Large model construction method for intelligent collaboration and decision-making of power time series

Abstract

The invention discloses a large model construction method for intelligent collaboration and decision-making of a power time sequence, and aims to solve the problems of complex time sequence analysis task, high technical threshold and lack of transparency of a prediction result in a power scene. Comprising the following steps: S1, constructing a diversified power time sequence data set and enhancing characteristics thereof; s2, generating a data set with context description; s3, designing a time sequence thinking chain, and finishing alignment of the power time sequence data and the natural language mode; s4, constructing a flexible cue word template, and generating a diversified question and answer training set; and S5, carrying out model adaptation and supervision fine tuning by adopting a low-rank adaptation technology, and constructing a power time sequence large model with intelligent collaboration and decision-making capabilities. According to the method, the technical threshold of power time sequence analysis is remarkably reduced, the accuracy and transparency of a prediction result are improved, the method can be widely applied to the fields of power load prediction, anomaly detection, energy optimization scheduling and the like, and technical support is provided for intelligent transformation of the power industry.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent management of power systems, and specifically relates to a large model construction method for intelligent coordination and decision-making of power time series, which belongs to the intersection of big data analysis, artificial intelligence and power engineering technology. Background Art

[0002] With the rapid development of artificial intelligence, large language models (LLMs) have demonstrated excellent multimodal understanding capabilities and can interact with users through natural language. However, in the scenario of intelligent coordination and decision-making of power time series, existing models still have obvious deficiencies 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 sector.

[0003] Power time series data is a core data type in power system operation and management, covering household electricity consumption trends, regional power grid load changes, power generation equipment status monitoring, etc. 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 on users' professional knowledge and programming skills, making it difficult to achieve efficient and intuitive use.

[0004] In recent years, researchers have tried to apply large language models to the field of power time series analysis. However, existing research has mostly focused on using large language models to extract time series features to enhance prediction capabilities, while ignoring its natural language interaction potential. At present, the following key challenges are still faced in power scenarios. Diversity of task scenarios: The requirements of power time series tasks are diverse, including load forecasting, equipment monitoring, energy optimization, etc. The existing methods are insufficient in scene coverage and adaptability, and need to be customized in a targeted manner, which increases the difficulty of deployment; Diversity of inquiry methods: When using the system, power operation and maintenance personnel, dispatchers, and decision makers may express similar requirements in a variety of ways through natural language. The fixed prompt templates of existing methods are difficult to adapt, affecting the accuracy and consistency of the prediction results; Insufficient reasoning ability: In power scenarios, users not only need the model to provide prediction results, but also hope that the model explains the basis for the prediction, such as the cause of load changes or equipment abnormalities. It is difficult for existing methods to clearly express reasoning logic through natural language, which reduces its value in actual decision-making. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the existing technologies in power scenarios, the present invention proposes an intelligent collaboration and decision-making method that combines 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 integrating the feature extraction capability of time series data with the natural language interaction capability. It not only lowers the threshold for use, but also provides transparent and explanatory prediction results, significantly improving the intelligence level of the power system in scenarios such as load forecasting, anomaly detection, equipment maintenance and optimal scheduling, and promoting the development of the power industry towards automation and intelligence.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a large model construction method for intelligent coordination and decision-making of power time series, comprising the following steps:

[0007] S1. Construct a diversified power time series data set, collect real power time series data sets from multiple fields, and construct synthetic power time series data with multiple characteristics such as trend type, stable type, periodic type and composite type;

[0008] S2, generate a data set with context description, construct multi-type time series tasks according to power demand, classify and process the power time series data in S1 by time period, extract key features by feature decomposition, and add context information;

[0009] S3. Design a time series thinking chain to complete the alignment of power time series data and natural language modality, match the natural language analysis description with the time series data one by one, and obtain a modally aligned power time series data pair;

[0010] S4. Construct a flexible prompt word template to generate a diverse question-answering training set. The template includes three parts: data placeholder, data enhancement, and data embedding.

[0011] S5. Use low-rank adaptation technology to perform model adaptation and supervised fine-tuning to build a large power time series model with intelligent collaboration and decision-making capabilities.

[0012] Furthermore, in step S1, household electricity consumption data, regional power grid load data and power generation equipment status monitoring data are obtained from multiple public data sets 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 removed abnormal data are labeled to form an abnormal sample library for subsequent use in model training and improvement of anomaly detection capabilities; in order to make up for the lack of descriptiveness of real data, synthetic power time series data with typical characteristics are generated based on statistical modeling and simulation methods.

[0013] Furthermore, in step S2, in order to meet the diverse task requirements in power scenarios, the collected power time series data is processed and task designed in a targeted manner, including time series data classification and time period processing; time series feature decomposition and key feature extraction; and power time series data is classified according to different scenario requirements, mainly including load data, equipment monitoring data and anomaly detection data.

[0014] When processing data in different time periods, the time window technology is used to segment the data. Assume that the time window size is T, and the time series {x t} is divided into n subsequences {X i}, where T is variable: X i ={x iT ,x iT+1 ,…,x (i+1)T-1}, i = 0, 1, ..., n-1, combining the multimodal characteristics of time series, trend decomposition and cycle decomposition are used to extract the key features of time series data, and generate a context description in natural language for each piece of time series data.

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

[0016] S31. For different tasks in power scenarios, decomposition strategies are used to decompose complex tasks into more manageable sub-problems, including but not limited to extracting long-term trend of time series; identifying periodic patterns and quantifying repeated features; weighting and fusing multiple prediction results to generate final values, time series The trend T t , period S t and residual R t Composition x t =T t +S t +R t , where T t represents the trend component, S t represents the periodic component, R t represents random noise;

[0017] S32. Construct a temporal thinking chain based on natural language description, complete trend identification and cycle extraction through step-by-step reasoning, and generate natural language description;

[0018] S33. Analyze trends, cycles, and residual features, generate natural language descriptions of anomalies or causal relationships, align data with language modalities, provide a clear analysis process and recommended decision-making basis, and ultimately form an alignment structure: <original time series segment, corresponding thought chain description>.

[0019] Furthermore, in step S4, in order to improve the model's understanding and generalization ability of power sequence tasks, the design and application of flexible prompt word templates include the following specific steps:

[0020] S41. Define universal identifiers to abstract data content and support different power scenario requirements, such as:

[0021] Time-related identifiers: e.g. <date>Indicates the timestamp,<time_range> Indicates a time range; numerically related identifiers: such as<value_series> represents time series data, <forecast>Indicates the prediction result; inference related identifier: such as <reasoning>Represents the analysis process, <cause>Indicates causal relationships; special scenario identifiers: such as <anomaly>Indicates an abnormal event. <trend>Represents trend description. Through the unified definition of placeholders, templates can be flexibly reused in multiple scenarios;

[0022] S42. Manually design prompt word templates and combine them with large language models (LLMs) to generate question-answer pairs with various expression styles. Use large model verification technology to verify the accuracy of the templates and improve the template coverage capability.

[0023] S43, the time series features after mode alignment (such as trend term T t , periodic term S t ) embedded template, for example: the trend of the data is<T_trend> , the period is<S_cycle> ;

[0024] S44. Embed contextual information and thought chain descriptions in the template, such as time period<date_start> to<date_end> , data source <region>, predicting the future <period>The value of <forecast>, because <reasoning>;

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

[0026] Furthermore, in step S5, in order to improve the adaptability and efficiency of the large model in the power timing task, the specific process of model adaptation and supervision fine-tuning includes the following steps:

[0027] S51. Based on the large-scale autoregressive language model, a model with efficient mathematical reasoning and natural language understanding capabilities is selected as the base to handle the complex calculation and feature analysis of time series data;

[0028] S52. In the process of supervised fine-tuning, in order to reduce the computational overhead and avoid interference with the original knowledge of the model, the low-rank adapter (LoRA) technology is used to adjust the weight matrix W∈R of the pre-trained model. d×k Perform 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, A and B are the parameters learned during fine-tuning, and the pre-trained weight W is fixed during fine-tuning, and only the newly added low-rank matrices A and B are optimized;

[0029] S53. During the supervision and fine-tuning process, the task objectives are used as the driving force, and ultimately a large model for intelligent coordination and decision-making of power time series is obtained.

[0030] The beneficial effects of the present invention are:

[0031] (1) The present invention makes power time series analysis easier to use through data set construction and prompt word template design, supporting non-professional users to complete complex tasks through natural language.

[0032] (2) The temporal thinking chain reasoning and LoRA fine-tuning technology are used to significantly improve the accuracy and multi-task coordination capabilities of the model in tasks such as load forecasting and anomaly detection.

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

[0034] The method proposed in the present invention significantly lowers the technical threshold for power time series analysis, while improving the accuracy and transparency of the prediction results. It can be widely used in power load forecasting, energy optimization scheduling, equipment status monitoring and other fields, providing strong technical support for the intelligent transformation of the power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Attached Figure 1 Flowchart of the method for building a large model for intelligent collaboration and decision-making of power time series provided by the present invention.

[0036] Attached Figure 2 A schematic diagram of the building blocks of the modally aligned power time series data pair provided by the present invention.

[0037] Attached Figure 3 Schematic diagram of the construction module of the diversified question-answer pair training set provided by the present invention.

[0038] Attached Figure 4 Schematic diagram of the model adaptation and supervised fine-tuning module provided by the present invention. DETAILED DESCRIPTION

[0039] A typical embodiment of a large model construction method for intelligent coordination and decision-making of power time series of the present invention is described in detail below, and the method is further described in detail. It is necessary to point out here that the following embodiments are only used to further illustrate the method and cannot be understood as limiting the scope of protection of the method. Personnel skilled in the art in the field can make some non-essential improvements and adjustments to the method based on the above content of the method, which still fall within the scope of protection of the present invention.

[0040] The present invention proposes a large model construction method for intelligent coordination and decision-making of power time series. Figure 1 As shown, the basic steps are as follows:

[0041] S1. Construct a diversified power time series data set, collect real power time series data sets from multiple fields, covering real time series scenarios such as household electricity consumption, regional power grid load, and power generation equipment status monitoring, and construct synthetic power time series data with multiple characteristics such as trend type, stable type, periodic type, and composite type;

[0042] S2, generate a data set with context description, construct a variety of different types of time series tasks (such as power load forecasting, anomaly detection, causal analysis, etc.) according to the actual needs of power scenarios, and perform data preprocessing. The power time series data set collected in S1 is classified and divided into time periods, and key features are extracted by combining the time series feature decomposition method. On this basis, context information is added to time series data of different types and time periods to obtain a power time series data set with context description;

[0043] S3. Design a time series thinking chain to align power time series data with natural language modalities, decompose complex time series tasks in power scenarios into simpler sub-problems, such as trend item and cycle item decomposition and prediction result fusion, and accurately express the characteristic patterns and causal relationships of power time series data with context descriptions through natural language descriptions, realize modal alignment of time series and natural language, and obtain modal-aligned data pairs;

[0044] S4. Construct flexible prompt word templates and generate diversified question-answer training sets to achieve efficient modeling of power timing tasks. The template design includes the following elements:

[0045] 1) Data placeholder: Use a universal identifier (such as <date>、<value_series>、 <reasoning>etc.) to replace specific data and improve template reusability;

[0046] 2) Diversification enhancement: Combining manual design, LLMs simulation and verification technology, expanding the task instruction set, and expanding the coverage of semantic understanding of time series data through multiple rounds of prompt generation and simulation strategies;

[0047] 3) Data embedding: embed the time series data features, timestamps, and natural language descriptions after modality alignment into the template to generate a diverse question-answer pair training set;

[0048] This template design can adapt to the characteristics of power timing, support a wide range of scenarios, and respond to different questioning methods;

[0049] S5. Use low-rank adaptation technology to perform model adaptation and supervised fine-tuning to build a large model with intelligent collaboration and decision-making capabilities. Select a base model that is good at handling mathematical reasoning tasks as the initial model to provide a solid foundation for power time series tasks, fix most of the pre-training parameters, and only update the newly added low-rank matrix parameters; improve the model's ability to understand the context of power time series data, so that it adapts to the characteristic distribution of power time series data, and finally obtain a large dialogue model with intelligent collaboration and decision-making capabilities for power time series.

[0050] In steps S1 to S3 of the embodiment of the present invention, the construction process of the modally aligned power time series data pair is as follows: Figure 2 shown.

[0051] In step S1 of the embodiment of the present invention, household electricity consumption data, regional power grid load data and power generation equipment status monitoring data are obtained from multiple public data sets and actual power systems, and the following processing is performed: outliers and missing values ​​are removed by statistical methods. For missing values ​​in time series, different processing strategies are adopted according to different missing rates, and the removed abnormal data are marked to form an abnormal sample library for subsequent model training and improvement of abnormal detection capabilities. In order to make up for the lack of descriptiveness of real data, synthetic power time series data with typical characteristics are generated based on statistical modeling and simulation methods. The anomalies marked for time series x include: sudden increase anomalies (such as abnormal load increase in a short period of time); sudden drop anomalies (such as power failure or shutdown of equipment causing a sharp drop in load); periodic anomalies (such as atypical behaviors that occur repeatedly in a specific time period). Trend-type time series data are synthesized using a linear trend model; periodic time series data are synthesized based on a sine function; stationary time series data are synthesized by generating a stationary sequence through an autoregressive moving average model (ARMA); and composite time series data are synthesized by combining the above models.

[0052] In step S2 of the embodiment of the present invention, in order to meet the diverse task requirements in the power scenario, the collected power time series data is processed and task designed in a targeted manner, including time series data classification and time period processing, time series feature decomposition and key feature extraction.

[0053] Step S2 of this embodiment is specifically as follows:

[0054] S21. Classify the power time series data according to different scenario requirements, mainly including the following categories:

[0055] Load data: divided into daily load, weekly load and monthly load series according to time segments.

[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 the sampling frequency.

[0057] Anomaly detection data: Mark possible abnormal areas according to historical operation records and extract subsequences containing abnormal patterns.

[0058] S22. When processing data in different time periods, the time window technology is used to segment the data. Assume that the time window size is T, and the time series {x t } is divided into n subsequences {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. Combining the multimodal characteristics of time series, the following method is used to decompose time series data and extract key features:

[0060] Trend decomposition: Extract long-term trend items based on the moving average method: Among them, T t is the trend component at time t, and k is the window size.

[0061] Periodic decomposition: Extract the main periodic components using Fourier transform: Among them, S k represents the amplitude of frequency k.

[0062] S24. Generate a contextual description in natural language for each segment of time series data, including: data source information (such as household electricity consumption, regional load, etc.); time interval information (such as date range, seasonal characteristics); task-related descriptions (such as prediction targets, anomaly types, inferred causal relationships, etc.).

[0063] In step S3 of the embodiment of the present invention, in response to the complexity of the power time series data, a reasoning mechanism based on the time series thinking chain is designed, which includes the following specific steps:

[0064] 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:

[0065] Trend analysis: extract the long-term trend of time series and model its development direction; Periodicity analysis: identify the periodic patterns in time series and quantify their repetitive characteristics; Forecast result fusion: weighted fusion of multiple forecast results to generate the final forecast value, time series The trend T t , period S t and residual R t Composition: x t =T t +S t +R t , where T t represents the trend component, S t represents the periodic component, R t represents random noise.

[0066] S32. To improve the model's reasoning ability for complex time series tasks, a time series thinking chain based on natural language description is constructed. The model is guided to gradually complete the analysis through step-by-step reasoning, and the model is used to identify long-term change trends, identify major cycle patterns, and generate natural language descriptions.

[0067] S33. Analyze trend changes, cycles, and residual features to generate natural language descriptions of potential anomalies or causal relationships. Ensure seamless integration of data modalities and language modalities, and provide a clear analysis process and recommended decision-making basis, ultimately forming the following alignment structure:

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

[0069] In step S4 of this embodiment, Figure 3 To improve the model’s understanding and generalization capabilities of power timing tasks, a flexible prompt word template is designed and applied, which includes the following specific steps:

[0070] S41. The data placeholder abstracts the data content by defining a common identifier to support task requirements in different power scenarios. Common identifiers include: time-related identifiers: <date>Indicates the timestamp,<time_range> Indicates a time range; numerically related identifiers: such as<value_series> represents time series data, <forecast>Indicates the prediction result; inference related identifier: such as <reasoning>Represents the analysis process, <cause>Indicates causal relationships; special scenario identifiers: such as <anomaly>Indicates an abnormal event. <trend>Indicates 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 tasks such as power load forecasting, anomaly detection, and causal analysis, general prompt word templates are manually designed to ensure that they adapt to task requirements and data characteristics. Large language models (LLMs) are used to generate multiple language expressions, and question-answer pairs of different styles are generated through guided prompts. For templates and question-answer pairs generated through imitation, the accuracy of the generated content is verified using another language model or manual proofreading.

[0073] S43, in order to support the modeling of diversified power time series tasks, the time series data after modal alignment is embedded in the prompt word template. The key features extracted from the time series data (such as the trend term T t , periodic term S t ) are embedded in the templates respectively, for example: the trend of the following data is<T_trend> , the period is<S_cycle> .

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

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

[0076] Step S5 of the embodiment of the present invention is as follows: Figure 4 As shown, the adaptability and efficiency of the large model in the power timing task are improved. 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, able to handle complex numerical calculations and feature analysis involved in time series data; natural language processing ability, and efficient understanding of the diverse instructions input by users.

[0078] S52. In the process of supervised fine-tuning, in order to reduce the computational overhead and avoid interference with the original knowledge of the model, the low-rank adapter (LoRA) technology is used to adjust the weight matrix W∈R of the pre-trained model. d×k Perform 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 item updated during fine-tuning, r is the rank of the low-rank matrix, A and B are the parameters learned during fine-tuning, and the pre-trained weight W is fixed during fine-tuning, and only the newly added low-rank matrices A and B are optimized.

[0079] S53. During the supervision and fine-tuning process, the task objectives are used as the driving force, and ultimately a large model for intelligent coordination and decision-making of power time series is obtained.

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

[0081] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0082] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.< / 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 intelligent coordination and decision-making of power time series, characterized by: The following steps are involved: S1. Construct a diversified power time series data set, collect real power time series data sets from multiple fields, and construct synthetic power time series data with multiple characteristics such as trend type, stable type, periodic type and composite type; S2, generate a data set with context description, construct multi-type time series tasks according to power demand, classify and process the power time series data in S1 by time period, extract key features by feature decomposition, and add context information; S3. Design a time series thinking chain to complete the alignment of power time series data and natural language modality, match the natural language analysis description with the time series data one by one, and obtain a modally aligned power time series data pair; S4. Construct a flexible prompt word template to generate a diverse question-answering training set. The template includes three parts: data placeholder, data enhancement, and data embedding. S5. Use low-rank adaptation technology to perform model adaptation and supervised fine-tuning to build a large power time series model with intelligent collaboration and decision-making capabilities.

2. The method for constructing a large model based on intelligent coordination and decision-making for power time series according to claim 1 is characterized in that: In step S1, household electricity consumption data, regional power grid load data and power generation equipment status monitoring data are obtained from multiple public data sets and actual power systems, and outliers and missing values ​​are removed using statistical methods. For missing values ​​in time series, different processing strategies are adopted according to different missing rates; The eliminated abnormal data are labeled to form an abnormal sample library; synthetic power time series data with typical characteristics are generated based on statistical modeling and simulation methods.

3. The method for constructing a large model based on intelligent coordination and decision-making for power time series according to claim 2 is characterized in that: In the step S2, the collected power time series data is processed and task designed in a targeted manner, including time series data classification and time period processing; Time series feature decomposition and key feature extraction; classify power time series data according to different scenario requirements, mainly including load data, equipment monitoring data and anomaly detection data; When processing data in different time periods, the time window technology is used to segment the data. Assume that the time window size is T, and the time series {x t } is divided into n subsequences {X i }, where T is variable: X i ={x iT ,x iT+1 ,…,x (i+1)T-1 }, i = 0, 1, ..., n-1, combining the multimodal characteristics of time series, trend decomposition and cycle decomposition are used to extract the key features of time series data, and generate a context description in natural language for each piece of time series data.

4. The method for constructing a large model based on intelligent coordination and decision-making for power time series according to claim 3 is characterized in that: In step S3, a reasoning mechanism based on a temporal thinking chain is designed, specifically: S31. For different tasks in power scenarios, decomposition strategies are used to decompose complex tasks into more manageable sub-problems, including but not limited to extracting long-term trend of time series; identifying periodic patterns and quantifying repeated features; Weighted fusion of multiple prediction results to generate final value, time series The trend T t , period S t and residual R t Composition x t =T t +S t +R t , where T t represents the trend component, S t represents the periodic component, R t represents random noise; S32. Construct a temporal thinking chain based on natural language description, complete trend identification and cycle extraction through step-by-step reasoning, and generate natural language description; S33. Analyze trends, cycles, and residual features, generate natural language descriptions of anomalies or causal relationships, align data with language modalities, and provide a clear analysis process and recommended decision-making basis.

5. The method for constructing a large model based on intelligent coordination and decision-making for power time series according to claim 4 is characterized in that: In step S4, a flexible prompt word template is designed and applied, specifically: S41. Define universal identifiers to abstract data content and support different power scenario requirements; S42. Manually design prompt word templates and combine them with large language models (LLMs) to generate question-answer pairs with various expression styles. Use large model verification technology to verify the accuracy of the templates and improve the template coverage capability. S43, the time series features after mode alignment (such as trend term T t , periodic term S t )Embed template; S44, embed context information and thought chain description in the template; S45. Combine templates and data to generate a diverse question-answer pair training set.

6. The method for constructing a large model based on intelligent coordination and decision-making for power time series according to claim 5 is characterized in that: The step S5 is specifically as follows: S51. Select a pre-trained large-scale autoregressive language model as the base model, which is required to have efficient mathematical reasoning capabilities, be able to handle complex numerical calculations and feature analysis involved in time series data, and have efficient understanding capabilities for the diverse instructions input by users; S52. In the process of supervised fine-tuning, in order to reduce the computational overhead and avoid interference with the original knowledge of the model, the low-rank adapter (LoRA) technology is used to adjust the weight matrix W∈R of the pre-trained model. d×k Perform low-rank decomposition into two parts W+ΔW=W+A·B,A∈R d×r ,B∈Rr ×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 weight W is fixed, and only the newly added low-rank matrices A and B are optimized. S53. The supervision and fine-tuning process is driven by task objectives, and ultimately intelligent coordination and decision-making for power time series data are obtained.

Images