A large model construction method for power time series intelligent collaboration and decision

Abstract

The application discloses a kind of big model construction methods for power time sequence intelligent cooperation and decision, to solve the problems of complex time series analysis task, high technical threshold and lack of transparency of prediction result in power scene.It includes: S1, construct diversified power time series dataset and enhance its characteristics;S2, generate dataset with context description;S3, design time series thinking chain, complete power time series data and natural language modal alignment;S4, construct flexible prompt word template, generate diversified question and answer training set;S5, use low rank adaptation technology for model adaptation and supervised fine-tuning, construct power time sequence big model with intelligent cooperation and decision ability.The method of the application significantly reduces the technical threshold of power time series analysis, improves the accuracy and transparency of prediction results, and can be widely used in power load forecasting, anomaly detection and energy optimization scheduling, etc., providing technical support for intelligent transformation of power industry.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent management of power systems, and specifically relates to a large model construction method for power time series intelligent collaboration and decision-making, belonging to the cross-field of big data analysis, artificial intelligence and power engineering technology. BACKGROUND

[0002] With the rapid development of artificial intelligence, large language models (LLMs) have shown excellent multi-modal understanding capabilities and can interact with users through natural language. However, in the context of power time series intelligent collaboration and decision-making, existing models still have obvious shortcomings in processing time series data, including difficulty in generating accurate predictions, providing misleading results, and failing to meet diverse needs, which limits their practical application in the power field.

[0003] Power time series data is a core data type in power system operation and management, covering trends such as household electricity consumption, regional power grid load changes, and power generation equipment state monitoring. However, existing power time series analysis methods, including classical statistical models (such as ARIMA, GARCH, etc.) and deep learning models (such as LSTM, DLinear, Transformer, etc.), although they perform well in capturing complex patterns, they have high requirements for users' professional knowledge and programming skills, making it difficult to achieve efficient and intuitive use.

[0004] In recent years, researchers have attempted to apply large language models to the field of power time series analysis. However, existing research has focused on using large language models to extract time series features to enhance prediction capabilities, while ignoring their natural language interaction potential. Currently, there are still the following key challenges in the power context. Diversity of task scenarios: Power time series tasks have diverse needs, including load prediction, equipment monitoring, and energy optimization. Existing methods lack coverage and adaptability in scenarios, requiring targeted customization, which increases deployment difficulty; Diversity of inquiry methods: Power operation and maintenance personnel, dispatchers, and decision-makers may express similar needs in various ways through natural language when using the system. The fixed prompt templates of existing methods are difficult to adapt, affecting the accuracy and consistency of prediction results; Inadequate reasoning ability: In the power context, users not only need the model to provide prediction results, but also want the model to explain the basis for the prediction, such as load change reasons or equipment anomaly reasons. Existing methods are difficult to clearly express reasoning logic through natural language, reducing their value in actual decision-making. SUMMARY

[0005] In view of the above deficiencies in the prior art in the power scenario, the present application proposes an intelligent collaboration and decision-making method combining a large language model and time series analysis technology. This method can realize intelligent processing of multiple scenarios and multiple tasks in the power system by fusing the feature extraction capability of time series data and the natural language interaction capability. It not only reduces the use threshold, but also provides transparent and interpretable prediction results, significantly improves the intelligent level of the power system in load forecasting, anomaly detection, equipment maintenance and optimal scheduling scenarios, and promotes the development of the power industry towards automation and intelligence.

[0006] To achieve the above-mentioned application purposes, the technical solution adopted by the present application is as follows: a large model construction method for power time series intelligent collaboration and decision-making, comprising the following steps:

[0007] S1, constructing a diversified power time series dataset, collecting real power time series datasets from multiple fields, and constructing synthetic power time series data with multiple characteristics such as trend type, stationary type, periodic type and composite type;

[0008] S2, generating a dataset with context description, constructing multiple types of time series tasks according to power demand, classifying and processing power time series data in S1 by time period, extracting key features combined with feature decomposition, and adding context information;

[0009] S3, designing a time series thinking chain to complete the alignment of power time series data and natural language modal, one-to-one correspondence between natural language analysis description and time series data, and obtaining modal-aligned power time series data pairs;

[0010] S4, constructing a flexible prompt word template to generate a diversified question and answer training set, the template including data placeholders, data augmentation and data embedding;

[0011] S5, using low-rank adaptation technology for model adaptation and supervised fine-tuning to construct a power time series large model with intelligent collaboration and decision-making capability.

[0012] Further, in step S1, household electricity data, regional power grid load data and power generation equipment state monitoring data are obtained from multiple public datasets and actual power systems, statistical methods are used to remove outliers and missing values, different processing strategies are adopted for missing values in time series according to different missing rates, the removed abnormal data are labeled to form an abnormal sample library for subsequent model training and improvement of anomaly detection capability, and synthetic power time series data with typical characteristics are generated based on statistical modeling and simulation methods to make up for the lack of describability of real data.

[0013] Further, in step S2, in order to meet the diversified task requirements in the power scenario, the collected power time series data is processed and designed for specific tasks, including time series data classification and time period processing; time series feature decomposition and key feature extraction; the power time series data is classified according to different scene requirements, mainly including load data, equipment monitoring data and abnormal detection data.

[0014] When the data is processed by time period, the data is divided by using a time window technology, the size of the time window is T, the time series {x t} is divided into n subsequences {X i}, wherein T is variable: X i ={x iT ,x iT+1 ,…,x (i+1)T-1}, i=0, 1,…, n-1, in combination with the multi-modal characteristics of the time series, trend decomposition and periodic decomposition are used to extract key features of the time series data, and a context description in natural language form is generated for each time series data.

[0015] Further, in step S3, in order to cope with the complexity of the power time series data, a reasoning mechanism based on the time series thinking chain is designed, including the following specific steps:

[0016] S31, for different tasks in the power scenario, a decomposition strategy is used to decompose complex tasks into more manageable sub-problems, including but not limited to extracting long-term trends of time series; identifying periodic patterns, quantifying repetitive features; weighting and fusing multiple prediction results to generate the final value, the time series may be composed of trend T t , period S t and residual R t x t =T t +S t +R t , wherein T t represents the trend component, S t represents the periodic component, and R t represents random noise;

[0017] S32, construct a time series thinking chain based on natural language description, complete trend identification, period extraction and generate natural language description through step-by-step reasoning;

[0018] S33, analyze the trend, period and residual characteristics, generate natural language description of abnormality or causality, realize the alignment of data and language modal, provide clear analysis process and recommended decision basis, and finally form the alignment structure: <original time series segment, corresponding thinking chain description>.

[0019] Further, in the step S4, to improve the understanding and generalization ability of the model for the power time series task, the design and application of the flexible prompt template include the following specific steps:

[0020] S41, define a general identifier to abstract data content, support different power scene needs, for example:

[0021] Time-related identifiers: such as <date>represents a timestamp, <time_range> represents a time range; value-related identifiers: such as <value_series> represents time series data, <forecast>represents the predicted outcome; inference-related identifiers: such as <reasoning>representing the analysis process, <cause>Indicates a causal relationship; Special scenario identifier: e.g. <anomaly>an abnormal event is indicated, <trend>Indicates the trend description. Through the unified definition of placeholders, the template can be flexibly reused in multiple scenarios;

[0022] S42, manually design prompt word templates and generate question and answer pairs of various expression styles in combination with large language models (LLMs). The accuracy of the template is verified by using large model kernel calibration technology to improve the coverage ability of the template;

[0023] S43, embed the modality-aligned time sequence characteristics (such as trend item T t , cycle item S t ) into the template, for example: the trend of the data is <T_trend>, and the cycle is <S_cycle>;

[0024] S44, embed context information and thought chain description in the template, for example: time period <date_start> to <date_end>, data source <region>, predict future <period>the value of <forecast>, the cause is <reasoning>;

[0025] S45. Combine templates and data to generate a diverse question-answer pair training set:<C,X,Y> Where C represents the context, X represents the input time-series data, and Y represents the output prediction or analysis result. By adjusting the prompt words and generation style, it adapts to the different expression needs of users in various scenarios.

[0026] Furthermore, in step S5, to improve the adaptability and efficiency of the large model in power time-series tasks, the specific process of model adaptation and supervised fine-tuning includes the following steps:

[0027] S51. Based on a large-scale autoregressive language model, a model with efficient mathematical reasoning and natural language understanding capabilities is selected as the foundation for processing complex calculations and feature analysis of time series data.

[0028] S52. During supervised fine-tuning, to reduce computational overhead and avoid interference with the model's original knowledge, a low-rank adapter (LoRA) technique is used to adjust the weight matrix W∈R of the pre-trained model. d×k Perform a low-rank decomposition into two parts: W + ΔW = W + A·B, A∈R d×r ,B∈R r×k ,r<<min(d,k), where ΔW is the weight adjustment term updated during fine-tuning, r is the rank of the low-rank matrix, and A and B are the parameters learned during fine-tuning. During fine-tuning, the pre-trained weights W are fixed, and only the newly added low-rank matrices A and B are optimized.

[0029] S53. During the monitoring and fine-tuning process, the task objectives are the driving force, and a large model for intelligent collaboration and decision-making in power time series is finally obtained.

[0030] The beneficial effects of this invention are as follows:

[0031] (1) This invention makes power time series analysis easier to use by constructing datasets and designing prompt word templates, and supports non-professional users to complete complex tasks through natural language.

[0032] (2) By adopting temporal thinking chain reasoning and LoRA fine-tuning technology, the accuracy and multi-task collaboration capabilities of the model in tasks such as load forecasting and anomaly detection are significantly improved.

[0033] (3) Provide a transparent reasoning process and result analysis, generate decision-making basis including trends, cycles and causal relationships, and enhance users' trust in the model output.

[0034] The method provided by the application significantly reduces the technical threshold of power time series analysis, while improving the accuracy and transparency of the prediction results, and can be widely applied in the fields of power load prediction, energy optimization scheduling, equipment state monitoring and the like, thereby providing strong technical support for the intelligent transformation of the power industry. BRIEF DESCRIPTION OF DRAWINGS

[0035] FIG. 1 is a flowchart of a large model construction method for power time series intelligent collaboration and decision-making provided by the application. Figure 1 FIG. 1 is a flowchart of a large model construction method for power time series intelligent collaboration and decision-making provided by the application.

[0036] FIG. 1 is a flowchart of a large model construction method for power time series intelligent collaboration and decision-making provided by the application. Figure 2 FIG. 1 is a flowchart of a large model construction method for power time series intelligent collaboration and decision-making provided by the application.

[0037] FIG. 1 is a flowchart of a large model construction method for power time series intelligent collaboration and decision-making provided by the application. Figure 3 FIG. 1 is a flowchart of a large model construction method for power time series intelligent collaboration and decision-making provided by the application.

[0038] FIG. 1 is a flowchart of a large model construction method for power time series intelligent collaboration and decision-making provided by the application. Figure 4 FIG. 1 is a flowchart of a large model construction method for power time series intelligent collaboration and decision-making provided by the application. DETAILED DESCRIPTION

[0039] The following will describe a typical embodiment of the large model construction method for power time series intelligent collaboration and decision-making in detail, and further describe the method. It is necessary to point out that the following embodiment is only used to further illustrate the method, and cannot be understood as a limitation on the protection scope of the method. Those skilled in the art can make some non-essential improvements and adjustments to the method according to the above method content, which still belongs to the protection scope of the application.

[0040] The application provides a large model construction method for power time series intelligent collaboration and decision-making, as shown in FIG. 1, and the basic steps are as follows: Figure 1

[0041] S1, constructing a diversified power time series dataset, collecting real power time series datasets from multiple fields, covering real time series scenes such as household electricity, regional power grid load and power generation equipment state monitoring, and constructing synthetic power time series data with multiple characteristics such as trend type, stationary type, periodic type and composite type;

[0042] S2, generating a dataset with context description, constructing multiple different types of time series tasks (such as power load prediction, anomaly detection, causal analysis and the like) according to the actual needs of the power scene, and performing data preprocessing, classifying and time periodizing the power time series datasets collected in S1, and extracting key features in combination with a time series feature decomposition method, and on this basis, adding context information to the time series data of different types and different time periods to obtain a power time series dataset with context description;

[0043] ​S3, design a timing thought chain, complete the alignment of power timing data and natural language mode, decompose the complex timing task in the power scene into simpler subproblems, such as trend item, periodic item decomposition and prediction result fusion, accurately express the feature mode and causal relationship of the power timing data with context description through natural language description, realize the modal alignment of timing and natural language, and obtain the modal aligned data pair;

[0044] S4, construct a flexible prompt word template, generate a diversified question and answer training set, to realize efficient modeling of the power timing task. The template design includes the following elements:

[0045] 1) Data placeholder: use a general identifier (such as <date>, <value_series> <reasoning>Substitute specific data to improve template reusability

[0046] 2) Diversified enhancement: combined with manual design, LLM imitation and kernel correction technology, expand the task instruction set, expand the coverage of temporal data semantic understanding through multi-round prompt generation and imitation strategy;

[0047] 3) Data embedding: embedding the modality-aligned temporal data features, timestamps and natural language descriptions into the template to generate a diversified question and answer pair training set;

[0048] This template design can not only adapt to the characteristics of power time series, but also support a wide range of scenarios and different questioning methods.

[0049] S5, using low-rank adaptation technology for model adaptation and supervised fine-tuning, constructing a large model with intelligent collaboration and decision-making ability, selecting a base model good at handling mathematical reasoning tasks as the initial model, providing a solid foundation for power time series tasks, fixing most of the pre-training parameters, and only updating the newly added low-rank matrix parameters; improve the model's contextual understanding of power time series data, adapt to the feature distribution of power time series data, and ultimately obtain a dialogue large model with intelligent collaboration and decision-making ability for power time series.

[0050] In steps S1-S3 of the embodiment of the present application, the construction process of the modality-aligned power time series data pair is as shown in Figure 2

[0051] In step S1 of the embodiment of the present application, household electricity data, regional power grid load data and power generation equipment state monitoring data are obtained from multiple public data sets and actual power systems, and the following processing is performed: remove outliers and missing values by statistical methods, for missing values in time series, different processing strategies are taken according to different missing rates, label the removed abnormal data, and form an abnormal sample library for subsequent model training and improvement of abnormal detection capability. To make up for the lack of describability of real data, synthetic power time series data with typical characteristics are generated based on statistical modeling and simulation methods. The anomalies labeled for time series x include: sudden increase anomaly (such as abnormal load increase in a short period of time); sudden decrease anomaly (such as load sharp drop caused by equipment power failure or shutdown); periodic anomaly (such as atypical behavior repeatedly occurring in a specific period of time). Linear trend model is used to synthesize trend type time series data; based on sine function to synthesize periodic type time series data; through autoregressive moving average model (ARMA) to generate stationary sequence to synthesize stationary type time series data; combine the above models to synthesize composite type time series data.

[0052] ​In step S2 of the embodiment of the present application, in order to meet the diversified task requirements in the power scenario, the collected power time series data is processed and designed for specific tasks, including time series data classification and time period processing, time series feature decomposition and key feature extraction.

[0053] Step S2 in the embodiment is specifically:

[0054] S21, classify the power time series data according to different scene requirements, mainly including the following categories:

[0055] Load data: according to time segmentation, daily load, weekly load and monthly load sequence.

[0056] Equipment monitoring data: divided into high-frequency data (such as second-level monitoring) and low-frequency data (such as minute or hour-level monitoring) according to sampling frequency.

[0057] Abnormal detection data: according to historical operation records, mark possible abnormal areas, and extract sub-sequences containing abnormal patterns.

[0058] S22, when processing data by time period, use time window technology to cut the data, set the time window size as T, and the time series {x t} is divided into n sub-sequences {X i}, where T is variable, X i ={x iT ,x iT+1 ,…,x (i+1)T-1}, i=0,1,…,n-1.

[0059] S23, combined with the multi-modal characteristics of time series, the following methods are used to decompose time series data and extract key features:

[0060] Trend decomposition: based on the moving average method to extract the long-term trend item: Where T t is the trend component of time t, and k is the window size.

[0061] Periodic decomposition: use Fourier transform to extract the main periodic component: Where S k represents the amplitude of frequency k.

[0062] S24, generate context description in natural language form for each time series data, including: source information of the data (such as household electricity, regional load, etc.); time interval information (such as date range, seasonal characteristics); task-related description (such as prediction target, abnormal type, inference causal relationship, etc.).

[0063] In step S3 of the embodiment of the present application, in order to cope with the complexity of power time series data, a reasoning mechanism based on time series thinking chain is designed, including the following specific steps:

[0064] S31, for different tasks of power scenarios, a decomposition strategy is adopted to decompose complex tasks into more manageable sub-problems, including but not limited to:

[0065] Trend analysis: extract the long-term trend of time series, model its development direction; periodicity analysis: identify the periodic pattern in the time series, quantify its repetitive characteristics; prediction result fusion: weight fusion of multiple prediction results, generate the final prediction value, time series Trend T t , periodicity S t and residual R t can be composed of: x t =T t +S t +R t , wherein T t represents the trend component, S t represents the periodic component, and R t represents random noise.

[0066] S32, to improve the reasoning ability of the model for complex time series tasks, construct a time series thinking chain based on natural language description. Through step-by-step reasoning, the model gradually completes the analysis, identifies long-term trends and main periodic patterns, and generates natural language descriptions.

[0067] S33, analyze the trend change, periodicity and residual characteristics, and generate natural language descriptions of potential anomalies or causal relationships. Ensure seamless integration of data modalities and language modalities, and provide clear analysis process and recommended decision basis, finally form the following alignment structure:

[0068] <Original time series segment, corresponding thinking chain description>

[0069] Step S4 of the embodiment, as shown in Figure 3 To improve the understanding and generalization ability of the model for power time series tasks, flexible prompt word templates are designed and applied, including the following specific steps:

[0070] S41, the data placeholder abstracts the data content by defining a general identifier, supports the task requirements in different power scenarios, and the commonly used identifiers include: time-related identifiers such as <date>represents a timestamp, <time_range> represents a time range; value-related identifiers: such as <value_series> represents time series data, <forecast>represents the predicted outcome; inference-related identifiers: such as <reasoning>representing the analysis process, <cause>Indicates a causal relationship; Special scenario identifier: e.g. <anomaly>an abnormal event is indicated, <trend>Indicates the trend description. Through the unified definition of placeholders, templates can be flexibly reused in multiple scenarios.

[0071] S42, to improve the coverage and generation quality of the prompt word template, the following enhancement methods are combined:

[0072] For power load forecasting, anomaly detection and causal analysis tasks, manually design general prompt word templates to ensure they adapt to task requirements and data characteristics; use large language models (LLMs) to generate multiple language expressions, and generate different styles of question and answer pairs through guided prompts; use another language model or manual proofreading to verify the accuracy of the generated content for templates and question and answer pairs generated by imitation.

[0073] S43, to support the modeling of diversified power time series tasks, the modal aligned time series data is embedded in the prompt word template. The key features (such as trend items T t , cycle items S t ) extracted from the time series data are embedded in the template, for example: the trend of the following data is <T_trend> and the cycle is <S_cycle>.

[0074] S44, embed the context description and thought chain reasoning corresponding to each time series data in the template. For example, the context: the time period is <date_start> to <date_end>, the load data is from the regional power grid, the time series data is <time_series>, and the future <period>the timing data. The prediction result is that the time series is <forecast>, by inference, the reason is <reasoning>.

[0075] S45. Combine the embedded templates with the data to generate a diverse question-answer pair training set:<C,X,Y> Where C represents the context description, X represents the input time series data, and Y represents the output prediction or analysis result.

[0076] Step S5 of this embodiment of the invention, as follows Figure 4 As shown, to improve the adaptability and efficiency of large models in power time-series tasks, the specific process of model adaptation and supervised fine-tuning includes the following steps:

[0077] S51. Select a pre-trained large-scale autoregressive language model as the base model. The selected model must meet the following characteristics: efficient mathematical reasoning ability, capable of handling complex numerical calculations and feature analysis involved in time series data; natural language processing ability, capable of efficiently understanding diverse user input commands.

[0078] S52. During supervised fine-tuning, to reduce computational overhead and avoid interference with the model's original knowledge, a low-rank adapter (LoRA) technique is used to adjust the weight matrix W∈R of the pre-trained model. d×k Perform a low-rank decomposition into two parts: W + ΔW = W + A·B, A ∈ R d×r ,B∈R r×k ,r<<min(d,k), where ΔW is the weight adjustment term updated during fine-tuning, r is the rank of the low-rank matrix, and A and B are the parameters learned during fine-tuning. During fine-tuning, the pre-trained weights W are fixed, and only the newly added low-rank matrices A and B are optimized.

[0079] S53. During the monitoring and fine-tuning process, the task objectives are the driving force, and a large model for intelligent collaboration and decision-making in power time series is finally obtained.

[0080] The method proposed in this invention significantly lowers the technical threshold for power time series analysis while improving the accuracy and transparency of prediction results. It can be widely applied in fields such as power load forecasting, energy optimization scheduling, and equipment condition monitoring, providing strong technical support for the intelligent transformation of the power industry.

[0081] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0082] Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and that the inventive principles are not limited to these particular embodiments. Other variations and modifications can be made to the embodiments without departing from the spirit and scope of the inventive principles.< / reasoning> < / forecast> < / period> < / trend> < / anomaly> < / cause> < / reasoning> < / forecast> < / date> < / reasoning> < / date> < / reasoning> < / forecast> < / period> < / region> < / trend> < / anomaly> < / cause> < / reasoning> < / forecast> < / date>

Claims

1. A large model construction method for power time series intelligent collaboration and decision, characterized in that, Comprising the following steps: S1, constructing a diversified power time series dataset: collecting real power time series datasets from multiple fields, while constructing synthetic power time series data; S2, generating a dataset with context description, according to the power demand, constructing multi-type time series tasks, classifying, time period processing, extracting key features combined with feature decomposition, and adding context information; step S2 is specifically: S21, classify the power time series data according to different scene requirements, including the following categories: load data: according to time segmentation, daily load, weekly load and monthly load sequence; equipment monitoring data: divided into high frequency data and low frequency data according to sampling frequency; anomaly detection data: according to historical operation records, label possible abnormal area, extract subsequence containing abnormal mode; S22, when the data is processed by time period, the data is cut by using time window technology; S23, combined with the multi-modal characteristics of time series, the key features are extracted by using trend decomposition and period decomposition technology: based on moving average method to extract long-term trend item, using Fourier transform to extract main period component; S24, generate context description in natural language form for each time series data, including: data source information; time interval information; task related description; S3, design a time sequence thinking chain, complete the alignment of power time sequence data and natural language mode, one-to-one correspondence between natural language analysis description and time sequence data, get the modal aligned power time sequence data pair, in step S3, design a reasoning mechanism based on time sequence thinking chain, specifically: S31, trend analysis: extract the long-term trend of time series, model its development direction; periodicity analysis: identify the periodic pattern in the time series, quantify its repetitive characteristics; prediction result fusion: weight fusion of multiple prediction results, generate the final prediction value, time series Can be composed of trend T t , cycle S t And residual error R t : x t =T t +S t +R t , wherein T t Indicates the trend component, S t Indicates the periodic component, R t Indicates random noise; S32, construct a time sequence thinking chain based on natural language description, guide the model to complete the analysis step by step through step-by-step reasoning, identify the long-term trend and the main periodic pattern, and generate a natural language description; S33, analyze the trend change, cycle and residual error characteristics, and generate natural language description of potential anomalies or causal relationships; S4, construct flexible prompt word template, generate diversified question and answer training set, the template includes data placeholder, data enhancement and data embedding three parts, in step S4, design and apply flexible prompt word template, specifically: S41, the data placeholder abstracts data content by defining a general identifier, supporting task requirements in different power scenarios; S42, manually design general prompt word template, use large language model to generate multiple language expressions, construct general prompt word template and question and answer pairs with different language styles, and then use another language model or artificial correction to verify the accuracy of the generated content; S43, embed the modality aligned time series data features into the general prompt word template; S44, embed the context description and thinking chain reasoning corresponding to each time series data into the general prompt word template; S45, combine the embedded general prompt word template with the power time series data to generate a diversified question and answer pair training set; S5, using the diversified question and answer pair training set in step S4, using low rank adaptation technology to adapt and supervise the fine tuning of the model, and constructing a power time series large model with intelligent collaboration and decision-making ability.

2. The method of claim 1, wherein, In step S1, household electricity data, regional power grid load data and power generation equipment state monitoring data are obtained from multiple public datasets and actual power systems, and statistical methods are used to remove outliers and missing values. For missing values in time series, different processing strategies are adopted according to different missing rates; The excluded abnormal data is labeled to form an abnormal sample library; based on statistical modeling and simulation method, synthetic power time series data with typical characteristics are generated.

3. The method of claim 1, wherein the method is characterized by, The step S5 is specifically: S51, select a pre-trained large-scale autoregressive language model as a base model; S52、in the supervision fine-tuning process, using low rank adapter (LoRA) technology, the weight matrix W of the pre-training model is decomposed into two parts d ×k Low rank decomposition is divided into two parts W+ΔW=W+A·B,A∈R d×r ,B∈R r×k ,r<<min(d,k), wherein, ΔW is the weight adjustment item updated during fine-tuning, r is the rank of the low rank matrix, A and B are parameters learned during fine-tuning; in fine-tuning, the pre-training weight W is fixed, and only the newly added low rank matrix A and B are optimized; S53, in the supervision fine tuning process, driven by task target, finally obtain intelligent collaboration and decision-making for power time series data.

Images