Wind power prediction method and device based on pre-trained big language model

Abstract

The invention relates to a wind power prediction method and device based on a pre-trained large language model, and belongs to the field of wind power generation, and the method comprises the steps: dividing a wind power data sequence into a plurality of time sequence patches; dividing the time sequence patch into a plurality of equal-length subsequences; extracting data features of the equal-length subsequences; converting the data features into semantic cue words; adding a text instruction to the semantic prompt word to obtain a semantic instruction; splicing the time sequence patch and the corresponding position code, time code and semantic instruction to obtain a fusion sample corresponding to the time sequence patch; inputting the plurality of fusion samples into a large language model for pre-training, and performing parameter adjustment on the large language model to obtain a wind power prediction model; and inputting the fusion sample of the new wind power data into the wind power prediction model, and outputting a wind power prediction value. According to the method, under the condition that the number of historical operation data samples is small, the large language model is pre-trained through the fusion samples, and high-precision wind power prediction is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and in particular to a wind power prediction method and device based on a pre-trained large language model. Background Art

[0002] As the global energy landscape shifts toward low- and zero-carbon energy, wind power, a key pillar of this transition, is experiencing rapid growth in newly installed capacity. Accurately forecasting the wind power output of newly constructed wind farms is crucial. This helps reduce grid instability and risk caused by wind power fluctuations, effectively mitigating the negative impact of wind power uncertainty on the grid.

[0003] Currently, most mainstream wind power forecasting methods rely on extensive historical power data for training and supervision. However, the limited historical operational data samples available during the initial operation phase of newly built wind farms make traditional forecasting models susceptible to performance bottlenecks. This significantly reduces the accuracy of current wind power forecasting methods, making high-precision wind power forecasting extremely difficult.

[0004] Large language models (LLMs) have remarkable few-shot learning capabilities. However, LLMs cannot be directly applied to wind power forecasting for several reasons: Firstly, LLMs are designed to process discrete tokens, while historical wind farm operating data is continuous time series data. The data structures of the two differ significantly, making it impossible to directly apply LLM data processing methods to wind power forecasting. Secondly, LLMs do not learn how to interpret time series patterns or perform related reasoning during pre-training, making it difficult to understand temporal information, which is crucial for time series forecasting. Summary of the Invention

[0005] In order to solve the above problems and predict wind power when there are few historical operation data samples of wind farms, the applicant has made the present invention, providing a wind power prediction method and device based on a pre-trained large language model.

[0006] In a first aspect, the present invention provides a wind power prediction method based on a pre-trained large language model, comprising: The wind power data sequence is divided into multiple time series patches; the position code corresponding to the time series patch is determined according to the sequence position of the wind power data in the time series patch; the time code corresponding to the time series patch is determined according to the timestamp of the wind power data in the time series patch; the time series patch is divided into multiple equal-length subsequences; the data features of each equal-length subsequence are extracted; the data features are converted into semantic prompt words that can be understood by a large language model; text instructions are added to the semantic prompt words to obtain semantic instructions; the time series patch and the corresponding position code, time code and semantic instruction are spliced to obtain a fused sample corresponding to the time series patch; the multiple fused samples are input into the large language model for pre-training, and the parameters of the large language model are adjusted to obtain a wind power prediction model; the fused sample of new wind power data is input into the wind power prediction model, and the wind power prediction model outputs a wind power prediction value.

[0007] In a second aspect, the present invention provides a wind power prediction device based on a pre-trained large language model, comprising: A data fusion module is configured to divide a wind power data sequence into a plurality of time series patches; determine a position code corresponding to the time series patch according to the sequence position of the wind power data in the time series patch; determine a time code corresponding to the time series patch according to the timestamp of the wind power data in the time series patch; divide the time series patch into a plurality of equal-length subsequences; extract data features of each equal-length subsequence; convert the data features into semantic prompt words understandable by a large language model; add text instructions to the semantic prompt words to obtain semantic instructions; and splice the time series patch and the corresponding position code, time code, and semantic instruction to obtain a fused sample corresponding to the time series patch; The model optimization module is used to input multiple fusion samples into the large language model for pre-training and adjust the parameters of the large language model to obtain a wind power prediction model; The power prediction module is used to input the fusion sample of the new wind power data into the wind power prediction model, and the wind power prediction model outputs the wind power prediction value.

[0008] Based on the same inventive concept, the present invention also provides a method for constructing a wind power prediction model based on a pre-trained large language model, comprising the following steps: The wind power data sequence is divided into multiple time series patches; the position code corresponding to the time series patch is determined according to the sequence position of the wind power data in the time series patch; the time code corresponding to the time series patch is determined according to the timestamp of the wind power data in the time series patch; the time series patch is divided into multiple equal-length subsequences; the data features of each equal-length subsequence are extracted; the data features are converted into semantic prompt words that can be understood by a large language model; text instructions are added to the semantic prompt words to obtain semantic instructions; the time series patch and the corresponding position code, time code and semantic instruction are spliced to obtain a fused sample corresponding to the time series patch; multiple fused samples are input into a large language model for pre-training, and the parameters of the large language model are adjusted to obtain a wind power prediction model.

[0009] Based on the same inventive concept, the present invention also provides a wind power prediction model based on a pre-trained large language model, comprising: A sequence discretization module is used to divide a wind power data sequence into multiple time series patches; determine the position code corresponding to the time series patch according to the sequence position of the wind power data in the time series patch; determine the time code corresponding to the time series patch according to the timestamp of the wind power data in the time series patch; divide the time series patch into multiple equal-length subsequences; extract data features of each equal-length subsequence; convert the data features into semantic prompt words that can be understood by a large language model; add text instructions to the semantic prompt words to obtain semantic instructions; splice the time series patch and the corresponding position code, time code and semantic instruction to obtain a fused sample corresponding to the time series patch; The pre-training and parameter adjustment module is used to input multiple fusion samples into the large language model for pre-training and adjust the parameters of the large language model to obtain a wind power prediction model.

[0010] Based on the same inventive concept, the present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the aforementioned wind power prediction method based on a pre-trained large language model, or implements the aforementioned wind power prediction model construction method based on a pre-trained large language model.

[0011] Based on the same inventive concept, the present invention also provides a computer storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed, the aforementioned wind power prediction method based on the pre-trained large language model is implemented, or the aforementioned wind power prediction model construction method based on the pre-trained large language model is implemented.

[0012] The beneficial effects of the above technical solution provided by the present invention include at least: The aforementioned wind power prediction method based on a pre-trained large language model can provide discrete fused samples suitable for large language model input. These fused samples include wind power data, positional encoding that preserves the relative position information of the time series, time encoding that reflects the dynamic characteristics of the time dimension, and semantic instructions that carry the high-level semantic information of the time series. These discrete fused samples not only integrate the local and global characteristics of the time series but also simultaneously capture dynamic temporal information and semantic context, providing rich input expressions for subsequent prediction tasks and enabling the wind power prediction model to better learn temporal dependencies. Even with a limited number of historical operating data samples, the fused samples can be used to pre-train the large language model, achieving highly accurate wind power predictions.

[0013] The above-mentioned method for constructing a wind power prediction model based on a pre-trained large language model can provide discrete fusion samples suitable for large language model input. The fusion samples include wind power data, position codes that retain the relative position information of the time series, time codes that reflect the dynamic characteristics of the time dimension, and semantic instructions that carry the high-level semantic information of the time series. The discrete fusion samples not only integrate the local and global characteristics of the time series, but also simultaneously capture dynamic time information and semantic context, providing rich input expressions for subsequent prediction tasks and allowing the wind power prediction model to better learn time dependencies. The wind power prediction model constructed using the above method can achieve high-precision wind power prediction when there are fewer historical operation data samples.

[0014] Other features and advantages of the present invention will be set forth in the description that follows.

[0015] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of a wind power prediction method based on a pre-trained large language model in an embodiment of the present invention; Figure 2 This is one of the schematic diagrams of the PowerMistral framework of the wind power prediction model in an embodiment of the present invention; Figure 3 This is the second schematic diagram of the PowerMistral framework of the wind power prediction model in an embodiment of the present invention; Figure 4 This is the third schematic diagram of the PowerMistral framework of the wind power prediction model in an embodiment of the present invention; Figure 5This is a schematic diagram of time code conversion in an embodiment of the present invention; Figure 6 Schematic diagram of a multi-head attention framework in an embodiment of the present invention; Figure 7 Schematic diagram of the LoRA fine-tuning process in an embodiment of the present invention; Figure 8 R for each model of 100% data set in the embodiment of the present invention 2 Comparison chart; Figure 9 R of each model of 5% data set in the embodiment of the present invention 2 Comparison chart; Figure 10 Schematic diagram of the impact of different prediction horizons on model performance in an embodiment of the present invention; Figure 11 Schematic diagram of the impact of different low-rank values on model performance in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0018] In order to solve the problems existing in the prior art, the embodiments of the present invention provide a wind power prediction method and device based on a pre-trained large language model, a method for constructing a wind power prediction model based on a pre-trained large language model, a wind power prediction model based on a pre-trained large language model, a computer program product and a computer storage medium.

[0019] The embodiment of the present invention provides a wind power prediction method based on a pre-trained large language model, the process of which is as follows: Figure 1 As shown, the following steps are included: Step S1: Divide the wind power data sequence into multiple time series patches; determine the position code corresponding to the time series patch according to the sequence position of the wind power data in the time series patch; determine the time code corresponding to the time series patch according to the timestamp of the wind power data in the time series patch; divide the time series patch into multiple equal-length subsequences; extract the data features of each equal-length subsequence; convert the data features into semantic prompt words that can be understood by a large language model; add text instructions to the semantic prompt words to obtain semantic instructions; splice the time series patch and the corresponding position code, time code and semantic instructions to obtain a fusion sample corresponding to the time series patch.

[0020] By pre-training on a large text corpus of tokens, large language models (LLMs) demonstrate remarkable few-shot learning capabilities, making them particularly suitable for scenarios where data is scarce, sensitive, or expensive. By fine-tuning an LLM, only a small dataset is required to adapt it to a specific task. Furthermore, compared to training a model from scratch, fine-tuning significantly reduces computing resource requirements because it only requires adjusting some weights and training on a smaller dataset.

[0021] The large language model is pre-trained and fine-tuned on a large-scale text corpus, capable of processing long texts and possessing strong context-awareness.

[0022] In some specific embodiments, Mistral-7B, a large language model based on the Transformer architecture, is used to perform small-sample wind power prediction. Mistral-7B constructed by the method of the present invention is called a wind power prediction model, namely PowerMistral. Mistral-7B used in this embodiment is an open source small and medium-sized LLM based on the Transformer architecture, and introduces grouped query attention (GQA) and sliding window attention (SWA) technologies, so that Mistral-7B can effectively solve the computational efficiency problem when processing long sequences while improving the reasoning speed, and significantly reduce memory requirements, making it suitable for real-time applications. Mistral-7B has 7 billion parameters and demonstrates excellent language understanding and reasoning capabilities through massive and diverse pre-training data. The model supports calls from multiple programming languages, making it easy to integrate with various systems.

[0023] Based on the powerful functions of Mistral-7B, this embodiment builds a framework for wind power prediction - PowerMistral. This framework fully utilizes the transfer learning capabilities of pre-trained LLM to effectively deal with the prediction challenges in small sample scenarios. By fine-tuning the model, training from scratch is avoided, which not only significantly reduces data requirements and computing overhead, but also ensures the accuracy of the prediction results. PowerMistral adopts a two-stage fine-tuning strategy. The framework is as follows: Figure 2 、 3 and 4. Figure 2 The following figure shows the basic framework of PowerMistral, where Q, K, and V represent the query matrix, key matrix, and value matrix in Multi-Head Attention (MHA) or Multi-Head Self Attention (MHSA), respectively. 、 、 They represent the parameters to be learned of Q, K, and V respectively. LoRA stands for low-rank fine-tuning, which is an efficient fine-tuning strategy. Figure 3 The figure shows a schematic diagram of data fusion, where the input patch represents the time series patch input for data fusion. The data sequence in the input patch is only an example, and such a time series patch of data is not necessarily used in implementation. Figure 4 The figure shows a two-stage fine-tuning strategy. The first stage is time series feature alignment, which aims to help the LLM adapt to the temporal characteristics of wind power data and ensure that the model can effectively handle time dependencies and temporal information. The second stage is forecast fine-tuning, which focuses on optimizing the model's performance in wind power time series forecasting tasks and improving forecast accuracy. In the time series feature alignment stage, InstanceNorm is used to capture the temporal patterns of the data. The error of time series patches is calculated to optimize the learning effect of time series features. In the forecast fine-tuning stage, RevIN (Reversible Instance Normalization) is introduced to restore the model output to the original data scale through a reversible mechanism, further reducing wind power forecast errors and improving model accuracy.

[0024] In the preprocessing process of wind power data series, in order to eliminate data distribution differences, the wind power data series is normalized.

[0025] After completing the normalization processing, the PatchTST method is further used to divide the wind power data series into several time series patches. Two consecutive patch areas can be overlapping or non-overlapping. The specific length is determined by the time series patch length and the horizontal sliding step size (the non-overlapping area between two consecutive patches).

[0026] In some specific embodiments, a wind power data sequence is divided into multiple time series patches, including the following steps: The wind power data sequence is divided in a non-overlapping manner to obtain N time series patches, where N = [(L - P) / S] + 2, where L represents the length of the wind power data sequence, P represents the time series patch length, and S represents the horizontal sliding step size on the wind power data sequence. (L - P) / S is rounded to the nearest integer. This segmentation method has three major advantages: 1) it aggregates the local fluctuation characteristics of the wind power data sequence into individual patches, preserving key information and compensating for the temporal dependency defects of point-by-point encoding; 2) it significantly reduces the number of input tokens, lowering computational complexity; and 3) each patch can be mapped into semantic prompts understandable by the LLM through language representation.

[0027] To enable the model to accurately capture the sequence position information in the input features, a positional encoding (PE) mechanism is introduced for each time series patch. PE uses sine and cosine functions to assign a unique position identifier to each feature, enabling the model to explicitly perceive the relative and absolute position information in the sequence. In some specific embodiments, determining the position encoding corresponding to the time series patch based on the sequence position of the wind power data in the time series patch includes the following steps: Determine the position embedding matrix ; pass , determine the position code corresponding to the timing patch, Represents a timing patch i Position encoding, pos represents the sequence position index of wind power data in the time series patch, i represents the temporal patch index, Indicates the timing patch length.

[0028] Currently, most single- or multi-variable prediction models focus on leveraging the variable's historical values to uncover dependencies across time and across variables, while neglecting the potential semantic information contained in timestamps. Timestamps are crucial indicators of the context of a time series, indirectly reflecting the fluctuation range of a variable's observed value at a given moment. For example, peaks and valleys in power load often occur at specific times of the day. Similarly, in meteorological terms, during the day, the ground is heated by strong solar radiation, resulting in frequent heat exchange with the atmosphere, enhanced turbulence, and relatively high wind speeds. At night, solar radiation weakens, the atmosphere becomes more stable, and wind speeds are generally lower. Based on the above analysis, timestamps are a potential constraint that limits or determines the observed values of variables. Combinations of different timestamps correspond to certain observed values of variables in a multidimensional vector space.

[0029] To fully exploit the potential information contained in timestamps, this embodiment uses a linear layer to perform temporal encoding on historical timestamps (minutes, hours, days, weeks) and map them to a latent space, where the encoded timestamps are input into the model.

[0030] In some specific embodiments, such as Figure 5 As shown, determining the time code corresponding to the time series patch according to the timestamp of the wind power data in the time series patch includes the following steps: Perform the first linear transformation on the timestamps of the wind power data in the time series patch to obtain the time data after dimensionality reduction; Normalizing and regularizing the time data and then performing convolution processing to obtain time data features; The time data feature is subjected to a second linear transformation to obtain a time code corresponding to the time series patch.

[0031] The normalized time series patch is represented by vector and further divided into I non-overlapping subsequences of equal length .in, Each timing patch Convert to basic embedded code through Patch Embedder In addition, by Extract local features and convert them into semantic information understandable by LLM, including data trends (rise, fall, steady, stop) and digital description (e.g., maximum, minimum, average, variance). Patch Embedder refers to an embedding coding method or a program corresponding to an embedding coding method.

[0032] In some specific embodiments, converting the data features into semantic prompt words understandable by a large language model includes the following steps: Call the attention mechanism in the Transformer library to respectively i Data trends Described with numbers Converted into semantic clues that can be understood by the large language model, where , , Represents a timing patch i The normalized vector, Express Extract data trend features, Express Perform digital description feature extraction, Express Extract the maximum value. Express Perform minimum value extraction. Express Extract the average value, Express Perform variance extraction.

[0033] On this basis, by adding text instructions (such as "text instructions for the data trend of this time segment" and "text instructions for the digital description of this time segment"), two complete semantic instructions are integrated. and .

[0034] In some specific embodiments, splicing the timing patch and the corresponding position code, time code, and semantic instruction to obtain a fused sample corresponding to the timing patch includes the following steps: pass , get the timing patch i The corresponding fusion samples, where Represents a timing patch i The fusion sample, Represents a splicing operation, Represents a timing patch i The basic embedding code of Represents a timing patch i Positional encoding, Represents a timing patch i Time code, Represents a timing patch i Semantic instructions for data trends, Represents a timing patch i Semantic instructions for digital descriptions.

[0035] The above fusion sample comprehensively represents the wind power data sequence from multiple perspectives, and its embedding includes the following four parts. 1) Basic features , retaining the local pattern of the wind power segment; 2) semantic feature embedding and , carrying high-level semantic information of the time series; 3) Position embedding , retaining the relative position information of time series segments; 4) Time embedding , reflecting the dynamic characteristics of the time dimension. Through this fusion method, we can not only integrate the local and global features of the time series, but also capture the dynamic time information and semantic context at the same time, providing rich input expressions for subsequent prediction tasks.

[0036] Step S2: Input multiple fusion samples into a large language model for pre-training, and adjust the parameters of the large language model to obtain a wind power prediction model.

[0037] The core architecture of Mistral-7B is based on the Transformer model with attention mechanism. The specific structural framework is shown in Figure 2 The model primarily consists of a multi-head self-attention mechanism (MHSA) and a position-based feed-forward network (FFN). To improve computational efficiency, the attention mechanism incorporates sliding window attention (SWA) and grouped query attention (GQA) for optimization. Furthermore, the model utilizes residual connections to ensure efficient information transfer and incorporates layer normalization (LayerNorm) to enhance training stability and convergence.

[0038] like Figure 6 As shown in Figure 2, Multi-Head Attention (MHA) is a core component of the Transformer architecture, used to capture correlations between different parts of the input data in sequence modeling. Its core approach is to parallelize the computation of multiple attention heads, enabling the model to focus on different features of the input data in different subspaces, thereby improving the model's ability to capture complex patterns.

[0039] If given H Attention heads, the calculation process of the multi-head attention mechanism is as follows: , , Where, 、 、 are the query matrix, key matrix, and value matrix of the input of the multi-head attention mechanism, to For independent attention heads; is the output projection matrix, 、 、 are the query matrix, key matrix, and value matrix corresponding to the h-th group of attention respectively. represents multi-head attention data, Represents a splicing operation, Represents the calculation of the attention head function.

[0040] A single attention head uses Scaled Dot-Product Attention, and the calculation formula is as follows: , Where, , , , 、 、 is the weight matrix; is the dimension of the query vector; is the input matrix of the multi-head attention mechanism.

[0041] As a parallel module of the multi-head self-attention mechanism, the position feedforward network acts on each sequence position and performs nonlinear transformation and normalization on the features through a two-layer fully connected network and a nonlinear activation function, which enhances the model's representation learning ability and enables it to better adapt to diverse input modes. The position feedforward network expression is as follows: , Where, is the input matrix of the position feedforward network, and is the weight matrix; and is the bias matrix.

[0042] The Add&Layer Norm module in the Transformer combines the dual advantages of residual connection and layer normalization, so that the output of each sub-layer can not only retain the key information of the original input, but also stabilize the feature distribution through normalization. The expression is as follows: ,

[0043] Where, It is the input matrix of Add&Layer Norm; is the output of the current sublayer; for The mean of for variance; 、 Learnable scaling and offset parameters; A smaller number, such as 0.001, ensures numerical stability. Representation layer normalization. Indicates the output of Add&Layer Norm.

[0044] In some specific embodiments, the LoRA (Low-Rank Adaptation) method is used to adapt the pre-trained LLM, which is particularly suitable for small sample scenarios. LoRA is an efficient fine-tuning strategy that freezes the pre-trained weights. , and perform low-rank decomposition on the attention layer weights to generate the parameter update matrix .like Figure 7 As shown, By low-rank matrix The product of and introduces only a small number of trainable parameters, which significantly reduces video memory and computational overhead while retaining model knowledge and avoiding the destruction of its capabilities due to parameter adjustments.

[0045]

[0046]

[0047] Where, is the new weight matrix; is the initial weight of the pre-trained model; is the incremental weight for a specific task; A and B are matrix decomposition factors respectively.

[0048] In some specific embodiments, multiple fusion samples are input into a large language model for pre-training, and the parameters of the large language model are adjusted, including the first stage and the second stage. Before dividing the wind power data series into multiple time series patches, the wind power data series is normalized; In the first stage, the offset of each temporal patch is linearly projected through the hidden features of the Transformer, and the prediction formula is: ,in, Represents the time series patch offset of the next time step in the time series predicted by the wind power prediction model, represents the true value of the corresponding time step, Represents the high-dimensional hidden features obtained by Transformer, and Loss represents and It should be noted that LoRA is only used in the pre-trained LLM of the first stage, while other modules in the first stage do not involve LoRA, but adopt the full parameter training strategy.

[0049] In the second stage, the prediction patches are spliced into prediction vectors through the Flatten operation, and the prediction vectors are rearranged through the Rearrange operation to adapt to the structure of the time series. Then, the rearranged prediction vectors are restored to the wind power data series prediction values using denormalization.

[0050] In the second stage, the model focuses on optimizing the wind power forecast value of the entire time series. Specifically, the output hidden features of the Transformer are first flattened to concatenate the patch-level prediction results into an overall prediction vector (as shown in Eq. ); Then, the Rearrange operation is applied to rearrange the flattened global prediction vector to adapt it to the structure of the specific time series. Then, the prediction results are restored to the original data distribution using RevIN. The above process is expressed as: , , , The training objective is to minimize the prediction curve With the true curve The error between the realization: , where Output time series future S A real sequence of wind power with complete time steps, is the single point prediction result of step s; is the single point true result of step s.

[0051] Step S3: inputting the fused sample of the new wind power data into the wind power prediction model, and the wind power prediction model outputs a wind power prediction value.

[0052] The above-described method of this embodiment can provide discrete fused samples suitable for input to a large language model. The fused samples include wind power data, position codes that preserve the relative position information of the time series, time codes that reflect the dynamic characteristics of the time dimension, and semantic instructions that carry the high-level semantic information of the time series. The discrete fused samples not only integrate the local and global characteristics of the time series, but also simultaneously capture dynamic temporal information and semantic context, providing rich input expressions for subsequent prediction tasks and enabling the wind power prediction model to better learn time dependencies. When historical operating data samples are scarce, the fused samples can be used to pre-train the large language model, achieving highly accurate wind power prediction.

[0053] In a specific implementation, an AutoDL deployment server was used to conduct wind power time series forecasting research based on the Mistral 7B pre-trained language model. Mistral 7B has a 32-layer Transformer encoder and supports context windows of 8192 tokens. The experimental hardware configuration consisted of an Intel Core i9-10850K CPU (3.60GHz), 64GB of RAM, and an NVIDIA GeForce RTX 3090 (24GB of VRAM). A two-stage fine-tuning strategy was employed: in the first stage, most of the Mistral 7B weights were frozen, and only the task-related LoRa adapter matrix (with a low-rank value set to 64) was updated. Attention weight updates were inserted to significantly reduce the parameter update range. Furthermore, 4-bit quantization (QLoRA) and FP16 mixed-precision training were combined to improve training efficiency. In the second stage, an 8-bit AdamW optimizer was used, with an initial learning rate of 2e-5, which was gradually reduced to 1e-6. The learning rate was dynamically adjusted using the CosineAnnealingLR strategy.

[0054] To verify the effectiveness of the proposed prediction model, operating data collected from six actual wind farms in Tianjin, China were selected for case verification, marked as WT1, WT2, WT3, WT4, WT5 and WT6. The data includes the actual power generation of each wind farm and meteorological information of wind speed at different heights from January 1, 2020 to December 31, 2020. For model training and testing, the first 70% of the data is used for training and the last 30% of the data is used for testing. All data comes from the wind farm SCADA system with a sampling interval of 10 minutes. In order to be consistent with the time granularity commonly used in actual engineering applications, the 10-minute interval data is converted into 1-hour data by taking the average of every 6 data points.

[0055] Models such as CNN, LSTM, GRU, PatchTST, FEDformer, Autoformer, Informer, and GPT4TS were selected for comparison.

[0056] CNN: A network that extracts local features from time series through convolutional and pooling layers. It is particularly good at capturing patterns with significant trends and cyclical patterns. Its local receptive field mechanism enables the model to efficiently handle short-term dependencies, while the global pooling operation further enhances the model's robustness and generalization capabilities.

[0057] LSTM: A network that uses its unique gating mechanism (such as input gate, forget gate, and output gate) to efficiently capture long-term dependency features in time series and successfully overcome the gradient vanishing problem in traditional recurrent neural networks.

[0058] GRU: A lightweight variant of LSTM that uses its simplified gating mechanism (update gate and reset gate) to efficiently capture long-term dependency features in time series.

[0059] PatchTST: A Transformer network optimized for time series data. It divides the time series into multiple subsequence blocks as input and improves prediction performance through self-supervised learning. This design not only efficiently captures local features but also reduces the computational complexity of processing long time series data.

[0060] FEDformer: A frequency-domain-enhanced Transformer model that replaces the traditional self-attention mechanism with Fourier transform and wavelet transform modules to decompose trend and periodic signals in time series. This frequency-domain enhancement enables the model to effectively capture the long-term structural information of time series.

[0061] Autoformer: A Transformer model that optimizes time series forecasting capabilities by decomposing trend and seasonal terms. It also introduces an autocorrelation mechanism to replace some self-attention modules, effectively capturing dependencies between subsequences and further improving the model's predictive performance.

[0062] Informer: A Transformer network designed for long time series. It uses a sparse self-attention mechanism to reduce computational complexity while retaining global feature modeling capabilities. It can handle extremely long sequence data and achieve a balance between performance and efficiency.

[0063] GPT4TS: A framework for applying large pre-trained language models to time series analysis. By freezing most of the model layers (such as self-attention layers and FFN layers) and only fine-tuning the position embedding and normalization layers, it retains the generalization knowledge of the original model and adapts it to time series tasks.

[0064] To ensure the robustness of the experimental results and reduce the influence of random factors, each model was run five times, and the average of the prediction results was taken as the final output. Subsequently, the prediction performance was comprehensively evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R².

[0065] We compared the proposed model with a baseline model using different training data ratios (100%, 10%, and 5%) and different datasets. The input history sequence length was 168, and the predicted future horizon was 24. Tables 1 through 3 summarize the MAE, RMSE, and R² metrics of each model on different wind farm datasets at different training data ratios.

[0066]

[0067]

[0068]

[0069] 1) 100% Data Scenario in Table 1: Given sufficient data, all models performed well, but PowerMistral maintained a clear advantage in both prediction accuracy and fitting ability. For example, in the WT1 dataset, PowerMistral achieved a MAE of 1.88 MW and a RMSE of 3.11 MW, significantly outperforming GPT4TS (2.63 MW and 3.32 MW) and the average values of Transformer models (such as PatchTST, FEDformer, and Informer) (3.1 MW and 3.9 MW). While traditional models (CNN, LSTM, and GRU) performed somewhat better, their MAE remained as high as 6.46 MW and their RMSE approached 8 MW. PowerMistral's R² exceeded 0.95 across all datasets, compared to 0.94 for GPT4TS and only 0.87 for traditional models, demonstrating that PowerMistral possesses superior prediction accuracy and pattern-capturing capabilities in large-data scenarios.

[0070] 2) 10% Data Scenario (Table 2): When the data size is reduced to 10%, PowerMistral's robustness becomes even more pronounced. For example, in the WT3 dataset, PowerMistral's MAE is 5.41 MW, approximately 12% and 33% lower than GPT4TS (6.12 MW) and the Transformer model (8.13 MW), respectively, and nearly 60% lower than the traditional model (MAE exceeding 13 MW). In terms of RMSE, PowerMistral's RMSE is 6.48 MW, 10% and 30% lower than GPT4TS (7.17 MW) and the Transformer model (9 MW), respectively, while the traditional model's MAE is over 12 MW. PowerMistral's R² value in this scenario remains above 0.88, outperforming GPT4TS (0.85), the Transformer model (0.81), and the traditional model (0.65), fully demonstrating its fitting ability and generalization performance under small sample conditions.

[0071] 3) 5% Data Scenario (Table 3): When the data size is further reduced to 5%, the R² values of traditional models (CNN, LSTM, and GRU) drop to negative values, indicating severe underfitting. While the Transformer model's R² remains positive, it is only around 0.53, significantly reducing predictive performance. In contrast, PowerMistral's R² remains above 0.84, demonstrating exceptional robustness. Taking the WT3 dataset as an example, PowerMistral's MAE and RMSE are 7.58 MW and 9.48 MW, respectively, significantly outperforming GPT4TS (7.82 MW and 9.82 MW) and the Transformer model (MAE exceeding 12 MW and RMSE exceeding 14 MW). Traditional models achieve as high as 23 MW and 29 MW, respectively. PowerMistral reduces error by approximately 33% compared to the Transformer model and by approximately 67% compared to traditional models, demonstrating significant advantages in extremely small sample scenarios.

[0072] Overall, PowerMistral, with its excellent performance under both full data volume and small sample conditions, has become the optimal model for wind farm power forecasting among the aforementioned comparison models. Compared to GPT4TS and mainstream Transformer models, PowerMistral has achieved comprehensive superiority in error control, fitting capabilities, and small sample processing, demonstrating its unique advantages and potential in the field of time series forecasting.

[0073] We conducted comprehensive ablation experiments and designed a series of model variants, including removing time embedding (E1), trend prompt embedding (E2), value prompt embedding (E3), and skipping the time alignment stage (E4) and prediction fine-tuning stage (E5). Through multi-dataset testing at 100% and 5% data scale, we demonstrated the impact of each component and training stage on model performance. The R² performance of the model is as follows: Figure 8 and Figure 9 shown.

[0074] In the ablation experiment, the wind power prediction model PowerMistral proposed in this invention has R² values of 0.9537-0.9645 and 0.8240-0.8429 in the 100% and 5% data scenarios, respectively. Figure 8), removing each module has little impact on performance. For example, on the WT5 dataset, the R² of the full model is 0.9541. Removing the time embedding (E1) reduces this to 0.9109 (a 4.53% decrease). Removing the trend prompt (E2) and the value prompt (E3) reduces the R² to 0.9134 (a 4.27% decrease) and 0.9185 (a 3.73% decrease), respectively. Removing the time alignment stage (E4) reduces the R² to 0.9248 (a 3.07% decrease). Removing the prediction fine-tuning stage (E5) has the greatest impact on performance, with the R² dropping to 0.8832 (a 7.43% decrease). Analysis shows that even with sufficient data, each module still plays a significant role in improving model performance. In particular, the prediction fine-tuning stage (E5) is crucial for its adaptability to complex tasks.

[0075] In the 5% data scenario ( Figure 9 ), the impact of each module on performance is significantly enhanced. For example, on the WT2 dataset, the R² of the wind power forecasting model PowerMistral was 0.8240. Removing the time embedding (E1) reduced it to 0.7131 (a decrease of 13.46%). Removing the trend prompt (E2) and the value prompt (E3) reduced the R² to 0.7189 (a decrease of 12.75%) and 0.7488 (a decrease of 9.10%), respectively. Skipping the time alignment stage (E4) reduced the R² to 0.6997 (a decrease of 15.08%), while removing the forecast fine-tuning stage (E5) caused performance to drop to 0.6573 (a decrease of 20.22%). On the WT4 dataset, removing the time embedding (E1) reduced the R² from 0.8423 to 0.7298, a similar impact to the time alignment stage (E4). This demonstrates that time embedding is crucial for the model to understand temporal characteristics and cyclical patterns.

[0076] Further analysis reveals that each module contributes uniquely and irreplaceably to performance improvement. In small sample size scenarios, the time embedding (E1) enhances the model's ability to capture periodic patterns by explicitly encoding temporal features; the prompting modules (E2 and E3) improve the model's perception of trends and numerical statistical properties; the time alignment stage (E4) maintains time series consistency, providing stable support for model training and prediction; and the prediction fine-tuning stage (E5), as a core module, plays a key role in adapting to small sample size distributions and improving generalization capabilities.

[0077] The above ablation experiments demonstrate the importance of each module at different data scales, confirming the rationality and complementarity of the module design. The omission of the temporal embedding, hint module, temporal alignment phase, or prediction fine-tuning phase significantly degrades model performance, further validating the model's efficiency and robustness in time series forecasting tasks.

[0078] A sensitivity analysis of the wind power prediction model PowerMistral was conducted to analyze the model performance under different prediction horizons (24, 96, 168). The results are as follows: Figure 10 As shown. Figure 10 As can be seen, the model performs exceptionally well for short time step predictions (168 / 24), with R² values remaining above 0.95 for all datasets, demonstrating the model's ability to accurately fit short-term time series characteristics. While the R² value decreases slightly with increasing prediction time steps (168 / 96 and 168 / 168), the model maintains high prediction accuracy on most datasets. In particular, the model's R² is largely unaffected by the increase in prediction time steps on the WT1 and WT4 datasets, demonstrating its robustness and superior performance when working with stable datasets.

[0079] Compare the effects of different low-rank values on the model prediction performance (e.g. Figure 11 As shown in ). The low-rank value is a key hyperparameter in the LoRA method, which determines the rank of the adapter matrix inserted into the pre-trained model weight matrix during fine-tuning. Figure 11 As can be seen, when the low-rank value is 16, the model's R² is slightly lower, decreasing by approximately 0.5%-1.0% compared to a low-rank value of 64. At a low-rank value of 64, the R² for all datasets reaches its peak, demonstrating the optimal performance balance. However, when the low-rank value is increased to 256, the prediction performance of some datasets (such as WT4, WT5, and WT6) decreases slightly, decreasing by approximately 0.2%-0.5% compared to a low-rank value of 64. This may be due to excessively high low-rank values leading to information redundancy, which affects the model's generalization ability. Overall, although the impact of changes in low-rank values on model performance is limited, choosing a moderate low-rank value (such as 64) can achieve a better balance between computing resources and performance requirements, making it a more robust configuration.

[0080] This paper proposes PowerMistral, a few-shot wind power forecasting framework based on a pre-trained large language model. Developed based on the Mistral-7B small and medium-sized pre-trained large language model, this framework efficiently completes wind power forecasting tasks without the need for training from scratch. The core approach includes a two-stage fine-tuning strategy: the first stage focuses on the temporal characteristics of wind power data, extracting key time series features through an adaptation mechanism; the second stage performs fine-tuning for specific forecasting tasks to further optimize forecasting performance. PowerMistral also encodes the temporal features of wind power series and combines them with tailored numerical features and trend feature text to effectively guide the model to improve its forecasting capabilities. The main conclusions are as follows: 1) PowerMistral demonstrates its unique advantages and potential in time series forecasting. PowerMistral surpasses baseline models in MAE, RMSE, and R² across six wind power datasets, both in full-data and small-sample scenarios. In particular, in the small-sample scenario of 5% data, PowerMistral reduces error by approximately 33% compared to the Transformer model and by approximately 67% compared to traditional models.

[0081] 2) Ablation experiments validated the design rationality of each module in the PowerMistral framework. The time embedding module effectively captures periodic features, the trend prompt and value prompt enhance trend and value perception, and the prediction fine-tuning phase improves generalization performance in small sample scenarios. Experimental results demonstrate that the synergy between these modules is the key to the performance improvement.

[0082] 3) PowerMistral demonstrates exceptional robustness across various prediction sequence step sizes, with R² consistently above 0.88, with only a slight decrease. Furthermore, selecting a moderately low rank value achieves a better balance between computational resources and performance requirements.

[0083] This embodiment proposes a Few-Shot (small sample) wind power prediction model based on a pre-trained large language model - PowerMistral. This model is developed based on the Mistral-7B small and medium-sized pre-trained large language model, and can efficiently complete the wind power prediction task without training from scratch. The core method includes a two-stage fine-tuning strategy: the first stage focuses on the time series characteristics of wind power data and extracts key time series features through an adaptation mechanism; the second stage performs fine-tuning for specific prediction tasks to further optimize the prediction performance. At the same time, PowerMistral encodes the time characteristics of the wind power series and combines tailored numerical features and trend feature text to effectively guide the model to improve its prediction capabilities. The main technical advantages of PowerMistral include: 1) PowerMistral demonstrates its unique advantages and potential in the field of time series forecasting. PowerMistral surpasses baseline models in MAE, RMSE, and R² across six wind power datasets, both in full-data and small-sample scenarios. In particular, in the small-sample scenario of 5% data, PowerMistral reduces error by approximately 33% compared to the Transformer model and by approximately 67% compared to traditional models.

[0084] 2) Ablation experiments validated the design rationality of each module in the PowerMistral framework. The time embedding module effectively captures periodic features, the trend prompt and value prompt enhance trend and value perception, and the prediction fine-tuning phase improves generalization performance in small sample scenarios. Experimental results demonstrate that the synergy between these modules is the key to the performance improvement.

[0085] 3) PowerMistral demonstrates exceptional robustness across various prediction sequence step sizes, with R² consistently remaining above 0.88 with only a slight decrease. Furthermore, selecting a moderately low rank value achieves a better balance between computational resources and performance requirements.

[0086] Those skilled in the art can change the above sequence without departing from the scope of protection of the present invention.

[0087] Another embodiment of the present invention provides a wind power prediction device based on a pre-trained large language model, comprising: A data fusion module is configured to divide a wind power data sequence into a plurality of time series patches; determine a position code corresponding to the time series patch according to the sequence position of the wind power data in the time series patch; determine a time code corresponding to the time series patch according to the timestamp of the wind power data in the time series patch; divide the time series patch into a plurality of equal-length subsequences; extract data features of each equal-length subsequence; convert the data features into semantic prompt words understandable by a large language model; add text instructions to the semantic prompt words to obtain semantic instructions; and splice the time series patch and the corresponding position code, time code, and semantic instruction to obtain a fused sample corresponding to the time series patch; The model optimization module is used to input multiple fusion samples into the large language model for pre-training and adjust the parameters of the large language model to obtain a wind power prediction model; The power prediction module is used to input the fusion sample of the new wind power data into the wind power prediction model, and the wind power prediction model outputs the wind power prediction value.

[0088] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0089] The aforementioned wind power prediction device, based on a pre-trained large language model, can provide discrete fused samples suitable for large language model input. These fused samples include wind power data, positional encoding that preserves the relative position information of the time series, time encoding that reflects the dynamic characteristics of the time dimension, and semantic instructions that carry the high-level semantic information of the time series. These discrete fused samples not only integrate the local and global characteristics of the time series but also simultaneously capture dynamic temporal information and semantic context, providing rich input expressions for subsequent prediction tasks and enabling the wind power prediction model to better learn temporal dependencies. Even with limited historical operating data samples, the fused samples can be used to pre-train the large language model, achieving highly accurate wind power predictions.

[0090] Based on the same inventive concept, the present invention also provides a method for constructing a wind power prediction model based on a pre-trained large language model, comprising the following steps: The wind power data sequence is divided into multiple time series patches; the position code corresponding to the time series patch is determined according to the sequence position of the wind power data in the time series patch; the time code corresponding to the time series patch is determined according to the timestamp of the wind power data in the time series patch; the time series patch is divided into multiple equal-length subsequences; the data features of each equal-length subsequence are extracted; the data features are converted into semantic prompt words that can be understood by a large language model; text instructions are added to the semantic prompt words to obtain semantic instructions; the time series patch and the corresponding position code, time code and semantic instruction are spliced to obtain a fused sample corresponding to the time series patch; multiple fused samples are input into a large language model for pre-training, and the parameters of the large language model are adjusted to obtain a wind power prediction model.

[0091] The above-mentioned method for constructing a wind power prediction model based on a pre-trained large language model can provide discrete fusion samples suitable for large language model input. The fusion samples include wind power data, position codes that retain the relative position information of the time series, time codes that reflect the dynamic characteristics of the time dimension, and semantic instructions that carry the high-level semantic information of the time series. The discrete fusion samples not only integrate the local and global characteristics of the time series, but also simultaneously capture dynamic time information and semantic context, providing rich input expressions for subsequent prediction tasks and allowing the wind power prediction model to better learn time dependencies. The wind power prediction model constructed using the above method can achieve high-precision wind power prediction when there are fewer historical operation data samples.

[0092] Based on the same inventive concept, the present invention also provides a wind power prediction model based on a pre-trained large language model, comprising: A sequence discretization module is used to divide a wind power data sequence into multiple time series patches; determine the position code corresponding to the time series patch according to the sequence position of the wind power data in the time series patch; determine the time code corresponding to the time series patch according to the timestamp of the wind power data in the time series patch; divide the time series patch into multiple equal-length subsequences; extract data features of each equal-length subsequence; convert the data features into semantic prompt words that can be understood by a large language model; add text instructions to the semantic prompt words to obtain semantic instructions; splice the time series patch and the corresponding position code, time code and semantic instruction to obtain a fused sample corresponding to the time series patch; The pre-training and parameter adjustment module is used to input multiple fusion samples into the large language model for pre-training and adjust the parameters of the large language model to obtain a wind power prediction model.

[0093] The aforementioned wind power prediction model based on a pre-trained large language model can provide discrete fusion samples suitable for large language model input. The fusion samples include wind power data, position codes that retain the relative position information of the time series, time codes that reflect the dynamic characteristics of the time dimension, and semantic instructions that carry the high-level semantic information of the time series. The discrete fusion samples not only integrate the local and global characteristics of the time series, but also simultaneously capture dynamic time information and semantic context, providing rich input expressions for subsequent prediction tasks and allowing the wind power prediction model to better learn time dependencies. The wind power prediction model constructed using the above method can achieve high-precision wind power prediction when there are fewer historical operation data samples.

[0094] Based on the same inventive concept, the present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the aforementioned wind power prediction method based on a pre-trained large language model, or implements the aforementioned wind power prediction model construction method based on a pre-trained large language model.

[0095] Based on the same inventive concept, the present invention also provides a computer storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed, the aforementioned wind power prediction method based on the pre-trained large language model is implemented, or the aforementioned wind power prediction model construction method based on the pre-trained large language model is implemented.

[0096] Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention shall still fall within the scope of the patent coverage of the present invention.

Claims

1. A wind power prediction method based on a pre-trained large language model, characterized in that: The following steps are involved: Divide the wind power data series into multiple time series patches; Determining a position code corresponding to the time series patch according to the sequence position of the wind power data in the time series patch; Determining a time code corresponding to the time series patch according to a timestamp of the wind power data in the time series patch; Divide the temporal patch into multiple subsequences of equal length; Extract the data features of each equal-length subsequence; Converting the data features into semantic clues that can be understood by a large language model; Adding text instructions to the semantic prompt words to obtain semantic instructions; splicing the temporal patch and the corresponding position code, time code and semantic instruction to obtain a fused sample corresponding to the temporal patch; Multiple fusion samples are input into a large language model for pre-training, and the parameters of the large language model are adjusted to obtain a wind power prediction model; The fused sample of the new wind power data is input into the wind power prediction model, and the wind power prediction model outputs a wind power prediction value.

2. The method according to claim 1, wherein Dividing the wind power data series into multiple time series patches includes the following steps: The wind power data sequence is divided into N time series patches in a non-overlapping manner, where N = [(L - P) / S] + 2, L represents the length of the wind power data sequence, P represents the length of the time series patch, and S represents the horizontal sliding step on the wind power data sequence.

3. The method according to claim 1, wherein Determining a position code corresponding to a time series patch according to a sequence position of wind power data in the time series patch includes the following steps: Determine the position embedding matrix ; pass , determine the position code corresponding to the timing patch, Represents a timing patch i Position encoding, pos represents the sequence position index of wind power data in the time series patch, i represents the temporal patch index, Indicates the timing patch length.

4. The method according to claim 1, wherein Determining the time code corresponding to the time series patch according to the timestamp of the wind power data in the time series patch includes the following steps: Perform the first linear transformation on the timestamps of the wind power data in the time series patch to obtain the time data after dimensionality reduction; Normalizing and regularizing the time data and then performing convolution processing to obtain time data features; The time data feature is subjected to a second linear transformation to obtain a time code corresponding to the time series patch.

5. The method according to claim 1, wherein Converting the data features into semantic prompt words that can be understood by a large language model includes the following steps: Call the attention mechanism in the Transformer library to respectively i Data trends Described with numbers Converted into semantic clues that can be understood by the large language model, where , , Represents a timing patch i The normalized vector, Express Extract data trend features, Express Perform digital description feature extraction, Express Extract the maximum value. Express Perform minimum value extraction. Express Extract the average value, Express Perform variance extraction.

6. The method according to claim 1, wherein The timing patch and the corresponding position code, time code and semantic instruction are spliced to obtain a fused sample corresponding to the timing patch, including the following steps: pass , get the timing patch i The corresponding fusion samples, where Represents a timing patch i The fusion sample, Represents a splicing operation, Represents a timing patch i The basic embedding code of Represents a timing patch i Positional encoding, Represents a timing patch i Time code, Represents a timing patch i Semantic instructions for data trends, Represents a timing patch i Semantic instructions for digital descriptions.

7. The method according to claim 1, wherein Multiple fusion samples are input into the large language model for pre-training, and the parameters of the large language model are adjusted, including the first and second stages. Before dividing the wind power data series into multiple time series patches, the wind power data series is normalized; In the first stage, the offset of each temporal patch is linearly projected through the hidden features of the Transformer, and the prediction formula is: ,in, Represents the time series patch offset of the next time step in the time series predicted by the wind power prediction model, represents the true value of the corresponding time step, Represents the high-dimensional hidden features obtained by Transformer, and Loss represents and The error between In the second stage, the prediction patches are spliced into prediction vectors through the Flatten operation, and the prediction vectors are rearranged through the Rearrange operation to adapt to the structure of the time series. Then, the rearranged prediction vectors are restored to the wind power data series prediction values using denormalization.

8. A wind power prediction device based on a pre-trained large language model, characterized in that: include: A data fusion module is used to divide the wind power data sequence into multiple time series patches; according to the sequence position of the wind power data in the time series patch, determine the position code corresponding to the time series patch; Determine the time code corresponding to the time series patch based on the timestamp of the wind power data in the time series patch; divide the time series patch into multiple equal-length subsequences; extract data features of each equal-length subsequence; convert the data features into semantic prompt words that can be understood by a large language model; add text instructions to the semantic prompt words to obtain semantic instructions; splice the time series patch and the corresponding position code, time code and semantic instruction to obtain a fused sample corresponding to the time series patch; The model optimization module is used to input multiple fusion samples into the large language model for pre-training and adjust the parameters of the large language model to obtain a wind power prediction model; The power prediction module is used to input the fusion sample of the new wind power data into the wind power prediction model, and the wind power prediction model outputs the wind power prediction value.

9. A method for constructing a wind power prediction model based on a pre-trained large language model, characterized in that: The following steps are involved: Divide the wind power data series into multiple time series patches; Determining a position code corresponding to the time series patch according to the sequence position of the wind power data in the time series patch; Determining a time code corresponding to the time series patch according to a timestamp of the wind power data in the time series patch; Divide the temporal patch into multiple subsequences of equal length; Extract the data features of each equal-length subsequence; Converting the data features into semantic clues that can be understood by a large language model; Adding text instructions to the semantic prompt words to obtain semantic instructions; splicing the temporal patch and the corresponding position code, time code and semantic instruction to obtain a fused sample corresponding to the temporal patch; Multiple fusion samples are input into a large language model for pre-training, and the parameters of the large language model are adjusted to obtain a wind power prediction model.

10. A wind power prediction model based on a pre-trained large language model, characterized in that: include: Sequence discretization module, used to divide the wind power data sequence into multiple time series patches; Determining a position code corresponding to the time series patch according to the sequence position of the wind power data in the time series patch; Determine the time code corresponding to the time series patch based on the timestamp of the wind power data in the time series patch; divide the time series patch into multiple equal-length subsequences; extract data features of each equal-length subsequence; convert the data features into semantic prompt words that can be understood by a large language model; add text instructions to the semantic prompt words to obtain semantic instructions; splice the time series patch and the corresponding position code, time code and semantic instruction to obtain a fused sample corresponding to the time series patch; The pre-training and parameter adjustment module is used to input multiple fusion samples into the large language model for pre-training and adjust the parameters of the large language model to obtain a wind power prediction model.

11. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, it implements the wind power prediction method based on a pre-trained large language model as described in any one of claims 1 to 7, or implements the wind power prediction model construction method based on a pre-trained large language model as described in claim 9.

12. A computer storage medium, characterized in that The computer storage medium stores computer executable instructions, which, when executed, implement the wind power prediction method based on a pre-trained large language model as described in any one of claims 1 to 7, or implement the wind power prediction model construction method based on a pre-trained large language model as described in claim 9.

Images