specific setting method is not limited. For time feature encoding, the time feature encoding of the first sampling time point in the M time points can be defined as 1, the time encoding of the second sampling time point can be defined as 2, and the like, which will not be repeated here.

[0084] Optionally, the weather forecast training sample includes a second time dimension and a meteorological category dimension, wherein the second time dimension includes a plurality of second time instants for weather sampling, the plurality of second time instants are different by a second set time period, and the meteorological category dimension includes second time feature encoding and a plurality of weather forecast data types, the plurality of weather forecast data types include forecast wind condition data and other weather forecast data, the forecast wind condition data includes forecast wind speed data and forecast wind direction data, and the other weather forecast data includes at least one of forecast air pressure data, forecast temperature data, and forecast humidity data.

[0085] It can be understood that the weather forecast training sample can be understood as a numerical weather forecast two-dimensional array (weather feature two-dimensional array), and the weather forecast two-dimensional array includes a second time dimension and a meteorological category dimension, as shown in Figure 7 Figure 7 A two-dimensional structure diagram of a weather forecast training sample provided by the embodiment of the present disclosure. Figure 7 ​​The second time dimension is the middle transverse direction, where 1, 2, 3, …, N represent the 1st moment, the 2nd moment, the 3rd moment, …, the Nth moment, respectively, and the difference between adjacent moments is 15 minutes, that is, the second set time period can also be set to 15 minutes. The first set time period and the second set time period can be the same or different. The weather category dimension includes a second time feature code and a plurality of weather forecast data types. For a plurality of categories of weather forecast data that affect wind power, including but not limited to: forecast wind condition data and other weather forecast data, wherein the forecast wind condition data includes forecast wind speed data and forecast wind direction data, and the other weather forecast data includes but is not limited to one or more of the following: forecast air pressure data, forecast temperature data, and forecast humidity data. Figure 7 Each row in the middle represents a type of weather forecast data. For each weather type, the forecast data at N time points is arranged in chronological order to obtain a one-dimensional array, which can also be called a sequence. The sequence of forecast data for each weather type is arranged according to the rule that the same sampling time point corresponds to the same time dimension to obtain a numerical weather prediction two-dimensional array.

[0086] It can be understood that the training process of the basic large model is as follows: copying the parameters of the time series large model, pre-training using transfer learning, and updating the parameters of all layers of the time series large model E times using the training data of multiple source wind farms. Wherein, set 1≤E≤10000, the specific value of E is determined by the "early stopping method", which is an effective strategy to prevent overfitting in the process of training machine learning models. Specifically, the current training data can be divided into a training set and a validation set, the model parameters of the time series large model are updated using the training set, and the accuracy is verified using the validation set. When the accuracy of the validation set decreases continuously for p times, the iteration process is terminated and the time series large model is no longer updated. Specifically, multiple source wind farms are selected in turn to pre-train the time series prediction large model using transfer learning, and the training process of all source wind farms is as follows: first, determine the current wind farm in the multiple source wind farms. The training data of each wind farm in the source domain is sufficient for data-driven time series large models. Wherein, the current wind farm refers to the wind farm whose corresponding training data is used to pre-train the time series model in the current round, and all training data of the current wind farm is referred to as the current training data. Then, the optimal parameters of the time series large model under the current wind farm are calculated, which includes the steps of S502 to S505.

[0087] S502, select a group of training samples from the sufficient training samples.

[0088] Wherein, a group of training samples includes historical wind power training samples, weather forecast training samples, and actual wind power training samples.

[0089] It is understood that, on the basis of S501, a group of training samples is randomly selected from the sufficient training samples under all weather conditions of the current wind farm, denoted as group A training samples, each group of training samples includes a historical wind power training sample and a weather forecast training sample, wherein A is an integer greater than or equal to 1, which can be set according to demand. In addition, the set A should not be greater than the total number of training data contained in the current wind farm.

[0090] S503, input the feature parameters of the historical wind power training sample and the weather forecast training sample into the time series large model to obtain a corresponding wind power prediction training sample.

[0091] Wherein, the time series large model has model parameters trained based on the previous wind farm corresponding training sample set.

[0092] It is understood that, on the basis of S502, the feature parameters of the A historical wind power training samples and the A weather forecast training samples in the A training samples are input into the time series large model to obtain the corresponding A wind power prediction training samples.

[0093] S504, calculate the first average prediction error between the wind power prediction training sample and the actual wind power training sample.

[0094] It is understood that, on the basis of S503, the wind power prediction training sample refers to the predicted value output by the large model, and the actual wind power training sample refers to the actual value. Calculate the average prediction error between the A predicted values and the actual values, denoted as the first average prediction error. Wherein, the error calculation method can adopt mean square error, mean absolute error, etc., which is not limited here.

[0095] S505, update all layer parameters of the time series large model according to the first average prediction error to obtain the optimal parameters under all weather conditions of the current wind farm, until the training of the time series large model by the last wind farm corresponding training data set is completed, and the basic large model for wind power prediction is obtained.

[0096] It is understood that, on the basis of S504, update all layer parameters of the large model according to the first average prediction error to obtain the optimal parameters under all weather conditions of the current wind farm, wherein the parameter update algorithm can adopt the parameter update method in the traditional neural network training process, such as Adam, other possible parameter update methods are not described.

[0097] It can be understood that after the optimal parameter calculation is completed, the optimal parameter is taken as the final parameter of the time series large model. After the training data of the plurality of source wind farms in the source domain is sequentially traversed, the optimal parameter at this time is taken as the final parameter of the time series large model, and a wind power prediction large model base (basic large model) is obtained. At this time, the parameters of the basic large model are the parameters after the last pre-training update of the time series prediction large model.

[0098] The wind power prediction method based on the time series large model provided in the embodiments of the present disclosure pre-trains the time series large model through the massive training data of the plurality of source wind farms, so that the time series large model is applicable to wind power prediction, and a basic large model conforming to different geographical environments, meteorological conditions and wind turbine configurations is obtained, thereby significantly improving the generalization ability and robustness of the model.

[0099] On the basis of the above-mentioned embodiments, Figure 8 For the flowchart of the wind power prediction method based on the time series large model provided in the embodiments of the present disclosure, the feature parameters of the historical wind power training sample and the weather forecast training sample are optionally input into the time series large model to obtain a corresponding group of wind power prediction training samples, which specifically includes the following steps as shown in Figure 8

[0100] S801, the input feature extraction layer is used to extract features from the historical wind power training sample and the weather forecast training sample to obtain power feature training samples and weather feature training samples with time sequence and variable distinction.

[0101] It can be understood that the power feature training sample (power feature two-dimensional array) and the weather feature training sample (weather feature two-dimensional array) both include a time dimension and a feature dimension. The time dimension corresponds to the time dimension of the historical power feature two-dimensional array and the weather feature two-dimensional array. Each dimension in the feature dimension represents a feature extracted, and the number of feature dimensions can be set as a power of 2, such as 8, 16, 32, 64, ….

[0102] It can be understood that in the pre-training of the time series large model, the feature extraction layer uses a multi-patch (Patch) size input projection module to pre-process the historical wind power training sample (historical power two-dimensional array) and the weather forecast training sample (numerical weather forecast two-dimensional array). The specific steps are as follows: the input two-dimensional array is divided into non-overlapping patches, different patch sizes are selected according to the data sampling frequency, and each patch is mapped to a unified high-dimensional embedding representation by linear projection to generate historical power feature two-dimensional arrays and weather feature two-dimensional arrays with time sequence and variable distinction. The feature extraction layer can not only effectively capture the trend, period and local features in the data, but also provide a unified high-dimensional embedding representation for subsequent other layers.​

[0103] S802, input the power feature training sample and the weather feature training sample into the time sequence modeling layer, perform deep feature fusion through the multi-head attention module and the time sequence modeling module included in the time sequence modeling layer, and obtain a fused time sequence context feature training sample.

[0104] Understandably, on the basis of S801, the time sequence modeling layer adopts a Transformer structure, which is specially used for long short-term dependence modeling of time sequences. First, the layer uses an arbitrary variable attention mechanism to convert the input historical power feature two-dimensional array and the weather feature two-dimensional array into unified sequence representation, and assigns an independent variable ID to each variable to encode the relationship between variables. Subsequently, the full-automatic attention mechanism is used in combination with the RoPE to enhance the modeling of time sequence dependence between different time steps, and a fused time sequence context feature two-dimensional array (time sequence context feature training sample) is obtained. Secondly, in order to better fuse multi-variable information, the time sequence modeling layer can also introduce a binary attention bias to ensure the invariance of the variable index and realize efficient modeling of any number of variables. In addition, in the calculation process, the time sequence modeling layer adopts a feedforward network structure, in which the nonlinear activation function is replaced by SwiGLU instead of the traditional ReLU to improve the expression ability. At the same time, RMSNorm and query-key normalization are used to improve the stability and training efficiency of the deep Transformer. After multi-layer Transformer stacking, the time sequence modeling layer outputs the fused time sequence context feature two-dimensional array, which provides sufficient semantic information for the final power mapping.

[0105] S803, using the prediction output mapping layer, mapping the high-dimensional time sequence context feature training sample to the prediction distribution parameter space to obtain a wind power prediction training sample.

[0106] It can be understood that, on the basis of S802, the prediction output mapping layer utilizes a multi-patch size output projection module to map the high-dimensional context features output by the time series modeling layer to the prediction distribution parameter space, thereby generating the final wind power prediction sequence (wind power prediction training sample). The specific implementation includes: first, a multi-patch size output projection mechanism is adopted to adapt to the data characteristics of different time resolutions, ensuring that high-frequency and low-frequency time series information can be fully utilized. Subsequently, linear projection is used to map the hidden state of the Transformer encoder to the parameter space of the prediction distribution. In the output module, Moirai adopts a hybrid probability distribution scheme to adapt to the possible asymmetry, multimodality and uncertainty of wind power data. Specifically, the prediction distribution can be composed of multiple parameterized distributions (such as Student's distribution, negative binomial distribution, lognormal distribution or low-variance normal distribution), and the optimal fitting is achieved by learning the weights and parameters. Finally, the prediction output mapping layer decodes the parameters of the prediction distribution to output the wind power prediction training sample.

[0107] The wind power prediction method based on a time series large model provided by the embodiments of the present disclosure can predict accurate wind power.

[0108] On the basis of the above-mentioned embodiments, Figure 9 The flowchart of the wind power prediction method based on a time series large model provided by the embodiments of the present disclosure is shown in the figure, and optionally, the parameters of the basic large model are fine-tuned using the second training data set to obtain a target large model, which specifically includes the following steps as shown in the figure. Figure 9 As shown in the figure.

[0109] It can be understood that the parameters of the basic large model are updated by U steps using the training data set of the target wind farm, where U is an integer greater than or equal to 1, which can be set according to requirements. The parameter update of the basic large model is as follows.

[0110] S901, select a set number of training samples from the second training data set.

[0111] It can be understood that M training samples are randomly selected from all training samples of the target wind farm, where M is the set number, M is an integer greater than or equal to 1, and the set M should not be greater than the total number of training samples contained in the target wind farm, which can be set according to requirements.

[0112] S902, input the historical power training array and the weather forecast training array in the set number of training samples into the basic large model respectively to obtain the corresponding set number of wind power prediction training arrays.

[0113] Understandably, on the basis of S901, the historical power training array and the weather forecast training array in the selected M training samples are input into the base large model respectively, and M corresponding wind power prediction sequences (wind power prediction training array) are obtained.

[0114] S903, calculate the second average prediction error between the set number of wind power prediction training arrays and the set number of wind power actual training arrays.

[0115] Understandably, on the basis of S902, the average prediction error between the M wind power prediction training arrays (predicted values) and the corresponding M wind power actual training arrays (actual values) is calculated, wherein the error calculation method can adopt mean square error, mean absolute error, etc.

[0116] S904, update the parameters of the base large model according to the second average prediction error to obtain a target large model.

[0117] Understandably, on the basis of S903, the parameters of the base large model are updated according to the second average prediction error to obtain a target large model.

[0118] It is understood that the training process of the time series large model to obtain the base large model and the training process of the base large model to obtain the target large model are similar, which will not be repeated here.

[0119] The wind power prediction method based on the time series large model provided by the embodiment of the present disclosure trains a target large model dedicated to a target wind farm.

[0120] Figure 10 A flowchart of a wind power prediction method based on a time series large model provided by the embodiment of the present disclosure, specifically comprising the following steps as shown in Figure 10 as shown in the following steps:

[0121] S101, obtain a time series pre-training large model Moirai.

[0122] S102, obtain source wind farm training data.

[0123] S103, pre-train the time series prediction large model Moirai using sufficient training samples of multiple source wind farms across the country in turn, to obtain a wind power prediction large model base.

[0124] S104, obtain a historical power two-dimensional array and a real-time weather forecast two-dimensional array of a target wind farm.

[0125] S105, input the historical power two-dimensional array and the real-time weather forecast two-dimensional array of the target wind farm into the wind power prediction large model base, to obtain the real-time wind power prediction sequence of the target wind farm.

[0126] S106, obtain the small sample training data of the target wind farm.

[0127] S107, fine-tune the wind power prediction large model base parameters by using the limited training sample of the target wind farm, to obtain the exclusive wind power prediction large model of the target wind farm.

[0128] S108, obtain the historical power two-dimensional array and the real-time weather forecast two-dimensional array of the target wind farm.

[0129] S109, input the historical power two-dimensional array and the real-time weather forecast two-dimensional array of the target wind farm into the exclusive wind power prediction large model of the target wind farm, to obtain the real-time wind power prediction sequence of the target wind farm.

[0130] It can be understood that the specific implementation of S101 to S109 described above can refer to the embodiments described above, and will not be repeated here.

[0131] Figure 11 A structure diagram of a wind power prediction device based on a time series large model is provided for the embodiments of the present disclosure. The wind power prediction device based on the time series large model provided by the embodiments of the present disclosure can perform the processing flow provided by the wind power prediction method based on the time series large model, as shown in Figure 11 The device 1100 includes:

[0132] A first obtaining unit 1101 is configured to obtain a time series large model and a first training data set of a plurality of source wind farms.

[0133] A training unit 1102 is configured to pre-train the time series large model by using the first training data set, to obtain a basic large model for wind power prediction.

[0134] A second obtaining unit 1103 is configured to obtain a historical wind power array and a future weather forecast array of a target wind farm, wherein the plurality of source wind farms do not include the target wind farm.

[0135] A prediction unit 1104 is configured to input the historical wind power array and the future weather forecast array as an input of a target large model, to predict a future wind power array of the target wind farm, wherein the target large model is the basic large model or is obtained by fine-tuning the basic large model by using a second training data set of the target wind farm.

[0136] Optionally, the device 1100 is further configured to:

[0137] obtaining a second training data set of the target wind farm, wherein a sample data amount and a sampling period of the second training data set are less than those of the first training data set;

[0138] in a case where the second training data set includes at least one historical wind power training array, fine-tuning parameters of the basic large model by using the second training data set to obtain a target large model, wherein the target large model is dedicated to wind power prediction of the target wind farm; or

[0139] in a case where the second training data set does not include the historical wind power training array, directly taking the basic large model as the target large model.

[0140] Optionally, the training unit 1102 is configured to:

[0141] inputing the training data sets corresponding to the plurality of wind farms in the first training data set into the time series large model in sequence, for a current wind farm in the plurality of wind farms, wherein the current wind farm contains sufficient training samples under all weather conditions;

[0142] selecting a group of training samples from the sufficient training samples, wherein the group of training samples includes historical wind power training samples, weather forecast training samples and actual wind power training samples;

[0143] inputing the feature parameters of the historical wind power training samples and the weather forecast training samples into the time series large model to obtain a group of corresponding wind power prediction training samples, wherein the time series large model has model parameters trained based on the training sample set corresponding to the previous wind farm;

[0144] calculating a first average prediction error between the wind power prediction training samples and the actual wind power training samples;

[0145] updating all layer parameters of the time series large model according to the first average prediction error to obtain optimal parameters under all weather conditions of the current wind farm, until the training of the time series large model by the training data set corresponding to the last wind farm is completed, to obtain a basic large model for wind power prediction.

[0146] wherein the time series large model includes an input feature extraction layer, a time series modeling layer and a prediction output mapping layer.

[0147] Optionally, the training unit 1102 is configured to:

[0148] extracting features of the historical wind power training samples and the weather forecast training samples by using the input feature extraction layer to obtain power feature training samples and weather feature training samples with time sequence and variable distinction;

[0149] The power feature training sample and the weather feature training sample are input into a time sequence modeling layer, and deep feature fusion is performed on the time sequence modeling layer through a multi-head attention module and a time sequence modeling module included in the time sequence modeling layer to obtain a fused time sequence context feature training sample;

[0150] The high-dimensional time sequence context feature training sample is mapped to a prediction distribution parameter space by using a prediction output mapping layer to obtain a wind power prediction training sample.

[0151] Optionally, the apparatus 1100 is further configured to:

[0152] select a set number of training samples from the second training data set;

[0153] input the historical power training array and the weather forecast training array in the set number of training samples into the base large model respectively to obtain a set number of wind power prediction training arrays corresponding thereto;

[0154] calculate a second average prediction error between the set number of wind power prediction training arrays and the set number of wind power actual training arrays;

[0155] update the parameters of the base large model according to the second average prediction error to obtain a target large model.

[0156] The historical wind power training sample and the weather forecast training sample are two-dimensional arrays.

[0157] The historical wind power training sample includes a first time dimension and a power category dimension, the first time dimension includes a plurality of first time points for power sampling, the plurality of first time points are different by a first set time period, and the power category dimension includes a first time feature code, historical power data and a wind farm number, the first time feature code corresponding to the first time dimension.

[0158] The weather forecast training sample includes a second time dimension and a meteorological category dimension, the second time dimension includes a plurality of second time points for weather sampling, the plurality of second time points are different by a second set time period, and the meteorological category dimension includes a second time feature code and a plurality of weather forecast data types, the plurality of weather forecast data types including forecast wind condition data and other weather forecast data, the forecast wind condition data including forecast wind speed data and forecast wind direction data, and the other weather forecast data including at least one of forecast air pressure data, forecast temperature data and forecast humidity data.

[0159] Optionally, the second obtaining unit 1103 is configured to:

[0160] obtain a historical wind power array of the target wind farm within a historical set time period and a future weather forecast array within a future set time period;

[0161] The historical wind power array and the future weather forecast array are taken as inputs of the target large model to predict a future wind power array of the target wind farm in a future set period.

[0162] Figure 11 The wind power prediction device based on the time series large model of the illustrated embodiment can be used to implement the technical solutions of the method embodiment described above, and has similar implementation principles and technical effects, which will not be described here again.

[0163] Figure 12 The structural schematic diagram of the electronic device provided by the embodiment of the present disclosure is shown in FIG. 12. The following specifically refers to Figure 12 which shows a structural schematic diagram suitable for implementing the electronic device 1200 in the embodiment of the present disclosure. The electronic device 1200 in the embodiment of the present disclosure can include but is not limited to a mobile terminal such as a mobile phone, a notebook computer, a digital broadcast receiver, a PDA (Personal Digital Assistant), a PAD (Tablet Personal Computer), a PMP (Portable Multimedia Player), a vehicle terminal (such as a vehicle navigation terminal), a wearable electronic device, and the like, and a fixed terminal such as a digital TV, a desktop computer, a smart home device, and the like. Figure 12 The electronic device shown is only an example, and should not bring any limitation to the functions and use range of the embodiment of the present disclosure.

[0164] As shown in Figure 12 , the electronic device 1200 can include a processing device 1201 (such as a central processor, a graphics processor, etc.) which can perform various appropriate actions and processes to implement the wind power prediction method based on the time series large model of the embodiment as described in the present disclosure according to the program stored in the read-only memory (ROM) 1202 or the program loaded from the storage device 1208 into the random access memory (RAM) 1203. In the RAM 1203, various programs and data required for the operation of the electronic device 1200 are also stored. The processing device 1201, the ROM 1202, and the RAM 1203 are connected to each other through a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0165] Generally, the following devices can be connected to the I / O interface 1205: an input device 1206 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 1207 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 1208 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1209. The communication device 1209 can allow the electronic device 1200 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 12Electronic device 1200 is illustrated with various devices, but it is understood that not all of the illustrated devices are required to be implemented or present. More or fewer devices can alternatively be implemented or present.

[0166] In particular, according to embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, embodiments of the present 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 executing the methods illustrated by the flowcharts, thereby implementing the wind power prediction method based on time series large model as described above. In such embodiments, the computer program can be downloaded and installed from a network through the communication device 1209, or installed from the storage device 1208, or installed from the ROM 1202. When the computer program is executed by the processing device 1201, the above-mentioned functions defined in the methods of the embodiments of the present disclosure are performed.

[0167] It should be noted that the computer-readable medium described above in the present disclosure can be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. The computer-readable storage medium, for example, can be, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any suitable combination thereof. More specific examples of the computer-readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device. In the present disclosure, the computer-readable signal medium can include a data signal propagated in a baseband or as a carrier wave in a propagated data signal, which bears computer-readable program code. Such propagated data signal can take many forms, including but not limited to, an electromagnetic signal, an optical signal, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium that is not a storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained in the computer-readable medium can be transmitted by any suitable medium, including but not limited to, wire, cable, RF (radio frequency), or the like, or any suitable combination thereof.

[0168] In some embodiments, the client and server can communicate using any currently known or later developed network protocol, such as HTTP (HyperText Transfer Protocol), and can be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or later developed network.

[0169] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0170] Optionally, when the above one or more programs are executed by the electronic device, the electronic device may also execute other steps described in the above embodiments.

[0171] Computer program code for performing the operations of the present disclosure may 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, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0172] The computer program product of the first aspect can include one or more non-transitory computer-readable media storing instructions that, when executed, cause one or more processors to perform the operations of the method of the first aspect. The computer program product of the first aspect can include a computer-readable medium storing instructions that, when executed, cause one or more processors to perform the operations of the method of the first aspect.

[0173] The units described in the embodiments of the present disclosure can be implemented by software, or by hardware, or by a combination of software and hardware. In some cases, the names of the units do not constitute a limitation on the units themselves.

[0174] The functions described in this description above can be implemented in hardware, software, or any combination thereof. If implemented in software, the functions can be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media include both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media can be any available media that can be accessed by a general purpose or special purpose computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code means in the form of instructions or data structures and that can be accessed by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor. Also, functional

[0175] In the context of the present disclosure, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0176] It needs to be explained that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or gateway including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or gateway. Without more limitations, the element defined by the statement "a wind power prediction based on a time series big model" does not exclude the existence of another same element in the process, method, article or gateway including the element.

[0177] The above description is only a specific implementation of the present disclosure, enabling those skilled in the art to understand or implement the present disclosure. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present disclosure. Therefore, the present disclosure will not be limited to these embodiments described herein, but will conform to the widest range consistent with the principles and novel features disclosed herein.

Claims

1. A wind power prediction method based on a time series large model, characterized in that: include: Obtain a first training data set of a large time series model and multiple source wind farms, wherein the first training data set includes a large number of historical wind power training samples, weather forecast training samples, and corresponding actual wind power training samples for each source wind farm, and the historical wind power training samples and the weather forecast training samples refer to two-dimensional arrays, wherein the historical wind power training samples include a first time dimension and a power category dimension, the first time dimension includes multiple first moments for power sampling, and the multiple first moments differ by a first set period, the power category dimension includes a first time feature code, historical power data, and a wind farm number, and the first time feature code corresponds to the first time dimension; the weather forecast training samples include a second time dimension and a meteorological category dimension, wherein the second time dimension includes multiple second moments for weather sampling, and the multiple second moments differ by a second set period, and the meteorological category dimension includes a second time feature code and multiple meteorological forecast data types; Pre-training the time series large model using the first training data set to obtain a basic large model for wind power prediction; wherein the basic large model includes a feature extraction layer, a wind power time series modeling layer, and a wind power mapping layer, the feature extraction layer is used to extract features of wind power and weather forecast data, the wind power time series modeling layer is used to fuse the extracted features, and the wind power mapping layer is used to map the fused features to a prediction distribution parameter space to generate a final wind power prediction sequence; Obtaining a historical wind power array for the past T hours and a future weather forecast array for the next N hours of a target wind farm, and calculating wind power forecast data for the next N hours using the historical wind power array as a primary input and the future weather forecast array as a dynamic covariate input, wherein the multiple source wind farms do not include the target wind farm; The historical wind power array and the future weather forecast array are used as inputs of a target large model to predict the future wind power array of the target wind farm in the next N hours, wherein the target large model is the basic large model.

2. The method according to claim 1, characterized in that Before obtaining the historical wind power array and the future weather forecast array of the target wind farm, the method further includes: Acquire a second training data set of the target wind farm, wherein the sample data volume and sampling period of the second training data set are smaller than those of the first training data set; In a case where the second training data set includes at least one historical wind power training array, the parameters of the basic large model are fine-tuned using the second training data set to obtain a target large model, wherein the target large model is specifically used for wind power prediction of the target wind farm; or In the case that the second training data set does not include a historical wind power training array, the basic large model is directly used as the target large model.

3. The method according to claim 1, characterized in that The method of pre-training the time series large model using the first training data set to obtain a basic large model for wind power prediction includes: Inputting the training data sets corresponding to the plurality of wind farms in the first training data set into the time series large model in sequence, for a current wind farm among the plurality of wind farms, wherein the current wind farm contains sufficient training samples under all weather conditions; Selecting a group of training samples from the sufficient training samples, wherein the group of training samples includes historical wind power training samples, weather forecast training samples, and actual wind power training samples; Inputting characteristic parameters of the historical wind power training samples and the weather forecast training samples into the time series large model to obtain a set of corresponding wind power prediction training samples, wherein the time series large model has model parameters trained based on a corresponding training sample set of a previous wind farm; Calculating a first average prediction error between the wind power prediction training samples and the wind power actual training samples; All layer parameters of the time series large model are updated according to the first average prediction error to obtain the optimal parameters under all weather conditions of the current wind farm, until the training of the time series large model with the training data set corresponding to the last wind farm is completed to obtain the basic large model for wind power prediction.

4. The method according to claim 3, characterized in that The time series large model includes an input feature extraction layer, a time series modeling layer, and a prediction output mapping layer. The feature parameters of the historical wind power training samples and the weather forecast training samples are input into the time series large model to obtain a set of corresponding wind power prediction training samples, including: Using the input feature extraction layer to perform feature extraction on the historical wind power training samples and the weather forecast training samples to obtain power feature training samples and weather feature training samples with time series and variable distinguishability; Inputting the power feature training samples and the weather feature training samples into the time series modeling layer, performing deep feature fusion through the multi-head attention module and the time series modeling module included in the time series modeling layer to obtain fused time series context feature training samples; The prediction output mapping layer is used to map the high-dimensional temporal context feature training samples to the prediction distribution parameter space to obtain wind power prediction training samples.

5. The method according to claim 2, characterized in that The method of fine-tuning the parameters of the basic large model using the second training data set to obtain a target large model includes: Selecting a set number of training samples from the second training data set; Inputting the historical power training array and the weather forecast training array in the set number of training samples into the basic large model respectively to obtain the corresponding set number of wind power prediction training arrays; Calculating a second average prediction error between the set number of wind power prediction training arrays and the set number of wind power actual training arrays; The parameters of the basic large model are updated according to the second average prediction error to obtain a target large model.

6. The method according to claim 4, characterized in that The multiple weather forecast data types include forecast wind condition data and other weather forecast data, the forecast wind condition data includes forecast wind speed data and forecast wind direction data, and the other weather forecast data includes at least one of forecast air pressure data, forecast temperature data and forecast humidity data.

7. The method according to claim 1, characterized in that The step of obtaining the historical wind power array and future weather forecast array of the target wind farm includes: Obtaining a historical wind power array of a target wind farm within a set historical period and a future weather forecast array within a set future period; The method of using the historical wind power array and the future weather forecast array as inputs of a target large model to predict the future wind power array of the target wind farm includes: The historical wind power array and the future weather forecast array are used as inputs of the target large model to predict the future wind power array of the target wind farm within the future set time period.

8. A wind power prediction device based on a time series large model, characterized in that: include: a first acquisition unit, configured to acquire a first training data set of a large time series model and multiple source wind farms, wherein the first training data set includes a large number of historical wind power training samples, weather forecast training samples, and corresponding actual wind power training samples for each source wind farm, and the historical wind power training samples and the weather forecast training samples refer to two-dimensional arrays, wherein the historical wind power training samples include a first time dimension and a power category dimension, the first time dimension includes multiple first moments for power sampling, the multiple first moments differ by a first set period, the power category dimension includes a first time feature code, historical power data, and a wind farm number, and the first time feature code corresponds to the first time dimension; the weather forecast training samples include a second time dimension and a meteorological category dimension, wherein the second time dimension includes multiple second moments for weather sampling, the multiple second moments differ by a second set period, and the meteorological category dimension includes a second time feature code and multiple meteorological forecast data types; a training unit, configured to pre-train the time series large model using the first training data set to obtain a basic large model for wind power prediction; wherein the basic large model includes a feature extraction layer, a wind power time series modeling layer, and a wind power mapping layer; the feature extraction layer is configured to extract features of wind power and weather forecast data; the wind power time series modeling layer is configured to fuse the extracted features; and the wind power mapping layer is configured to map the fused features to a prediction distribution parameter space to generate a final wind power prediction sequence; a second acquisition unit, configured to acquire a historical wind power array of the target wind farm in the past T hours and a future weather forecast array for the next N hours, and calculate wind power forecast data for the next N hours using the historical wind power array as a primary input and the future weather forecast array as a dynamic covariate input, wherein the multiple source wind farms do not include the target wind farm; The prediction unit is configured to use the historical wind power array and the future weather forecast array as inputs of a target large model to predict the future wind power array of the target wind farm in the next N hours, wherein the target large model is the basic large model.

9. An electronic device, characterized in that: include: Memory; processor; as well as computer programs; The computer program is stored in the memory and is configured to be executed by the processor to implement the wind power prediction method based on a time series large model as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the wind power prediction method based on a time series large model as claimed in any one of claims 1 to 7 are implemented.

Images