Wind power prediction method driven by large language model and combined with physical prediction field

Abstract

The application provides a wind power prediction method driven by a large language model and fusing a physical prediction field, and relates to the field of wind power prediction.The application fuses physical prediction field data as historical input and future correction, can effectively utilize the large language model neural network architecture to learn the data mode of the historical input, predict future wind power, and greatly improve the prediction accuracy.The proposed adaptive attention decomposition method is used for multivariate information decoupling, which can effectively solve the information coupling and high-dimensional evolution complexity caused by the coupling of multiple variable information in a traditional model embedding, effectively aggregate the correlation between multiple variables, separate the embedding encoding, and improve the prediction performance.In addition, the autoregressive prediction method is used to predict the future wind power, which is consistent with the pre-training method of the large language model, fully develops the large language model capability, and comprehensively improves the wind power prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power prediction, in particular to a wind power prediction method driven by a large language model and fusing a physical prediction field. BACKGROUND

[0002] With the deep dependence of modern industry on fossil fuels, the concentration of carbon dioxide in the atmosphere continues to rise, the greenhouse effect continues to intensify, and the situation of climate change is becoming increasingly severe. In this context, wind energy, as a highly potential clean energy, has attracted widespread attention due to its zero emissions, conservation of natural resources, and environmental friendliness to facilities. However, the intermittency and volatility of wind energy lead to unstable power output, which not only challenges the reliability of power grid operation, but also significantly increases the system dispatching cost, severely restricting the large-scale application and promotion of wind energy. Therefore, accurately predicting the trend of wind power has become the key to breaking this bottleneck.

[0003] Wind power prediction has always been a key area of research. Early research mainly relied on statistical models. For example, traditional methods such as autoregressive moving average (ARMA) and autoregressive integrated moving average (ARIMA) build the mapping relationship between input and output variables through statistical techniques. Although these methods are easy to apply, they rely too much on statistical parameters, limiting their ability to handle non-stationary time series and model complex nonlinear relationships. In addition to statistical-based methods, traditional prediction methods also include finite element simulation prediction based on physical processes. This method predicts future weather conditions by accurately solving partial differential equations, and numerical weather prediction (NWP) is a typical representative. Although the wind speed prediction calculated by the physical simulation-based method is relatively accurate, the calculation consumption of this method is large, and due to reasons such as spatial resolution, there is still deviation, which is not good for wind power prediction.

[0004] With the rapid development of artificial intelligence technology, a new generation of wind power prediction models based on deep neural networks has emerged and quickly become a hot topic in the research field. Deep neural networks, with their outstanding ability to model complex nonlinear relationships, have significantly improved the accuracy and robustness of wind power prediction, showing significant advantages over traditional methods. Among the many neural network methods, large language models have achieved the best results in time series prediction tasks, showing great application potential. Using large language models for wind power prediction can take advantage of their large context window to efficiently learn historical data patterns, while their powerful prediction capabilities and excellent prediction performance make them stand out. However, wind power prediction tasks are not simply time series prediction. They do not have the independence of variables, and there are complex correlations between different variables, which need to be effectively extracted to achieve accurate prediction. This makes the application of large language models in wind power prediction face many challenges. Although large language models can efficiently learn historical data patterns using their large context window, and have powerful prediction capabilities and excellent prediction performance, how to overcome the difficulties brought by variable correlations in wind power prediction is still a key problem that needs to be solved. SUMMARY

[0005] (I) Technical problems solved

[0006] To solve the technical problems of the prior art, the present application provides a large language model driven wind power prediction method fusing physical prediction fields, which solves the technical problem of difficulty in perceiving the correlation of variables.

[0007] (II) Technical solutions

[0008] To achieve the above purpose, the present application is realized by the following technical solutions:

[0009] A large language model driven wind power prediction method fusing physical prediction fields, comprising:

[0010] Obtaining and preprocessing the historical power of the wind farm, the historical climate variables and the future climate variables corresponding to the physical prediction field data;

[0011] Based on the historical power of the wind farm and the historical climate variables, a historical input is constructed, and based on the future climate variables, a future input is constructed;

[0012] The historical input and the future input are non-overlappingly segmented, and the segmented variables are feature extracted and semantically encoded to obtain corresponding historical embedding encoding and future embedding encoding respectively;

[0013] The historical embedding encoding and the future embedding encoding are adaptively decomposed, and through position encoding and multi-head scaled dot product attention mechanism, corresponding decoupling embedding encoding is obtained respectively;

[0014] The decoupled embedding encoding of history is used as input to a pre-trained large language model to predict the decomposed embedding encoding of future moments, and then the two are merged.

[0015] By using a multi-head cross-attention mechanism, the decoupled embedding code is fused to correct the merged embedding code, so as to decode and obtain the predicted power of the wind farm at future moments.

[0016] The predicted power is recoded and concatenated with the corresponding historical embedding code to obtain a new historical embedding code. A new future embedding code is introduced, and the above operation is repeated. The outputs of each operation are concatenated along the time dimension to obtain the final autoregressive prediction result.

[0017] Preferably, for the input embedding E, it is represented as along the time dimension. P sequence The adaptive attention decomposition includes:

[0018] By introducing learnable embeddings, adaptive attention decomposition utilizes a multi-head scaled dot product attention mechanism to compute the decomposed embeddings, as shown in the following equation:

[0019]

[0020] in, This indicates a learnable embedding. Represents the set of real numbers. K Indicates the number of decomposed embeddings. D Indicates the dimension of a vector; Indicates positional embedding, V Indicates the number of input variables; Q, K, and V represent the query matrix, key matrix, and value matrix, respectively; matrix W Q W K W V Both represent learnable weight matrices used to project the input embedding and the learnable embedding into the attention space; D' Indicates the dimension of the key space; i Indicates the index of the input embedding;

[0021] The decoupled embedding code after decomposition is computed using attention, as shown in the following equation:

[0022]

[0023] in, It represents the decoupled variable encoding; Attention represents the attention function; softmax represents the activation function; superscript T Indicates transpose;

[0024] Collect various input embeddings Corresponding decoupling embedding , along the time dimension P recombination to obtain the final decoupled embedding code .

[0025] Preferably, the historical decoupled embedding code is input into the pre-trained large language model to predict the decomposition embedding code at the future time, including:

[0026] Align the dimensions using a linear mapping layer, as shown in the following formula:

[0027]

[0028] wherein, represents the embedding after dimensionality increase, LinearUp represents the dimensionality increase linear mapping function; represents the historical decoupled embedding code;

[0029] as input into the pre-trained large language model to obtain the prediction result at the future time , as shown in the following formula:

[0030]

[0031] wherein, LLMPredict represents the large language model prediction function;

[0032] Dimensionality reduction mapping is performed on the prediction result, as shown in the following formula:

[0033]

[0034] wherein, LinearDown represents the dimensionality reduction linear mapping function; represents the embedding after dimensionality reduction, represents the decomposition embedding code at the future time.

[0035] Preferably, the decomposition embedding code is merged using a linear mapping layer.

[0036] Preferably, the future decoupled embedding code is fused to correct the merged embedding code through a multi-head cross-attention mechanism, including:

[0037] Calculate the cross-attention matrix, as shown in the following formula:

[0038]

[0039] wherein, Q c , K c and V c represent the query matrix, key matrix and value matrix of cross-attention, all represent the learnable weight matrix thereof, ​represents its attention dimension; represents the merged prediction embedding; represents the decoupled embedding encoding of the future, is its only element; represents the real set;

[0040] perform cross-attention calculation, as shown in the following formula:

[0041]

[0042] wherein, represents the modified embedding encoding; CrossAttention represents the cross-attention function; softmax represents the activation function; the superscript T represents the transpose.

[0043] Preferably, a multi-layer perceptron is used to extract features and semantically encode the segmented variables.

[0044] Preferably, a multi-layer perceptron decoder is used to decode the modified embedding encoding.

[0045] A large language model driven wind power prediction system fusing a physical prediction field, comprising:

[0046] An acquisition module is configured to acquire and preprocess historical power of a wind farm, historical climate variables and future climate variables corresponding to physical prediction field data of the wind farm;

[0047] A construction module is configured to construct a historical input based on the historical power of the wind farm and the historical climate variables, and construct a future input based on the future climate variables;

[0048] An encoding module is configured to perform non-overlapping segmentation on the historical input and the future input, and extract features and semantically encode the segmented variables to obtain corresponding historical embedding encoding and future embedding encoding, respectively;

[0049] A decoupling module is configured to perform adaptive attention decomposition on the historical embedding encoding and the future embedding encoding, and obtain corresponding decoupled embedding encoding through position encoding and multi-head scaled dot-product attention mechanism;

[0050] A prediction module is configured to use the historical decoupled embedding encoding as an input of a pre-trained large language model, predict the decomposition embedding encoding at a future time, and merge the decomposition embedding encoding;

[0051] A correction module is configured to correct the merged embedding encoding by fusing the future decoupled embedding encoding through a multi-head cross-attention mechanism, to obtain the predicted power of the wind farm at the future time through decoding;

[0052] An autoregressive module is configured to splice the predicted power after recoding with the corresponding historical embedding code to obtain a new historical embedding code, introduce a new future embedding code, repeat the above operation, and splice the outputs of each time along the time dimension to obtain a final autoregressive prediction result.

[0053] A storage medium storing a computer program for large language model driven fusion of physical prediction field for wind power prediction, wherein the computer program causes a computer to execute the wind power prediction method as described above.

[0054] An electronic device comprising:

[0055] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs comprising instructions for executing the wind power prediction method as described above.

[0056] (Three) beneficial effects

[0057] The present application provides a large language model driven fusion of physical prediction field for wind power prediction method. Compared with the prior art, it has the following beneficial effects:

[0058] 1. In order to increase the length of the wind power prediction interval, the present application fuses physical prediction field data as historical input and future correction, which can effectively utilize the large language model neural network architecture to learn the data pattern of the historical input, predict the future wind power, and greatly improve the prediction accuracy.

[0059] 2. In order to solve the information coupling problem caused by the fusion of physical prediction field data, the present application proposes an adaptive attention decomposition method for multivariate information decoupling. This method can effectively solve the information coupling and high-dimensional evolution complexity caused by the coupling of multiple variable information in a traditional model, effectively aggregate the relevance between multiple variables, separate the embedding code, and improve the prediction performance.

[0060] 3. In order to solve the problem of insufficient development of the ability of the existing large language model method, the present application uses an autoregressive prediction method to predict future wind power, which is consistent with the pre-training method of the large language model, and fully develops the ability of the large language model. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0062] Figure 1 A block diagram of a large language model driven wind power prediction method fusing a physical prediction field provided by an embodiment of the present application;

[0063] Figure 2 A flowchart of a large language model driven wind power prediction method fusing a physical prediction field provided by an embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0065] The embodiments of the present application provide a large language model driven wind power prediction method fusing a physical prediction field, which solves the technical problem of difficulty in perceiving the relevance of variables and comprehensively improves the accuracy of wind power prediction.

[0066] The technical solutions in the embodiments of the present application are as follows to solve the above technical problems:

[0067] The embodiments of the present application use a large language model to fuse a physical prediction field for autoregressive prediction of wind power. Specifically:

[0068] (1) The large language model neural network architecture adopted is superior to general neural network methods and effectively improves the prediction accuracy. It effectively solves the problem that the prediction ability of the neural network structure used in the existing wind power prediction method is weak, cannot fully perceive the time dependence and variable relevance of wind power, and has insufficient modeling ability for complex relationships.

[0069] (2) The joint prediction problem of the physical prediction field and historical wind power information is focused on, which can not only perform high-precision prediction but also perform medium and short-term wind power prediction (24 hours to 168 hours). It solves the problem that most of the existing wind power prediction technologies either directly give power prediction based on physical prediction data or directly give prediction based on historical data. The former ignores the time dependence and has poor prediction accuracy; the latter ignores the variable relevance and can only be used for ultra-short-term prediction (less than or equal to 4 hours).

[0070] (3) A special multi-variable fusion method, namely the Adaptive Attention Decomposition (AAD) method, is proposed, which can effectively decouple multiple input variables, aggregate associated information, simplify the time evolution of embedding, and thus improve the performance of wind power prediction. The existing technology fuses the embedding codes of multiple climate variables into one input into the neural network for prediction. This method highly couples variable information and complicates the evolution of embedding, making it difficult for the model to perceive the relevance of variables. Therefore, this method cannot simultaneously input too many variables, and the use of physical prediction field information is insufficient.

[0071] (4) The autoregressive prediction method is used to predict future wind power, which is consistent with the pre-training method of large language models, fully developing the capabilities of large language models. The existing technology based on large language models is mostly based on non-autoregressive prediction methods, i.e., predicting future N-hour wind power at once. This method cannot effectively utilize the prediction capabilities of large language models and does not conform to the assumed usage during the training of large language models.

[0072] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0073] First of all, the term "physical prediction field" in the embodiments of the present application specifically refers to the field information obtained by simulating and predicting the distribution of atmospheric data variables such as wind direction, wind speed, temperature, humidity, etc. at a specific time and space based on atmospheric physical processes and equations through numerical weather prediction models and other means.

[0074] Embodiment 1:

[0075] As shown in Figure 1 The embodiments of the present application provide a large language model driven wind power prediction method fusing physical prediction field, which comprises:

[0076] S1, obtaining and preprocessing the historical power of the wind farm, the historical climate variables and future climate variables corresponding to the physical prediction field data;

[0077] S2, constructing a historical input based on the historical power of the wind farm and the historical climate variables, and constructing a future input based on the future climate variables;

[0078] S3, non-overlappingly segmenting the historical input and the future input, and performing feature extraction and semantic encoding on the segmented variables to obtain corresponding historical embedding codes and future embedding codes, respectively;

[0079] S4, adaptively decompose the historical embedding code and the future embedding code, obtain the corresponding decoupled embedding codes through position encoding and multi-head scaled dot-product attention mechanism, respectively;

[0080] S5, input the historical decoupled embedding code into the pre-trained large language model, predict the decomposed embedding code at the future time, and merge it;

[0081] S6, correct the merged embedding code by fusing the future decoupled embedding code through multi-head cross-attention mechanism, to decode the predicted power of the wind farm at the future time;

[0082] S7, splice the predicted power after recoding with the corresponding historical embedding code to obtain new historical embedding code, introduce a new future embedding code, repeat the above operation and splice the outputs of each time along the time dimension to obtain the final autoregressive prediction result.

[0083] The embodiment of the application fuses a large language model with a physical prediction field to predict future wind power, and proposes an adaptive attention decomposition method to semantically decompose multiple input variables to decouple information and simplify prediction. At the same time, an autoregressive prediction strategy is introduced to strengthen the development and utilization of the prediction potential of the large language model, and the accuracy of wind power prediction is comprehensively improved.

[0084] As shown in Figure 2 , a flowchart of a large language model driven wind power prediction method fusing a physical prediction field is provided. Figure 2

[0085] Next, based on the overall algorithm flow of AADAR (Adaptive Attention Decomposition, AutoRegressive, translated as 'adaptive attention decomposition, autoregressive') shown in Figure 2 , the above scheme will be described in detail:

[0086] In step S1, the historical power of the wind farm, the corresponding historical climate variables and future climate variables of the physical prediction field data are obtained and preprocessed.

[0087] On the one hand, in the wind power prediction task, the power target variable and the wind speed and wind direction in the physical prediction field have a high correlation, but in general data, the unit of wind direction is usually angle, and a large jump will occur at 360 degrees, making it difficult for the neural network to learn and perceive the variable. In order to solve this problem, the wind direction variable is first vectorized, i.e. a single angle is converted into the projection of two unit vectors of horizontal and vertical to represent, and the calculation formula is shown as (1): ​

[0088]

[0089] wherein, represents the value of the wind direction expressed in angular units, and represent the projections of the horizontal and vertical unit vectors, respectively, and cos and sin are the cosine and sine functions, respectively.

[0090] On the other hand, in order to accelerate the fitting process of the neural network model, the data is normalized to [0, 1] by the maximum-minimum normalization method in this step, ensuring that the input amplitudes of each variable tend to be consistent and will not bring difficulties to the fitting of the neural network. The calculation formula of this process is shown in equation (2):

[0091]

[0092] wherein, x is the input value, x n is the normalized value, min and max are the minimum function and the maximum function, respectively.

[0093] The above two aspects of processing are collectively referred to as the "data preprocessing" link in the embodiments of the present application, which is a preparatory work before the model predicts the wind power. In subsequent steps S2-S7: In order to improve the accuracy of future wind power prediction, AADAR introduces future physical forecast field information, effectively utilizing the correlation between meteorological data and wind power generation. AADAR adopts a large language model prediction architecture with autoregression, fully utilizing the prediction ability of large language models. In order to better capture the changes in wind power generation, the adaptive attention decomposition (AAD) method is introduced. This method decouples the embedding at the semantic level, constructs latent variables to conform to the assumption of channel independence, thereby improving the prediction accuracy. In order to deal with the problem of time sequence misalignment within the label, AADAR only introduces future information through the cross-attention mechanism after the large language model generates the preliminary prediction. Through the above methods, AADAR achieves the most advanced performance in the task of wind power prediction.

[0094] For details, see the following:

[0095] In step S2, historical inputs are constructed based on historical power of the wind farm and historical climate variables, and future inputs are constructed based on future climate variables.

[0096] The dimension of the physical forecast field meteorological variable is defined as , the dimension of the time variable is , and the wind power generation is a single-channel variable with a dimension of 1. Therefore, the historical input constructed in this step can be represented as . Among them, represents the real set; Mis the lookback range, representing the past time steps for prediction. Since the input variables for future time steps are unknown for wind power generation, the future input constructed in this step can be represented as where N is the prediction range, and the embedding length L is consistent in the model. In addition, the target variable of wind power generation is defined as , which has the same length as .

[0097] In step S3, the historical input and the future input are non-overlappingly partitioned, and the partitioned variables are feature extracted and semantically encoded to obtain corresponding historical embedding codes and future embedding codes, respectively.

[0098] In this step, the input data is non-overlappingly partitioned with an interval of 24 hours ( L = 24). For any input variable , where V is the variable dimension, T is the time dimension, it is first partitioned in the time dimension into T = P x L , where P is the number of embedding partitions, L is the embedding length, and the partitioned variable is obtained.

[0099] Then, the variable is feature extracted and semantically encoded according to formula (3) using a multi-layer perception:

[0100]

[0101] where EncoderMLP is a multi-layer perception function, E is the output code, D is the semantic encoding dimension.

[0102] The historical input and the future input are encoded according to formula (3) to obtain corresponding historical embedding codes and future embedding codes .

[0103] In step S4, the historical embedding codes and the future embedding codes are adaptively decomposed into attention, and the corresponding decoupled embedding codes are obtained through position encoding and multi-head scaled dot-product attention mechanism.

[0104] In this step, the historical embedding codes and the future embedding codes are adaptively decomposed into attention to obtain and , respectively. Wherein, The length of the history input embedding. In addition, since AADAR is a self-recursive prediction method, the length of the future information is only 1. This process can be represented by formula (4):

[0105]

[0106] wherein, and represent the adaptive attention decomposition modules of the history input and the future input, respectively, and the weights of the two are not shared.

[0107] It should be noted that the above adaptive attention decomposition method can effectively decouple multiple input variables, aggregate associated information, simplify the time evolution of the embedding, and thus improve the performance of wind power prediction.

[0108] In an optional embodiment, for an input embedding E, represented as a sequence P along the time dimension , the related steps of the adaptive attention decomposition include:

[0109] First, by introducing a learnable embedding, the adaptive attention decomposition calculates the decomposed embedding using a multi-head scaled dot-product attention mechanism, as shown in formula (5):

[0110]

[0111] wherein, represents the learnable embedding, represents the set of real numbers, K represents the number of decomposed embeddings, D represents the vector dimension; represents the position embedding, V represents the number of input variables; Q, K, and V represent the query matrix, the key matrix, and the value matrix, respectively; matrices W Q , W K , and W V all represent learnable weight matrices for projecting the input embedding and the learnable embedding into the attention space; D' represents the dimension of the key space; i represents the index of the input embedding.

[0112] Second, the decomposed decoupled embedding encoding is calculated using attention, as shown in formula (6):

[0113]

[0114] wherein, is the decoupled variable encoding; Attention represents an attention function; softmax represents an activation function; superscript T represents a transpose.

[0115] Finally, the respective input embeddings are collected are recombined along the time dimension P to obtain the final decoupled embedding encoding .

[0116] In step S5, the historical decoupled embedding encoding is input into the pre-trained large language model to predict the decomposition embedding encoding at the future time, and the decomposition embedding encoding is combined.

[0117] In this step, the decoupled embedding is predicted by the large language model (corresponding to the linear dimension matching layer in Figure 2 ), and the predicted embedding encoding at the future time is given.

[0118] In order to enable the embedding to be accepted by the large language model, the dimension of the embedding must be aligned with the large language model, and therefore a linear mapping layer is used to align the dimension, as shown in equation (7):

[0119]

[0120] wherein represents the embedded embedding after dimension increase, LinearUp represents a dimension increase linear mapping function; represents the historical decoupled embedding encoding.

[0121] The is input into the large language model to obtain the prediction result at the future time , as shown in equation (8):

[0122]

[0123] wherein LLMPredict represents a large language model prediction function.

[0124] Since is the prediction result under the dimension of the large model, it needs to be mapped to a lower dimension, similar to equation (7), as shown in equation (9):

[0125]

[0126] wherein LinearDown represents a dimension reduction linear mapping function; represents the embedded embedding after dimension reduction, since the autoregressive prediction method is used in the embodiment of the present application, represents the decomposition embedding encoding at the future time.

[0127] It should be noted that the large language model solves the problem of insufficient prediction performance of the existing neural network model, improves the complex relationship modeling capability of the model, especially the time dependence and variable correlation. Moreover, the large language model applicable here has multiple, and any one large language model such as GPT-2 or Llama can achieve good performance. Therefore, the embodiments of the present application do not need to limit the type of large language model, and those skilled in the art can select it according to their needs.

[0128] Further, since the large language model has completed the prediction, the decomposed embedding codes can be combined into one embedding code for subsequent decoding work.

[0129] Correspondingly, the step also adopts a linear mapping layer to combine the decomposed embedding codes, as shown in formula (10):

[0130]

[0131] wherein, represents the combined prediction embedding, LinearMerge represents the linear mapping layer used to combine the embedding.

[0132] In step S6, the future decoupled embedding codes are fused through the multi-head cross attention mechanism to correct the combined embedding codes, so as to decode and obtain the predicted power of the wind farm at the future time.

[0133] The future physical forecast information is introduced as a correction in this step, aiming to greatly improve the prediction accuracy. Specifically, the future forecast information is fused through the multi-head cross attention mechanism for correction, and the related steps include:

[0134] First, the cross attention matrix is calculated, as shown in formula (11):

[0135]

[0136] wherein, Q c , K c and V c represent the query matrix, key matrix and value matrix of cross attention, all represent the learnable weight matrix thereof, represents the attention dimension thereof; represents the combined prediction embedding (i.e. future forecast information); represents the future decoupled embedding code, is the only element thereof; represents the real set.

[0137] Secondly, cross attention calculation is performed, as shown in formula (12):

[0138]

[0139] wherein, denotes the revised embedding code; CrossAttention denotes the cross-attention function; softmax denotes the activation function; the superscript T denotes transposition.

[0140] Finally, a multi-layer perceptron decoder is used to decode the revised embedding code to obtain the final predicted power , as shown in formula (13):

[0141]

[0142] wherein, MLPDecoder is a multi-layer perceptron decoding function.

[0143] In step S7, the predicted power is re-encoded and spliced with the corresponding historical embedding code to obtain a new historical embedding code, a new future embedding code is introduced, the above operations are repeated, and the outputs of each time are spliced along the time dimension to obtain the final autoregressive prediction result.

[0144] To realize autoregressive prediction, the predicted power is re-encoded, and the corresponding future input is spliced in the first dimension to obtain a new historical embedding code, and then a new future input corresponding to the future embedding code is introduced, steps S3 to S6 are repeated, and the outputs of each time are spliced along the time dimension to obtain the final autoregressive prediction result.

[0145] At this point, the embodiment of the present application has completed the entire process of the wind power prediction method driven by the large language model and integrated with the physical prediction field.

[0146] Embodiment 2:

[0147] The embodiment of the present application provides a wind power prediction system driven by a large language model and integrated with a physical prediction field, comprising:

[0148] An acquisition module is configured to acquire and preprocess historical power of a wind farm, historical climate variables and future climate variables corresponding to physical prediction field data.

[0149] A construction module is configured to construct historical input based on historical power and historical climate variables of the wind farm, and construct future input based on future climate variables.

[0150] The encoding module is configured to perform non-overlapping segmentation on the historical input and the future input, and perform feature extraction and semantic encoding on the segmented variables to obtain corresponding historical embedding encoding and future embedding encoding respectively.

[0151] The decoupling module is configured to perform adaptive attention decomposition on the historical embedding encoding and the future embedding encoding, and obtain corresponding decoupled embedding encoding through position encoding and multi-head scaled dot-product attention mechanism.

[0152] The prediction module is configured to input the historical decoupled embedding encoding into a pre-trained large language model to predict the decomposition embedding encoding at the future time, and combine the decomposition embedding encoding.

[0153] The correction module is configured to correct the combined embedding encoding by fusing the future decoupled embedding encoding through multi-head cross-attention mechanism, so as to decode the predicted power of the wind farm at the future time.

[0154] The autoregressive module is configured to splice the predicted power after re-encoding with the corresponding historical embedding encoding to obtain new historical embedding encoding, introduce a new future embedding encoding, repeat the above operations, and splice the outputs of each time along the time dimension to obtain the final autoregressive prediction result.

[0155] Embodiment 3:

[0156] The embodiment of the present application provides a storage medium which stores a computer program for large language model driven wind power prediction of fusion physical prediction field, wherein the computer program enables a computer to execute the wind power prediction method as described in embodiment 1.

[0157] Embodiment 4:

[0158] The embodiment of the present application provides an electronic device, comprising:

[0159] one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the program comprises a program for executing the wind power prediction method as described in embodiment 1.

[0160] It can be understood that the large language model driven wind power prediction system of fusion physical prediction field, the storage medium and the electronic device provided by the embodiment of the present application correspond to the large language model driven wind power prediction method of fusion physical prediction field provided by the embodiment of the present application, and the explanation, examples and beneficial effects of the related content can refer to the corresponding part in the wind power prediction method, which will not be repeated here.

[0161] In summary, compared with the prior art, the present application has the following advantages:

[0162] 1. To increase the length of the wind power prediction interval, the embodiment of the present application fuses physical prediction field data as historical input and future correction, which can effectively utilize the large language model neural network architecture to learn the data pattern of the historical input and predict the future wind power, achieving a significant improvement in prediction accuracy.

[0163] 2. To solve the information coupling problem caused by the fusion of physical prediction field data, the embodiment of the present application proposes an adaptive attention decomposition method for multivariate information decoupling. This method can effectively solve the information coupling and high-dimensional evolution complexity caused by the coupling of multiple variable information in a traditional model, effectively aggregate the relevance between multiple variables, separate the embedding encoding, and improve the prediction performance.

[0164] 3. To solve the problem of insufficient development of the capabilities of the existing large language model method, the embodiment of the present application uses an autoregressive prediction method to predict future wind power, which is consistent with the pre-training method of the large language model and fully develops the capabilities of the large language model.

[0165] It should be noted that, in this paper, relational terms such as first and second are used merely to distinguish one entity or action from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or actions. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device 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 device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0166] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the foregoing embodiments of the present application have been described in detail, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A wind power prediction method driven by a large language model and fusing a physical prediction field, characterized in that, The method comprises the following steps: acquiring and preprocessing historical climate variables corresponding to historical power and physical forecast field data of a wind farm and future climate variables; constructing historical input based on historical power and historical climate variables of the wind farm, and constructing future input based on the future climate variables; performing non-overlapping segmentation on the historical input and the future input, and performing feature extraction and semantic coding on the segmented variables to obtain corresponding historical embedding codes and future embedding codes; performing adaptive attention decomposition on the historical embedding codes and the future embedding codes, and obtaining corresponding decoupled embedding codes through position coding and multi-head scaled dot-product attention mechanism; inputting the historical decoupled embedding codes into a pre-trained large language model to predict the decomposition embedding codes at a future time, and merging the decomposition embedding codes; correcting the merged embedding codes through multi-head cross-attention mechanism by fusing the future decoupled embedding codes, and decoding to obtain the predicted power of the wind farm at the future time; For an input embedding E, denoted as a sequence along the time dimension P ;​ the adaptive attention decomposition comprises: the adaptive attention decomposition uses multi-head scaled dot-product attention mechanism to calculate the decomposed embedding by introducing a learnable embedding, as shown in the following formula: wherein, denotes a learnable embedding, denotes the set of real numbers, K denotes the number of decomposed embeddings, D denotes the vector dimension; denotes a position embedding, V denotes the number of input variables; Q, K and V denote a query matrix, a key matrix and a value matrix, respectively; the matrices W Q , W K , W V all denote learnable weight matrices for projecting the input embeddings and the learnable embeddings into an attention space; D' denotes the dimension of the key space; i denotes an index of the input embedding; the decomposed decoupled embedding codes are calculated using attention, as shown in the following formula: wherein, is the decoupled variable encoding; Attention denotes an attention function; softmax denotes an activation function; superscript T denotes a transpose; collecting the respective input embeddings corresponding decoupled embeddings along the time dimension P recombining, obtaining final decoupled embedding encoding .

2. The wind power prediction method of claim 1, wherein, after decoding to obtain the predicted power of the wind farm at the future time, the wind power prediction method further comprises: splicing the predicted power after recoding with the corresponding historical embedding codes to obtain new historical embedding codes, introducing a new future embedding code, repeating the above operations and splicing the outputs of each time along the time dimension to obtain the final autoregressive prediction result.

3. The wind power prediction method of claim 1, wherein, the inputting the historical decoupled embedding codes into a pre-trained large language model to predict the decomposition embedding codes at a future time comprises: aligning the dimensions using a linear mapping layer, as shown in the following formula: wherein, represents the embedding after dimensionality increase, LinearUp represents the linear mapping function of dimensionality increase; represents the decoupled embedding encoding of history; will be described below. as an input of a pre-trained large language model to obtain a prediction result of a future time as shown in the following formula: wherein, LLMPredict represents a large language model prediction function; performing dimension reduction mapping on the prediction result, as shown in the following formula: wherein LinearDown denotes a dimensionality reduction linear mapping function; denotes the embedding after dimensionality reduction, denotes the decomposed embedding encoding for the future time instant.

4. The wind power prediction method of claim 1, wherein, merging the decomposition embedding codes using a linear mapping layer.

5. The wind power prediction method of claim 1, wherein, the correcting the merged embedding codes through multi-head cross-attention mechanism by fusing the future decoupled embedding codes comprises: calculating the cross-attention matrix, as shown in the following formula: where Q c , K c and V c denote the query, key and value matrices of cross-attention, respectively, denote their learnable weight matrices, denote their attention dimensions; denote the merged predicted embedding; denote the future decoupled embedding encoding, is its unique element; denotes the set of real numbers; performing cross-attention calculation, as shown in the following formula: wherein, denotes the modified embedded code; CrossAttention denotes a cross-attention function; softmax denotes an activation function; superscript T denotes the transpose.

6. The wind power prediction method of claim 1, wherein the feature extraction and semantic coding of the segmented variables are performed using a multi-layer perceptron; and / or the modified embedding codes are decoded using a multi-layer perceptron decoder.

7. A wind power prediction system driven by a large language model and fusing physical prediction fields, characterized in that, The method comprises the following steps: an acquisition module for acquiring and preprocessing historical climate variables corresponding to historical power and physical forecast field data of a wind farm and future climate variables; a construction module for constructing historical input based on historical power and historical climate variables of the wind farm, and constructing future input based on the future climate variables; an encoding module for performing non-overlapping segmentation on the historical input and the future input, and performing feature extraction and semantic coding on the segmented variables to obtain corresponding historical embedding codes and future embedding codes; A decoupling module is configured to perform adaptive attention decomposition on the historical embedding code and the future embedding code, and obtain corresponding decoupled embedding codes through position encoding and multi-head scaled dot-product attention mechanism, respectively; A prediction module is configured to input the historical decoupled embedding code into a pre-trained large language model, predict the decomposition embedding code at the future time, and merge the decomposition embedding code; A correction module is configured to fuse the future decoupled embedding code through multi-head cross-attention mechanism, correct the merged embedding code, and decode the predicted power of the wind farm at the future time; An autoregressive module is configured to splice the predicted power after recoding with the corresponding historical embedding code, obtain new historical embedding code, introduce a new future embedding code, repeat the above operations, and splice the outputs of each time along the time dimension to obtain the final autoregressive prediction result.

8. A storage medium, characterized by The computer program for driving the large language model to fuse the physical prediction field of the wind power prediction, wherein the computer program enables the computer to perform the wind power prediction method according to any one of claims 1-6.

9. An electronic device, comprising: Comprise: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the program comprising a wind power prediction method for executing any one of claims 1-6.

Images