Fire power generation time series prediction method based on multi-modal pseudo graph large language model

Abstract

This disclosure provides a method for predicting time series data of thermal power generation based on a multimodal pseudograph large language model. It preprocesses and embeds segmented multi-channel time series data by channel independence, decomposes the time-domain signal into multi-frequency sub-sequences using Fast Fourier Transform, and reconstructs them into a two-dimensional pseudograph. Frequency domain features are extracted using a visual transformer, and global semantic alignment of the time-domain and frequency-domain embedding vectors is performed using a cross-modal predictor and contrastive loss. Simultaneously, dataset metadata, domain knowledge, statistics, and dominant frequency components are concatenated into a cue prefix, which, along with the time-frequency multimodal embedding, is input into a pre-trained large language model for prediction. This effectively integrates multimodal operating parameters and domain knowledge, significantly improving the prediction accuracy and long-term stability of key operating parameters of thermal power generating units. Its superiority over existing methods has been verified on multiple public datasets and actual thermal power unit data, while also enhancing the model's interpretability and industrial deployment adaptability.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and time series prediction technology for power systems, and in particular to a method for predicting the time series of thermal power generation based on a multimodal pseudograph large language model. Background Technology

[0002] Thermal power generation (including coal-fired, gas-fired, and biomass power generation) is a core component of the global power system, undertaking important tasks in base load and peak / frequency regulation. However, with the large-scale grid connection of renewable energy sources such as wind and solar power, the power grid places higher demands on the flexibility, response speed, and operational stability of thermal power generating units. Thermal power generating units typically operate under high-frequency variable load conditions, and their combustion processes, steam cycles, and boiler hydrodynamics exhibit complex characteristics such as strong nonlinearity, large time lag, and multivariate coupling. These characteristics make it difficult for traditional mechanistic modeling and physical simulation methods to accurately and in real-time reflect the dynamic behavior of the units. Therefore, utilizing data-driven methods, especially deep learning models, to perform high-precision long-term predictions of the operating parameters of thermal power generating units (such as boiler load, main steam temperature, drum water level, and BTU calorific value) has become a key technological path to improve the control quality and operational economy of the units.

[0003] In recent years, the field of time series forecasting has made significant progress, with the emergence of various deep learning models, including: Transformer-based models, such as Informer, Autoformer, and FEDformer, have improved long-term prediction accuracy by refining attention mechanisms and introducing seasonal-trend decomposition. However, these models are essentially still single-modal predictions, primarily utilizing the autoregressive characteristics of historical sequences and lacking the ability to integrate external factors (such as fuel characteristics, ambient temperature, and grid dispatch instructions).

[0004] Time series prediction methods based on Large Language Models (LLM), such as Time-LLM and LLaMA-Adapter, map time series data to the embedding space of LLM and leverage the powerful semantic understanding capabilities of LLM for prediction. While these methods perform well on some datasets, they still suffer from the following significant drawbacks: Limited modality: Most methods only use the raw time series as input, ignoring the structured information such as periodicity and trends inherent in the frequency domain, resulting in prediction accuracy limited to a single feature perspective; Poor interpretability: The LLM prediction process is typically a "black box" operation, making it difficult to explain which information in the input drives the prediction result, lacking explicit integration of physical laws and domain knowledge; Insufficient long-term prediction stability: As the prediction step size increases, error accumulation becomes severe, and the prediction results gradually deviate from the true trend, making them difficult to apply in real-world industrial scenarios.

[0005] Existing pseudo-graph or multimodal methods: While some recent work has attempted to convert time-series data into images or graph structures for processing, their image construction methods (such as directly plotting time-series curves) or graph structure definitions are too simplistic, failing to fully extract fine-grained features in the frequency domain. Furthermore, they lack semantic alignment during modality fusion, easily leading to modality conflicts or information loss. The information disclosed in the background section is merely for enhancing understanding of the background of this invention and may therefore contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, embodiments of this application provide a method for predicting time series data of thermal power generation based on a multimodal pseudograph large language model, comprising the following steps: S1: Collect time series data of multi-channel operating parameters of thermal power generator set, preprocess the time series data, divide it into multiple time segments according to the set window length and sliding step size, and obtain time domain block embedding vector through linear projection; S2: Perform a fast Fourier transform on each time segment to obtain an amplitude spectrum, divide the amplitude spectrum evenly into K frequency bands, perform an inverse fast Fourier transform on each frequency band to obtain a subsequence, arrange the K subsequences in rows to form a pseudo-image, and input the pseudo-image into a visual transformer to segment it into image blocks and extract frequency domain feature vectors. S3: The time segment is divided into time domain patches and mapped to time domain embedding vectors, and the frequency domain feature vector is mapped to frequency domain embedding vectors. The global semantics of the time domain embedding vectors and the frequency domain embedding vectors are aligned by a cross-modal predictor and a contrastive loss. S4: Extract statistics from the time segments, select the dominant frequency component in the frequency domain, and concatenate the generated text information into a prompt prefix; S5: The time domain embedding vector, the frequency domain embedding vector, and the embedding vector of the prompt prefix are concatenated and input into the pre-trained large language model. The model output is then processed by the prediction head to obtain the predicted value and output.

[0007] In conjunction with the first aspect, S1 specifically includes: S11: Perform instance normalization on the sequence of each channel to remove the temporal drift of the data distribution and retain the local statistical characteristics of the sequence; S12: Using a sliding window of length L, the long sequence is divided into multiple continuous segments of equal length with a step size S.

[0008] In conjunction with the first aspect, S2 specifically includes: S21: Perform a fast Fourier transform on the time segments to obtain a complex spectrum, and calculate the amplitude-frequency spectrum; S22: Divide the amplitude-frequency spectrum evenly into K frequency bands. For each frequency band, extract its corresponding amplitude value and retain the corresponding phase information to obtain the complex spectrum segment of that frequency band. S23: Perform inverse fast Fourier transform on the complex spectrum segment of each frequency band to obtain the time-domain subsequence; S24: Arrange the K time-domain subsequences in rows to form the pseudo-image; S25: The pseudo-image is segmented into multiple fixed-size image blocks, each image block is flattened and mapped to the embedding space through linear projection, and after adding position encoding, it is input into the encoder layer of the visual transformer for self-attention calculation to extract the frequency domain feature vector.

[0009] In conjunction with the first aspect, S3 specifically includes: S31: Divide the time segment into multiple time domain patches, perform linear mapping or MLP mapping on each patch to obtain the time domain embedding vector, whose dimension is consistent with the token embedding dimension of the large language model. S32: Convert the frequency domain feature vector into a frequency domain embedding vector consistent with the large language model through a linear mapping layer or adapter; S33: A cross-modal predictor is introduced, employing a joint embedding prediction architecture. The semantic distributions of temporal and frequency modalities are globally aligned in the batch-level top-level representation space. Overall semantic consistency is maximized through contrastive learning or mutual information. The contrastive loss is the cross-modal semantic consistency loss, calculated using the following formula: , Where N is the batch size. and Let be the time domain and frequency domain semantic representation vectors of the i-th sample, respectively, sim(·) be the cosine similarity function, and τ be the temperature hyperparameter.

[0010] In conjunction with the first aspect, S4 specifically includes: S41: Extract basic metadata from the current dataset and generate a text description; S42: For thermal power generation scenarios, domain knowledge is manually constructed or automatically extracted, and the domain knowledge is stored in the form of natural language text. S43: Construct explicit instruction text for the prediction task; S44: Perform trend-cycle decomposition on the time segment, extract trend components and cycle components, and calculate statistics, including mean, variance, first-order autocorrelation coefficient, skewness, kurtosis, trend slope and cycle length, and format them as numerical text. S45: Select the top M main frequency components with the largest amplitude from the amplitude-frequency spectrum, record their frequency values ​​and corresponding amplitude values, and generate a text description; S46: The text information generated in S41 to S45 above is concatenated in a fixed order to form the prompt prefix.

[0011] In conjunction with the first aspect, S5 specifically includes: S51: Concatenate the time-domain embedding vector, the frequency-domain embedding vector, and the embedding vector of the prompt prefix to obtain the complete input sequence of the large language model; S52: Input the concatenated input sequence into the pre-trained large language model. The parameters of the large language model are frozen or efficiently fine-tuned through the LoRA adapter, and the feature sequence is output. S53: Map the feature sequence to the target prediction dimension through a lightweight prediction head to obtain the prediction result; S54: The total loss function is a weighted average of the prediction loss and the cross-modal semantic consistency loss, expressed as: ,in, To predict losses, These are the weighting coefficients.

[0012] In conjunction with the first aspect, the prediction loss employs mean squared error or Huber loss, wherein the mean squared error is defined as: , Where N is the number of samples and P is the prediction step size. For the true value, These are predicted values.

[0013] In conjunction with the first aspect, the length L of the sliding window is less than or equal to 512, and the step size S is 1 or 2.

[0014] A second aspect of this disclosure provides an electronic device, comprising: One or more processors; A storage unit is used to store one or more programs that, when executed by one or more processors, enable the one or more processors to implement the method.

[0015] A third aspect of this disclosure provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method.

[0016] Beneficial Effects: This disclosure provides a method for predicting time series data of thermal power generation based on a multimodal pseudograph large language model. It preprocesses and embeds multi-channel time series data into segments with channel independence, decomposes the time-domain signal into multi-frequency sub-sequences using Fast Fourier Transform, and reconstructs them into a two-dimensional pseudograph. Frequency domain features are extracted using a visual transformer, and global semantic alignment of the time-domain and frequency-domain embedding vectors is achieved through a cross-modal predictor and contrastive loss. Simultaneously, dataset metadata, domain knowledge, statistics, and dominant frequency components are concatenated into a cue prefix, which, along with the time-frequency multimodal embedding, is input into a pre-trained large language model for prediction. This effectively integrates multimodal operating parameters and domain knowledge, significantly improving the prediction accuracy and long-term stability of key operating parameters of thermal power generating units. Its superiority over existing methods has been verified on multiple public datasets and actual thermal power unit data, while also enhancing the model's interpretability and industrial deployment adaptability. Attached Figure Description

[0017] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0018] Figure 1 This is a flowchart illustrating a method for predicting time series data of thermal power generation based on a multimodal pseudograph large language model, according to an embodiment of this application. Figure 2 The electronic device described in this application is an embodiment of the present application. Detailed Implementation

[0019] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0020] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0021] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0022] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0023] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0024] like Figure 1 The image shows a method for predicting the time series of thermal power generation based on a multimodal pseudograph large language model, as illustrated in the application embodiment. The method includes the following steps: S1: Collect time series data of multi-channel operating parameters of thermal power generator set, preprocess the time series data, divide it into multiple time segments according to the set window length and sliding step size, and obtain time domain block embedding vector through linear projection; S2: Perform a fast Fourier transform on each time segment to obtain an amplitude spectrum, divide the amplitude spectrum evenly into K frequency bands, perform an inverse fast Fourier transform on each frequency band to obtain a subsequence, arrange the K subsequences in rows to form a pseudo-image, and input the pseudo-image into a visual transformer to segment it into image blocks and extract frequency domain feature vectors. S3: The time segment is divided into time domain patches and mapped to time domain embedding vectors, and the frequency domain feature vector is mapped to frequency domain embedding vectors. The global semantics of the time domain embedding vectors and the frequency domain embedding vectors are aligned by a cross-modal predictor and a contrastive loss. S4: Extract statistics from the time segments, select the dominant frequency component in the frequency domain, and concatenate the generated text information into a prompt prefix; S5: The time domain embedding vector, the frequency domain embedding vector, and the embedding vector of the prompt prefix are concatenated and input into the pre-trained large language model. The model output is then processed by the prediction head to obtain the predicted value and output.

[0025] Specifically, S1 includes: S11: Perform instance normalization on the sequence of each channel to remove the temporal drift of the data distribution and retain the local statistical characteristics of the sequence; the normalized sequence is denoted as set. Where C is the number of channels, each The length is T0. S12: Using a sliding window of length L (preferably 512), the long sequence is divided into segments with a step size S (preferably S=1 or S=2). It is divided into multiple continuous segments of equal length, each segment being denoted as . ; S2 includes: S21: For the sequence Performing a Fast Fourier Transform (FFT) yields the complex spectrum F = FFT( ), where the dimension of F is L×1; calculate the amplitude-frequency spectrum A = |F|, which is the amplitude value vector, with a frequency range from 0 to half of the sampling frequency; S22: Divide the amplitude-frequency spectrum A uniformly into K frequency bands (the value of K is related to L, preferably K=4 or K=8); for each frequency band k (k=1,2,...,K), extract the amplitude value A corresponding to that frequency band. k And by preserving the corresponding phase information, the complex spectral segment F of that frequency band is obtained. k ; S23: Complex spectrum F for each frequency band k Perform an inverse fast Fourier transform (IFFT) to obtain the time-domain subsequence s. k =IFFT(F k ), each s k The length is L / K. It should be noted that the length of the subsequence after IFFT is not strictly equal to L / K. In actual processing, it is aligned to the standard length L / K by zero padding or pruning. S24: Divide the K subsequences Arranged row by row, they form a two-dimensional matrix of K rows × (L / K) columns, which is the pseudo-image. Each row in this pseudo-image represents a time-domain reconstructed signal for a frequency band. The row dimension of the image reflects different frequency bands, and the column dimension reflects the temporal evolution. S25: Remove pseudo-images The input is processed by a Vision Transformer (ViT) for feature extraction. Specifically: fake images The image is divided into multiple fixed-size patches, preferably 4×4 or 8×8 pixels. Each image patch is flattened and mapped to the embedding space via linear projection to obtain a patch embedding sequence. Position encoding is added, and self-attention calculation is performed on the encoder layer of ViT to extract frequency domain feature vectors. , where d is the embedding dimension (d is typically 768 or 1024). S3 includes: S31: Transfer the original sequence fragment The time domain is divided into multiple time-domain patches, each containing p consecutive time steps (p is preferably 16 or 32); a linear mapping or MLP mapping is performed on each patch to obtain the time-domain patch embedding vector. The dimension of this vector is consistent with the token embedding dimension of the large language model; S32: The frequency domain feature vector obtained in step S25 Transformed into a frequency domain embedding vector consistent with the large language model through a linear mapping layer or adapter. ; S33: Introducing a cross-modal predictor for time-domain embedded vectors With frequency domain embedding vector Perform global semantic alignment. Specifically: The cross-modal predictor employs a joint embedding prediction architecture, which globally aligns the semantic distributions of the two modalities in the top-level representation space at the batch level. Instead of imposing strict local correspondence constraints on the two modalities (i.e., not requiring each time-domain patch to correspond exactly to a specific frequency-domain patch), the overall semantic consistency is constrained by objectives such as contrastive learning or maximizing mutual information. The specific loss function is the cross-modal semantic consistency loss, defined as:

[0026] Where N is the batch size. and Let be the time domain and frequency domain semantic representation vectors of the i-th sample, respectively, sim(·) be the cosine similarity function, and τ be the temperature hyperparameter (preferably 0.1~0.5).

[0027] S4 includes: S41: Extract basic metadata of the current dataset, including: dataset name (e.g., ETTh1, ETTh2, ETTm1, Electricity, Traffic, or RGY), total number of sensors C, sampling interval, target prediction variable name, etc., and generate a text description; S42: For thermal power generation scenarios, manually construct or automatically extract domain knowledge. This domain knowledge includes, but is not limited to: energy conservation descriptions of boiler combustion processes, phase change laws of steam circulation, the influence of thermal inertia on parameter changes, and typical response relationships between load changes and main steam temperature. The knowledge is stored in natural language text format, preferably 50-200 characters in length. S43: Construct a clear prediction task instruction text, in the format "Please predict the next [P]time steps of the target sequence [TARGET_NAME] based on the input sequencelength [L].", where [P] is the prediction step size, [TARGET_NAME] is the target variable name, and [L] is the input sequence length; S44: For time-domain sequences Perform trend-cycle decomposition (using methods such as STL decomposition, moving average, or FFT filtering), extract the trend and cycle components, and calculate the following statistics: mean, variance, first-order autocorrelation coefficient, skewness, kurtosis, trend slope, and cycle length. Format these statistics as numerical text. S45: From the amplitude-frequency spectrum A, select the top M dominant frequency components with the largest amplitudes (M is preferably 3 to 5), and record their frequency values ​​and corresponding amplitude values. Generate a text description such as "Dominant frequencies: f1=0.01 Hz (amp=1.23), f2=0.05 Hz (amp=0.87),..."; S46: Concatenate the above text information in a fixed order to form the final prompt prefix text, Prompt_Prefix. The total length of this prompt prefix is ​​recommended not to exceed 200 tokens (approximately 150-300 characters) to avoid excessive length causing LLM attention distraction.

[0028] S5 includes: S51: Embed the time-domain patch obtained in step C into the vector. Frequency domain patch embedding vector The embedding vector of the prompt prefix text obtained in step D is concatenated to obtain the complete input sequence of the large language model; S52: Input the concatenated input sequence into the pre-trained large language model. The large language model is selected from GPT-2, LLaMA, Mistral, Gemma, or Qwen series models. During inference, the parameters of the large language model can be frozen (training only the embedding layer and reprogramming modules), or the parameters can be efficiently fine-tuned (e.g., using the LoRA adapter). The last layer of the large language model outputs a feature sequence; S53: The output features of the large language model are mapped to the target prediction dimension through a lightweight prediction head (usually a two-layer MLP with a hidden layer size of 256~1024) to obtain the prediction result; S54: The total loss function is a weighted average of the prediction loss and the cross-modal semantic consistency loss, expressed as:

[0029] in To predict loss, mean squared error (MSE) or Huber loss is typically used. MSE is defined as:

[0030] Where N is the number of samples and P is the prediction step size. For the true value, These are predicted values.

[0031] Electronic device 200 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 200 may include, but is not limited to, processor 201 and memory 202. Those skilled in the art will understand that... Figure 2 This is merely an example of electronic device 200 and does not constitute a limitation on electronic device 200. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0032] The processor 201 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0033] The memory 202 can be an internal storage unit of the electronic device 200, such as a hard disk or RAM of the electronic device 200. The memory 202 can also be an external storage device of the electronic device 200, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 200. Furthermore, the memory 202 can include both internal and external storage units of the electronic device 200. The memory 202 is used to store the computer program 203 and other programs and data required by the electronic device. The memory 202 can also be used to temporarily store data that has been output or will be output.

[0034] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for predicting time series data of thermal power generation based on a multimodal pseudograph large language model, characterized in that, Includes the following steps: S1: Collect time series data of multi-channel operating parameters of thermal power generator set, preprocess the time series data, divide it into multiple time segments according to the set window length and sliding step size, and obtain time domain block embedding vector through linear projection; S2: Perform a fast Fourier transform on each time segment to obtain an amplitude spectrum, divide the amplitude spectrum evenly into K frequency bands, perform an inverse fast Fourier transform on each frequency band to obtain a subsequence, arrange the K subsequences in rows to form a pseudo-image, and input the pseudo-image into a visual transformer to segment it into image blocks and extract frequency domain feature vectors. S3: The time segment is divided into time domain patches and mapped to time domain embedding vectors, and the frequency domain feature vector is mapped to frequency domain embedding vectors. The global semantics of the time domain embedding vectors and the frequency domain embedding vectors are aligned by a cross-modal predictor and a contrastive loss. S4: Extract statistics from the time segments, select the dominant frequency component in the frequency domain, and concatenate the generated text information into a prompt prefix; S5: The time domain embedding vector, the frequency domain embedding vector, and the embedding vector of the prompt prefix are concatenated and input into the pre-trained large language model. The model output is then processed by the prediction head to obtain the predicted value and output.

2. The method according to claim 1, characterized in that, S1 specifically includes: S11: Perform instance normalization on the sequence of each channel to remove the temporal drift of the data distribution and retain the local statistical characteristics of the sequence; S12: Using a sliding window of length L, the long sequence is cut into multiple continuous segments of equal length with a step size S.

3. The method according to claim 1, characterized in that, S2 specifically includes: S21: Perform a fast Fourier transform on the time segments to obtain a complex spectrum, and calculate the amplitude-frequency spectrum; S22: Divide the amplitude-frequency spectrum evenly into K frequency bands. For each frequency band, extract its corresponding amplitude value and retain the corresponding phase information to obtain the complex spectrum segment of that frequency band. S23: Perform inverse fast Fourier transform on the complex spectrum segment of each frequency band to obtain the time-domain subsequence; S24: Arrange the K time-domain subsequences in rows to form the pseudo-image; S25: The pseudo-image is segmented into multiple fixed-size image blocks, each image block is flattened and mapped to the embedding space through linear projection, and after adding position encoding, it is input into the encoder layer of the visual transformer for self-attention calculation to extract the frequency domain feature vector.

4. The method according to claim 1, characterized in that, S3 specifically includes: S31: Divide the time segment into multiple time domain patches, perform linear mapping or MLP mapping on each patch to obtain the time domain embedding vector, whose dimension is consistent with the token embedding dimension of the large language model. S32: Convert the frequency domain feature vector into a frequency domain embedding vector consistent with the large language model through a linear mapping layer or adapter; S33: A cross-modal predictor is introduced, employing a joint embedding prediction architecture. The semantic distributions of temporal and frequency modalities are globally aligned in the batch-level top-level representation space. Overall semantic consistency is maximized through contrastive learning or mutual information. The contrastive loss is the cross-modal semantic consistency loss, calculated using the following formula: ， Where N is the batch size. and Let be the time domain and frequency domain semantic representation vectors of the i-th sample, respectively, sim(·) be the cosine similarity function, and τ be the temperature hyperparameter.

5. The method according to claim 4, characterized in that, S4 specifically includes: S41: Extract basic metadata from the current dataset and generate a text description; S42: For thermal power generation scenarios, domain knowledge is manually constructed or automatically extracted, and the domain knowledge is stored in the form of natural language text. S43: Construct explicit instruction text for the prediction task; S44: Perform trend-cycle decomposition on the time segment, extract trend components and cycle components, and calculate statistics, including mean, variance, first-order autocorrelation coefficient, skewness, kurtosis, trend slope and cycle length, and format them as numerical text. S45: Select the top M main frequency components with the largest amplitude from the amplitude-frequency spectrum, record their frequency values ​​and corresponding amplitude values, and generate a text description; S46: The text information generated in S41 to S45 above is concatenated in a fixed order to form the prompt prefix.

6. The method according to claim 4, characterized in that, S5 specifically includes: S51: Concatenate the time-domain embedding vector, the frequency-domain embedding vector, and the embedding vector of the prompt prefix to obtain the complete input sequence of the large language model; S52: Input the concatenated input sequence into the pre-trained large language model. The parameters of the large language model are frozen or efficiently fine-tuned through the LoRA adapter, and the feature sequence is output. S53: Map the feature sequence to the target prediction dimension through a lightweight prediction head to obtain the prediction result; S54: The total loss function is a weighted average of the prediction loss and the cross-modal semantic consistency loss, expressed as: ,in, To predict losses, These are the weighting coefficients.

7. The method according to claim 6, characterized in that, The prediction loss uses mean squared error or Huber loss, where the mean squared error is defined as: ， Where N is the number of samples and P is the prediction step size. For the true value, These are predicted values.

8. The method according to claim 2, characterized in that, The length L of the sliding window is less than or equal to 512, and the step size S is 1 or 2.

9. An electronic device, characterized in that, include: One or more processors; A storage unit for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it can implement the method according to any one of claims 1 to 8.

Images