Load prediction method fusing large language model and agent technology

Abstract

The application relates to a load prediction method fusing a large language model and an agent technology. The method comprises the following steps: a large language model generates to-be-predicted data based on historical structured data and historical unstructured data in a power system; the prediction task is divided into multiple subtasks based on a load prediction demand of the power system, and the multiple subtasks are distributed to multiple agents; a model selection agent determines the adaptation degrees of multiple preset basic prediction models based on the to-be-predicted data and the load prediction demand, and determines a candidate prediction model from the multiple preset basic prediction models based on the adaptation degrees; a hyperparameter optimization agent optimizes model parameters of the candidate prediction model to obtain a target prediction model; and a prediction execution agent performs power load prediction on the to-be-predicted data based on the target prediction model and the load prediction demand to obtain a power load prediction result. The method can improve the accuracy of the prediction result.

Description

Technical Field

[0001] This application relates to the field of power load forecasting technology, and in particular to a load forecasting method that integrates large language models and intelligent agent technology. Background Technology

[0002] Power load forecasting is a core component of power system dispatching and operation, directly impacting grid security and power supply reliability.

[0003] Currently, the types of data used in power load forecasting are limited, and the forecasting models are also limited, which affects the accuracy of power load forecasting results.

[0004] Therefore, how to predict power load and improve the accuracy of the prediction results is a problem that needs to be solved. Summary of the Invention

[0005] Therefore, it is necessary to provide a load forecasting method, apparatus, computer equipment, and medium that integrates large language model and intelligent agent technology to address the above-mentioned technical problems, so as to improve the accuracy of power load forecasting results.

[0006] In a first aspect, this application provides a load forecasting method that integrates large language models and intelligent agent technology. The method is applied to a multi-agent power load forecasting system, which includes a large language model and multiple intelligent agents. The method includes:

[0007] The large language model is invoked to generate data to be predicted based on historical structured data and historical unstructured data in the power system; wherein, the historical structured data is data formed by monitoring the operation process of the power system based on historical time series, and the historical unstructured data is data constrained by the historical operation process of the power system using natural language as a carrier;

[0008] The large language model is invoked, and the prediction task is divided into multiple sub-tasks based on the load prediction requirements of the power system. The multiple sub-tasks are then assigned to multiple intelligent agents. The multiple intelligent agents include at least a model selection agent, a hyperparameter optimization agent, and a prediction execution agent.

[0009] The model selection agent is invoked to determine the fit of multiple preset basic prediction models based on the data to be predicted and the load prediction requirements, and to determine candidate prediction models from the multiple preset basic prediction models based on the fit.

[0010] The hyperparameter optimization agent is invoked to optimize the model parameters of the candidate prediction model, thereby obtaining the target prediction model;

[0011] The prediction execution agent is invoked to perform power load prediction on the data to be predicted based on the target prediction model and the load prediction requirements, thereby obtaining the power load prediction result.

[0012] In one embodiment, determining the fit of multiple preset basic prediction models based on the data to be predicted and the load prediction demand, and determining candidate prediction models from the multiple preset basic prediction models based on the fit, includes:

[0013] Determine the historical prediction accuracy and prior fit of the target basic prediction model; the target basic prediction model is any one of the multiple preset basic prediction models; the historical prediction accuracy is the prediction accuracy of the data to be predicted based on the target basic prediction model and the load prediction demand in historical prediction data; the prior fit is the fit of the large language model determined based on domain prior knowledge.

[0014] The fitness of the target basic prediction model is determined based on the historical prediction accuracy, the prior fitness, and the computational complexity of the target basic prediction model.

[0015] Based on the fit of the multiple preset basic prediction models, at least two candidate prediction models are determined from the multiple preset basic prediction models.

[0016] In one embodiment, optimizing the model parameters of the candidate prediction model to obtain the target prediction model includes:

[0017] Based on domain prior knowledge, the range of model parameters for each candidate prediction model is determined;

[0018] Based on the range of model parameters, the corresponding candidate prediction models are iteratively trained until the corresponding candidate prediction models converge, thereby obtaining the converged prediction model corresponding to each candidate prediction model.

[0019] Among the converged prediction models, the converged prediction model with the smallest model error is determined as the target prediction model.

[0020] In one embodiment, the plurality of agents further includes an error tracing agent; the method further includes:

[0021] The prediction execution agent is invoked to perform power load prediction on the data to be predicted based on the target prediction model and the load prediction demand, and an initial prediction result is obtained.

[0022] The error tracing agent is invoked to determine the prediction error based on the initial prediction result and the measured results corresponding to the data to be predicted in historical data; the large language model is then invoked to analyze the source of the error based on the prediction error to obtain the error type; the initial prediction result is then corrected based on the error type to obtain the power load prediction result.

[0023] In one embodiment, the plurality of intelligent agents further includes a security verification intelligent agent; the method further includes:

[0024] The error tracing agent is invoked to correct the initial prediction result based on the error type, thereby obtaining a corrected prediction result;

[0025] The security verification agent is invoked to verify the corrected prediction result. If the verification result is that the verification fails, the steps of invoking the error tracing agent to correct the corrected prediction result based on the measured result and invoking the security verification agent to verify the corrected prediction result are executed repeatedly until the verification result is that the verification passes. The corrected prediction result that passes the verification is then determined as the power load prediction result.

[0026] In one embodiment, generating the data to be predicted based on historical structured data and historical unstructured data from the power system includes:

[0027] Semantic features are extracted from the historical unstructured data to obtain multiple semantic feature vectors;

[0028] The multiple semantic feature vectors are weighted and fused to obtain the fused semantic features;

[0029] The fused semantic features and the historical structured data are time-series aligned and normalized to obtain the data to be predicted.

[0030] In one embodiment, the method further includes invoking a large language model to perform the following steps:

[0031] The operating status of the multiple intelligent agents is monitored to obtain the monitoring results of each intelligent agent;

[0032] Optimize tasks in agents whose monitoring results are abnormal;

[0033] Based on preset update conditions, the operating parameters of each of the intelligent agents are updated;

[0034] If the prediction results meet the preset conditions in multiple prediction periods, the operating parameters of each intelligent agent are determined as standard operating parameters, and power load standardization prediction is performed in the power load prediction system based on the standard operating parameters.

[0035] Secondly, this application also provides a load forecasting device integrating large language model and intelligent agent technology. The device is applied to a multi-agent power load forecasting system, which includes a large language model and multiple intelligent agents. The device includes:

[0036] The data processing module is used to call the large language model and generate data to be predicted based on historical structured data and historical unstructured data in the power system; wherein, the historical structured data is data formed by monitoring the operation process of the power system based on historical time series, and the historical unstructured data is data constrained by the historical operation process of the power system using natural language as a carrier;

[0037] The task allocation module is used to call the large language model, divide the prediction task into multiple sub-tasks based on the load prediction requirements of the power system, and allocate the multiple sub-tasks to multiple intelligent agents; the multiple intelligent agents include at least a model selection intelligent agent, a hyperparameter optimization intelligent agent, and a prediction execution intelligent agent;

[0038] The model selection module is used to call the model selection agent to determine the fit of multiple preset basic prediction models based on the data to be predicted and the load prediction requirements, and to determine the candidate prediction model from the multiple preset basic prediction models based on the fit.

[0039] The parameter optimization module is used to call the hyperparameter optimization agent to optimize the model parameters of the candidate prediction model to obtain the target prediction model;

[0040] The prediction module is used to call the prediction execution agent to perform power load prediction on the data to be predicted based on the target prediction model and the prediction requirements, and obtain the power load prediction result.

[0041] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in any of the embodiments of the first aspect described above.

[0042] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in any of the embodiments of the first aspect described above.

[0043] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in any of the embodiments of the first aspect described above.

[0044] In the above implementation process, the large language model generates the data to be predicted based on historical structured and unstructured data. This ensures a rich variety of data types involved in power load forecasting, effectively reducing the problem of low prediction accuracy caused by a single data type and significantly improving the accuracy of power load forecasting results. Furthermore, the large language model divides the forecasting task into multiple sub-tasks based on load forecasting requirements and assigns these sub-tasks to multiple agents for execution. Specifically, the model selection agent obtains the fit of multiple preset basic forecasting models based on the data to be predicted and load forecasting requirements, and then determines candidate forecasting models from these models based on their fit, achieving adaptive model selection and reducing the problem of low prediction accuracy caused by a single forecasting model. The hyperparameter optimization agent optimizes the model parameters of the candidate forecasting models to obtain the target forecasting model, improving the forecasting performance and thus facilitating higher accuracy in load forecasting results. Finally, the prediction execution agent performs forecasting based on the target forecasting model and load forecasting requirements to obtain the prediction results. By splitting the prediction task into multiple agents, the coupling of the power load prediction system is reduced, allowing each agent to run in parallel or independently, thus improving the efficiency of power load prediction. Furthermore, by using multiple types of data to participate in power load prediction and by adaptively selecting prediction models to predict the power load of the power system, the accuracy of power load prediction results is effectively improved. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of the application environment for a load prediction method that integrates large language models and intelligent agent technology, as provided in an embodiment of this application.

[0047] Figure 2 This is a flowchart illustrating a load prediction method that integrates large language models and intelligent agent technology, as provided in an embodiment of this application.

[0048] Figure 3 This is a schematic flowchart of a method for determining a candidate prediction model provided in an embodiment of this application;

[0049] Figure 4This is a schematic diagram of a load prediction device that integrates large language model and intelligent agent technology, provided in an embodiment of this application.

[0050] Figure 5 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0053] The load prediction method integrating large language models and intelligent agent technology provided in this application can be applied to, for example... Figure 1 The application environment shown. Figure 1 This is a schematic diagram illustrating the application environment of a load prediction method integrating large language models and intelligent agent technology provided in an embodiment of this application. Figure 1 As shown, terminal 102 communicates with server 104 via a network. The data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, and IoT devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart vehicle devices, projection devices, etc. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0054] In one exemplary embodiment, Figure 2 This is a flowchart illustrating a load forecasting method integrating large language models and agent technology, provided in an embodiment of this application. This method is applied to a multi-agent power load forecasting system, which can be deployed in... Figure 1In server 104, the multi-agent power load forecasting system may include a large language model and multiple agents, such as Figure 2 As shown, the method includes the following steps:

[0055] Step 201: Call the large language model to generate the data to be predicted based on the historical structured data and historical unstructured data in the power system.

[0056] Among them, historical structured data is data formed by monitoring the operation process of the power system based on historical time series, while historical unstructured data is data that uses natural language as a carrier to constrain the historical operation process of the power system.

[0057] For example, the server can acquire historical structured data and historical unstructured data of the power system during its historical operation. Specifically, historical structured data can be data generated by monitoring the operation of the power system based on historical time series, such as historical load time series data, meteorological time series data, and power grid operation parameter data. Here, the historical time series can be a series of consecutive time points or time periods in the past. Historical unstructured data can be data that uses natural language as a carrier to constrain the historical operation of the power system, such as power dispatch policy text, user electricity consumption behavior text, extreme weather warning text, and power grid fault record text. Here, the historical operation process can be the operation process of the power system during a certain period in the past.

[0058] As an example, the server can collect historical load time-series data of the power system in real time using the OPC Unified Architecture (OPC UA) protocol, with a sampling frequency of 15 minutes per point; obtain meteorological time-series data such as temperature, humidity, and wind speed through the meteorological department's API interface; and obtain historical structured data such as unit output, line power flow, and transformer load rate through the dispatching OMS system. For historical unstructured data, the server can periodically crawl power dispatching policy texts issued by relevant departments, collect user electricity consumption behavior text logs uploaded by smart meter terminals, access extreme weather warning texts issued by the meteorological department such as orange high temperature warnings and blue cold wave warnings, and power grid fault record texts including fault time, fault type, and affected area, thereby obtaining historical unstructured data.

[0059] Furthermore, the large language model generates the data to be predicted based on the acquired historical structured and unstructured data. Specifically, the large language model can extract semantic features and perform structuring transformation on the historical unstructured data, thereby converting the historical unstructured data into structured data, and then use the transformed historical unstructured data and the historical structured data together to generate the data to be predicted.

[0060] Step 202: Invoke the large language model, divide the forecasting task into multiple sub-tasks based on the load forecasting demand of the power system, and assign the multiple sub-tasks to multiple agents.

[0061] Multiple agents include at least a model selection agent, a hyperparameter optimization agent, and a prediction execution agent.

[0062] For example, the process of forecasting power load may include multiple stages, and the tasks corresponding to each stage are different due to varying forecasting needs. Therefore, a large language model can break down the forecasting task into multiple sub-tasks based on the load forecasting needs, and assign these sub-tasks to multiple agents, thereby enabling the corresponding agents to execute the sub-tasks.

[0063] Specifically, load forecasting requirements can include forecasting period, forecasting accuracy, forecasting scenarios, and data boundaries. Large language models can parse load forecasting requirements and break down forecasting tasks based on the parsing results, with a breakdown depth of 3 to 5 levels.

[0064] As an example, a large language model can include a domain-fine-tuned large language model and a core scheduling large language model. Both the core scheduling large language model and the domain-fine-tuned large language model are obtained by fine-tuning the general large language model through the power load forecasting domain corpus. The power load forecasting domain corpus includes power grid dispatching procedures, load forecasting technical specifications, historical forecasting cases, power market policies, and knowledge texts related to meteorology and load. The accuracy of the fine-tuned model in knowledge question answering in the power field is ≥95%.

[0065] Domain-fine-tuned large language models can be used to extract semantic features and transform structured data from historical unstructured data. Core scheduling large language models can divide prediction tasks into multiple sub-tasks according to load prediction requirements and assign these sub-tasks to multiple agents.

[0066] Specifically, when the dispatcher inputs the forecasting task as: "Predict the residential electricity load of a certain prefecture-level city in the next 24 hours, with an accuracy requirement of MAPE≤3%, for day-ahead dispatching planning", the core dispatching big language model first analyzes the load forecasting requirement mechanism and obtains: the forecasting period is 24 hours (day-ahead forecasting), the forecasting accuracy requirement is MAPE≤3%, the forecasting scenario is residential load forecasting, and the data boundary is the aggregated data of smart meters across the prefecture-level city and the corresponding meteorological data.

[0067] Then, the prediction task can be broken down into data governance sub-tasks, feature engineering sub-tasks, model training sub-tasks, and prediction execution sub-tasks. Multiple agents include at least a model selection agent, a hyperparameter optimization agent, and a prediction execution agent. Furthermore, these agents may also include a data governance agent and a feature engineering agent. Thus, the data governance sub-task can be executed by the data governance agent, the feature engineering sub-task by the feature engineering agent, the model training sub-task by the model selection agent and hyperparameter optimization agent, and the prediction sub-task by the prediction execution agent.

[0068] Specifically, the prediction task is broken down into multiple sub-tasks at four levels of depth: Level 1 is divided into four stages: data governance, feature engineering, model training, and prediction execution; Level 2 further breaks down data governance into data cleaning, data alignment, and missing value imputation sub-tasks; Level 3 further breaks down data cleaning into outlier detection, duplicate value removal, and dimensional unification; and Level 4 breaks down outlier detection into numerical anomaly detection based on the 3σ criterion and textual anomaly detection based on semantic understanding. The final result is a task set containing 18 sub-tasks.

[0069] Furthermore, the core scheduling large language model can adopt a token-based hybrid scheduling mechanism: the data governance subtask and the feature engineering subtask have a strong dependency relationship, that is, feature engineering needs to wait for data governance to be completed, and serial scheduling is adopted, with tokens being issued sequentially by the core scheduling large language model; model selection and hyperparameter optimization have a strong dependency relationship, but data governance and model training have a weak dependency relationship (partially parallelizable), so when data governance is 60% complete, the core scheduling large language model pre-issues reserve tokens to the feature engineering agent to achieve pipeline parallelism.

[0070] Furthermore, the core scheduling large language model can set maximum execution time thresholds for each subtask. As an example, the maximum execution time threshold for the data governance subtask is 15 minutes, for the feature engineering subtask it's 10 minutes, for the model training subtask it's 30 minutes, and for the prediction execution subtask it's 5 minutes. When an agent exceeds the corresponding maximum execution time threshold, the core scheduling large language model triggers task reallocation, migrating the subtask to a backup computing node.

[0071] Step 203: Call the model selection agent to determine the fit of multiple preset basic prediction models based on the data to be predicted and the load prediction requirements, and select candidate prediction models from the multiple preset basic prediction models based on the fit.

[0072] After the large language model distributes subtasks, the model selection subtask can be executed by the model selection agent. Specifically, the model training subtask can include the model selection subtask and the model hyperparameter training subtask.

[0073] As an example, a power load forecasting system may include a time-series basic model and a large language model-enhanced forecasting model. The time-series basic model may include an Autoregressive Integrated Moving Average (ARIMA) model, a Long Short-Term Memory (LSTM) network model, a Temporal Convolutional Network (TCN) model, a Transformer model, and an Informer model. The large language model-enhanced forecasting model includes a time-series large language model fine-tuned based on the power domain. In other words, multiple preset basic forecasting models may include ARIMA, LSTM, TCN, Transformer, Informer, and a time-series large language model fine-tuned based on the power domain.

[0074] The model selection agent can determine the fitness of each preset basic prediction model based on the data to be predicted and the load prediction requirements. Then, the preset basic prediction model with the highest fitness can be used as a candidate prediction model, or the top few preset basic prediction models with relatively high fitness can be used as candidate prediction models.

[0075] Step 204: Call the hyperparameter optimization agent to optimize the model parameters of the candidate prediction model and obtain the target prediction model.

[0076] Furthermore, the hyperparameter optimization agent performs the model hyperparameter training subtask. Specifically, the hyperparameter optimization agent optimizes the model parameters of candidate prediction models to obtain the target prediction model. As an example, the hyperparameter optimization agent can use the Bayesian optimization algorithm to optimize the model parameters of each candidate prediction model to obtain an optimized prediction model, and the optimized prediction model with the smallest error rate is determined as the target prediction model.

[0077] Step 205: Invoke the prediction execution agent to perform power load prediction on the data to be predicted based on the target prediction model and load prediction requirements, and obtain the power load prediction results.

[0078] Furthermore, the prediction execution agent performs prediction sub-tasks. Specifically, the prediction execution agent performs power load prediction on the data to be predicted based on the target prediction model and load prediction requirements, and obtains the power load prediction results.

[0079] In the above implementation process, the large language model generates the data to be predicted based on historical structured and unstructured data. This ensures a rich variety of data types involved in power load forecasting, effectively reducing the problem of low prediction accuracy caused by a single data type and significantly improving the accuracy of power load forecasting results. Furthermore, the large language model divides the forecasting task into multiple sub-tasks based on load forecasting requirements and assigns these sub-tasks to multiple agents for execution. Specifically, the model selection agent obtains the fit of multiple preset basic forecasting models based on the data to be predicted and load forecasting requirements, and then determines candidate forecasting models from these models based on their fit, achieving adaptive model selection and reducing the problem of low prediction accuracy caused by a single forecasting model. The hyperparameter optimization agent optimizes the model parameters of the candidate forecasting models to obtain the target forecasting model, improving the forecasting performance and thus facilitating higher accuracy in load forecasting results. Finally, the prediction execution agent performs forecasting based on the target forecasting model and load forecasting requirements to obtain the prediction results. By breaking down a prediction task into multiple sub-tasks and having each sub-task executed by multiple agents, the coupling of the power load prediction system is reduced. This allows each agent to execute its corresponding sub-task in a combined serial and parallel manner, improving the efficiency of power load prediction. Furthermore, by using multiple types of data in power load prediction and adaptively selecting prediction models to predict the power load of the power system, the accuracy of power load prediction results is effectively improved.

[0080] In one embodiment, based on the data to be predicted and the load forecasting requirements, the fit of multiple preset basic forecasting models is determined, and a candidate forecasting model is determined from the multiple preset basic forecasting models based on the fit. Figure 3 The method shown is used to obtain, Figure 3 This is a schematic flowchart of a method for determining a candidate prediction model provided in an embodiment of this application, as shown below. Figure 3 As shown, the method may include the following steps:

[0081] Step S1: Determine the historical prediction accuracy and prior fit of the target basic prediction model.

[0082] Among them, the target basic prediction model is any one of multiple preset basic prediction models; the historical prediction accuracy is the prediction accuracy of the data to be predicted based on the target basic prediction model and the load prediction requirements in the historical prediction data; and the prior fit is the fit of the large language model determined based on the domain prior knowledge.

[0083] Step S2: Determine the fitness of the target basic prediction model based on historical prediction accuracy, prior fitness, and computational complexity of the target basic prediction model.

[0084] Step S3: Based on the fit of multiple preset basic prediction models, determine at least two candidate prediction models from the multiple preset basic prediction models.

[0085] For example, in the process of determining candidate prediction models, the model selection agent can select any one of the multiple preset basic prediction models as the target basic prediction model.

[0086] Then, the prediction accuracy when predicting data of the same type as the data to be predicted based on the target basic prediction model and load prediction requirements in the historical prediction data is determined as the historical prediction accuracy; the fitness of the large language model determined based on domain prior knowledge is determined as the prior fitness.

[0087] Furthermore, the historical prediction accuracy, prior fit, and computational complexity of the target-based prediction model are weighted and summed to obtain the fit of the target-based prediction model.

[0088] As an example, the preset basic prediction model can be obtained by taking into account the time series length, nonlinearity, and data dimension of the data to be predicted, combined with the prediction cycle and accuracy requirements of the sub-task, through the following expression (1). Adaptability :

[0089] (1);

[0090] in, For the first The fit of a pre-defined basic prediction model. For the first The historical prediction accuracy of a pre-defined basic prediction model on the same type of dataset. For the first The computational complexity of a pre-defined basic prediction model, The first language model given by the large language model based on domain prior knowledge The prior fit of a pre-defined basic prediction model , , Let be the weight coefficient, and satisfy... , The value ranges from 0.4 to 0.6. The value ranges from 0.2 to 0.3. The value ranges from 0.2 to 0.3.

[0091] Furthermore, the top three preset basic prediction models with the highest fitness can be selected as candidate prediction models, which makes it easier for the hyperparameter optimization agent to determine the target prediction model from the candidate prediction models.

[0092] As an example, the feature engineering agent performs feature selection and dimensionality reduction on the data to be predicted output by the data governance agent, thereby outputting feature data. The attributes of this feature data may include a time series length of 168 time steps (7 days × 24 hours), a non-linearity calculated to be 0.78 (based on the Lyapunov exponent), and a data dimension of 128. The subtask requires a prediction period of 24 hours and an accuracy requirement of MAPE ≤ 3%. The model selection agent calls a prediction model library, which includes basic time series models such as ARIMA, LSTM (2 layers, 256 hidden layer dimensions), TCN (3 residual blocks, 3 convolutional kernels), Transformer (4 encoder layers, 8 attention heads), and Informer (ProbSparse attention, 2 encoder layers), as well as a time series large language model (Time-LLM architecture, input sequence length 336, output length 96) fine-tuned based on the power domain.

[0093] The model selection agent calculates the fitness of each preset basic prediction model according to the above expression (1): For the ARIMA model, the time series length fitness is 0.6 (short sequence advantage, but 168 steps is a medium length), the nonlinear fitness is 0.3 (ARIMA has weak nonlinear fitting ability), the data dimension fitness is 0.8 (low dimension advantage), and the overall fitness is 0.52; for the LSTM model, the time series length fitness is 0.85, the nonlinear fitness is 0.82, the data dimension fitness is 0.75, and the overall fitness is 0.81. For the Transformer model, the time series length fit is 0.9, the nonlinearity fit is 0.88, the data dimension fit is 0.85, and the overall fit is 0.88. For the Informer model, the time series length fit is 0.92, the nonlinearity fit is 0.85, the data dimension fit is 0.9, and the overall fit is 0.89. For the power sector fine-tuning time series large language model, the time series length fit is 0.95, the nonlinearity fit is 0.9, the data dimension fit is 0.88, and the overall fit is 0.91. The top three models with the highest overall fit (power sector fine-tuning time series large language model, Informer, and Transformer) are selected as candidate prediction models.

[0094] In the above implementation process, the model selection agent determines the fitness of each preset basic prediction model based on its historical prediction accuracy, prior fitness, and computational complexity. This multi-dimensional approach improves the fitness of the prediction model selection process. Furthermore, based on the fitness of each preset basic prediction model, at least two candidate prediction models are selected, facilitating the selection of the target prediction model from these candidates and enhancing the error tolerance and stability of the prediction model selection process.

[0095] In one embodiment, optimizing the model parameters of the candidate prediction model to obtain the target prediction model may include the following steps:

[0096] Step 1: Based on domain prior knowledge, determine the range of model parameters for each candidate prediction model.

[0097] Step 2: Iteratively train the corresponding candidate prediction models based on the model parameter range until the corresponding candidate prediction models converge, and obtain the converged prediction model corresponding to each candidate prediction model.

[0098] Step 3: Among all the converged prediction models, the converged prediction model with the smallest model error is determined as the target prediction model.

[0099] For example, after the model selection agent selects at least two candidate prediction models, the model parameters of each candidate prediction model can be optimized by the hyperparameter optimization agent to obtain the target prediction model.

[0100] Specifically, the hyperparameter optimization agent can generate the model parameter range and corresponding initial parameters for each candidate prediction model based on the prior domain knowledge of the core scheduling large language model. This method can effectively shorten the optimization range of model parameters.

[0101] Furthermore, a Bayesian optimization algorithm can be used to iteratively optimize the corresponding candidate prediction models within the parameter range of each candidate prediction model. The number of iterations can be set to 50-200 times, with the optimization objective being to minimize the Mean Absolute Percentage Error (MAPE). After every 10 iterations, the convergence of the current optimization result is determined by the core scheduling large language model. If it is determined to have converged prematurely, the optimization is terminated, and the optimal parameter combination is output. If it has not converged, the optimization step size is adjusted, and iteration continues until the corresponding candidate prediction model converges, thus obtaining the converged prediction model corresponding to each candidate prediction model. Further, the training and validation of the model to be optimized are completed based on the optimal parameter combination of each converged prediction model, and the converged prediction model with the lowest MAPE on the validation set is selected as the target prediction model.

[0102] As an example, the power domain fine-tuning time series large language model, Informer, and Transformer are used as candidate prediction models. Furthermore, the core scheduling large language model, based on prior knowledge, generates model parameters for the power domain fine-tuning time series large language model, including: learning rate [1e-5, 1e-3], batch size [16, 64], dropout rate [0.1, 0.3], and LoRA rank [4, 32]. The corresponding initial parameters are: learning rate 5e-4, batch size 32, dropout rate 0.1, and LoRA rank 16. Therefore, the parameter optimization space is reduced by 55% compared to the full space. Then, Bayesian optimization (Gaussian process surrogate model, acquisition function is EI) is used for iterative optimization, with the number of iterations set to 100, and the optimization objective is to minimize the validation set MAPE. Every 10 iterations, the core scheduling large language model performs convergence determination: in the 30th iteration, if the MAPE improvement is less than 0.05% for three consecutive iterations, it is determined to be early convergence, the optimization is terminated, and the optimal hyperparameter combination is output. The parameters at convergence of the candidate prediction model include: learning rate 2.3e-4, batch size 48, dropout rate 0.15, and LoRA rank 24. Furthermore, based on the optimal hyperparameter combination, the model is trained, and the MAPE on the validation set is 2.1%, the lowest among the three candidate prediction models. Therefore, this fine-tuned time-series large language model for the power sector is selected as the target prediction model.

[0103] In the above implementation process, the parameter range of each candidate prediction model is determined through the prior knowledge of the large language model. This effectively narrows the optimization space, and the candidate prediction models are iteratively trained based on the determined parameter range, improving the efficiency of training the candidate prediction models. This continues until the corresponding candidate prediction models converge, and the converged prediction model with the smallest error is determined as the target prediction model, thus improving the accuracy of the target prediction model.

[0104] In one embodiment, the plurality of agents further includes an error tracing agent; the method further includes the following steps:

[0105] Step 1: Invoke the prediction execution agent to perform power load prediction based on the target prediction model and load prediction requirements, and obtain the initial prediction results.

[0106] Step 2: Call the error tracing agent to determine the prediction error based on the measured results corresponding to the initial prediction result and the data to be predicted in the historical data; and call the large language model to analyze the source of the error based on the prediction error to obtain the error type; and correct the initial prediction result based on the error type to obtain the power load prediction result.

[0107] For example, the multiple agents in the power load forecasting system may also include an error tracing agent, which can perform error tracing and anomaly identification on the forecast results output by the forecast execution agent. Specifically, the forecast execution agent performs power load forecasting on the data to be forecasted based on the target forecasting model and load forecasting requirements, and obtains initial forecast results.

[0108] Furthermore, the error tracing agent calculates the error between the initial prediction result and the actual measurement result to obtain the prediction error. When the prediction error exceeds a preset error threshold, the anomaly identification process is triggered. As an example, the preset error threshold can be set to 5% to 15%.

[0109] Furthermore, based on the prediction error, a large language model is invoked to analyze the source of the error and obtain the error type. The initial prediction result is then corrected according to the error type to obtain the power load prediction result.

[0110] As an example, when the prediction error exceeds a preset error threshold, the core scheduling big language model performs semantic reasoning and attribution analysis on the error source, distinguishing between three error types: data anomaly, model fit anomaly, and external sudden factor anomaly, and generating corresponding error correction factors for different error types. Then, the power load forecast result is obtained by expressing (2):

[0111] (2);

[0112] in, The corrected power load forecast result at time t. The initial prediction result at time t, Let be the error correction factor at time t. The value range is 0.8 to 1.2.

[0113] Specifically, the predictive agent can invoke the target prediction model, input standardized feature data from the past 168 hours, and the model outputs point predictions for the next 24 hours (one prediction per hour, for a total of 24 points) and interval predictions (upper and lower bounds) with a 95% confidence level. For example, the predicted load at 14:00 tomorrow is 1523.5MW, with a 95% confidence interval of [1482.3MW, 1564.7MW].

[0114] The error tracing agent compares the predicted value with the measured value: if the measured load at 14:00 tomorrow is 1610.2MW, then the relative error is (1523.5-1610.2) / 1610.2=-5.38%, and the absolute value exceeds the preset error threshold of 5%, triggering the anomaly identification process. Based on the semantic reasoning ability of the large language model, the error tracing agent retrieves the related data in the shared memory module and finds that a "temporary staggered peak electricity consumption notice" was issued during this period, requiring some industrial enterprises to limit power production from 14:00 to 16:00. This policy text was not included in the feature dataset in time. The error tracing agent generates an error correction factor λ=1.057 (based on the load attenuation coefficient of the policy impact), and corrects the predicted value in real time through the above expression (2): the corrected power load prediction result = 1523.5×1.057=1610.3MW, and the error with the measured value is reduced to 0.006%.

[0115] In the above implementation process, by analyzing the sources of error, the error type is obtained, and the prediction results output by the prediction execution agent are corrected according to the error type. Different correction strategies can be adopted in a targeted manner to improve the accuracy of the correction results, that is, to improve the accuracy of the power load prediction results.

[0116] In one embodiment, the plurality of agents further includes a security verification agent; the method may also include the following steps:

[0117] Step 1: Call the error tracing agent to correct the initial prediction result based on the error type and obtain the corrected prediction result.

[0118] Step 2: Call the safety verification agent to verify the corrected prediction result. If the verification result is that the verification fails, then repeatedly execute the steps of calling the error tracing agent to correct the corrected prediction result based on the measured result, and calling the safety verification agent to verify the corrected prediction result, until the verification result is that the verification passes, and then determine the corrected prediction result that passes the verification as the power load prediction result.

[0119] For example, after the error tracing agent corrects the initial prediction result, a corrected prediction result is obtained. The corrected prediction result can also be verified by a security verification agent.

[0120] Specifically, security verification can include compliance verification and physical boundary verification. Regarding compliance verification: the error tracing agent can verify whether the corrected prediction results comply with relevant management regulations. For example, it verifies whether the corrected prediction results meet the requirements of the "Power Grid Dispatch Management Regulations" regarding the reporting time limit for day-ahead load forecasts (prediction results must be submitted before 12:00 on D-1 day) and the format specifications (must include 96 points of 15-minute level prediction values). Regarding physical boundary verification, the historical maximum load value of the error tracing agent is 1850MW, and the minimum is 980MW. The corrected prediction value of 1610.3MW is within the range of [882MW, 2035MW] (i.e., ±10% of the historical maximum value), and the verification passes, outputting the final power load prediction result. If the prediction value for a certain period is 2200MW, exceeding the physical boundary upper limit of 2035MW, the verification fails, triggering the error tracing agent to re-analyze whether there is abnormal weather or special event causing a sudden increase in load during that period, and then re-corrects until the verification passes.

[0121] In the above implementation process, through the iterative mechanism of error correction and safety verification, while pursuing high precision, the bottom line of safe operation is further ensured, thereby improving the robustness of the power system under extreme operating conditions.

[0122] In one embodiment, generating the data to be predicted based on historical structured data and historical unstructured data in the power system may include the following steps:

[0123] Step 1: Extract semantic features from historical unstructured data to obtain multiple semantic feature vectors.

[0124] Step 2: Weight and fuse multiple semantic feature vectors to obtain fused semantic features.

[0125] Step 3: Combine the fused semantic features with historical structured data, perform temporal alignment and normalization processing to obtain the data to be predicted.

[0126] For example, a domain-fine-tuned large language model can extract semantic features and perform structured transformation on unstructured data. Specifically, the domain-fine-tuned large language model performs entity recognition, relation extraction, and correlation determination on unstructured data, generating multiple semantic feature vectors.

[0127] As an example, the weight coefficients of the semantic feature vector can be determined by the following expression (3):

[0128] (3);

[0129] in, For the first The weight coefficients of each semantic feature, For the first A semantic feature vector, The core associated feature vector of the load, This is the cosine similarity calculation function. This represents the total number of semantic feature vectors.

[0130] Furthermore, multiple semantic feature vectors are weighted and fused according to their corresponding weight coefficients to obtain fused semantic features.

[0131] It will also integrate semantic features with historical structured data, perform time-series alignment, and... Normalization is performed to obtain the data to be predicted.

[0132] Specifically, the domain-fine-tuned large language model can be a model obtained by LoRA fine-tuning based on the Llama2-7B architecture and using power domain corpus. In the semantic feature extraction stage, the domain-fine-tuned large language model performs entity recognition on the extreme weather warning text "Affected by the subtropical high pressure, there will be continuous high temperatures above 38℃ for the next three days," extracting key entities such as "subtropical high pressure" (weather system entity), "three days" (time entity), and "38℃" (temperature entity); it determines the causal relationship between "high temperature weather" and "load increase" through relation extraction, with a correlation degree of 0.92; for the user electricity consumption behavior text "Industrial park suspends production from 14:00 to 18:00 due to equipment maintenance," it extracts "industrial park" (user type entity), "equipment maintenance" (behavioral entity), and "14:00-18:00" (time period entity), with a correlation degree of 0.85. The generated semantic feature vector has a dimension of 768. After calculation by the above expression (3), the weight of the semantic feature vector of the "extreme weather" category is 0.35, and the weight of the semantic feature vector of the "user behavior" category is 0.28. After Min-Max normalization, it is fused with the structured time series data to generate standardized data to be predicted. The data dimension is unified to 128, and the time window is the past 168 hours (7 days).

[0133] In the above implementation process, a large language model is used to perform entity recognition, relation extraction, and semantic feature extraction on historical unstructured data, transforming it into a standardized structured feature vector. This effectively solves the problem of the difficulty in coordinating the use of multi-source heterogeneous data in traditional power load forecasting, significantly improves the information richness of the feature dataset and the quality of the prediction input, and further improves the accuracy of power load forecasting results.

[0134] In one embodiment, the method further includes invoking a large language model to perform the following steps:

[0135] Step 1: Monitor the operating status of multiple agents and obtain the monitoring results for each agent.

[0136] Step 2: Optimize the tasks in the agents whose monitoring results are abnormal.

[0137] Step 3: Update the operating parameters of each agent based on the preset update conditions.

[0138] Step 4: If the prediction results in multiple prediction periods meet the preset conditions, the operating parameters of each agent are determined as standard operating parameters, and standardized power load prediction is performed in the power load prediction system based on the standard operating parameters.

[0139] For example, the core scheduling large language model can monitor the entire power load forecasting process, that is, monitor the operating status of multiple agents and obtain the monitoring results of each agent. The monitoring results can include the execution time of each agent's sub-tasks, data quality, model accuracy, error distribution, and the collaborative efficiency of multiple agents. When the monitoring result of a certain agent is abnormal, the task of the corresponding agent is optimized.

[0140] The system iteratively updates the agent's operating parameters based on preset update conditions. Specifically, the iterative update cycle can be set to a combination of a fixed cycle and a triggered update. The fixed cycle is 7 to 30 days, and the triggered update is triggered immediately when the relative error of the prediction result exceeds a preset error threshold for three consecutive time steps.

[0141] Furthermore, if the prediction results across multiple prediction periods meet the error threshold, the operating parameters of each agent can be defined as standard operating parameters, and standardized power load prediction can be performed in the power load prediction system based on these standard operating parameters. As an example, if the mean absolute percentage error (MAPE) is ≤3% and the maximum relative error is ≤8% over 10 consecutive prediction periods, the corresponding prediction process can be set as a fixed standardized template.

[0142] Specifically, the core scheduling big language model has a built-in power load forecasting domain knowledge base, which includes a data governance rule base, a model parameter base, a forecast scenario adaptation rule base, an anomaly handling rule base, and a security and compliance rule base. The multi-agent collaborative architecture also includes a shared memory module and a message bus. Each agent completes data interaction and instruction transmission through the message bus. The execution data, output results, and optimization parameters of all agents are synchronously stored in the shared memory module.

[0143] The core scheduling big language model can read all the data in the shared memory module in real time, realizing global monitoring of the execution status of each agent. When the execution of an agent is abnormal, the core scheduling big language model can immediately trigger the task reassignment and abnormal handling process. The load forecasting scenarios adapted to the method include residential load forecasting, commercial load forecasting, industrial load forecasting, park comprehensive load forecasting, and virtual power plant aggregated load forecasting.

[0144] For different prediction scenarios, the core scheduling big language model can call fixed standardized templates. Only the target prediction task and related data for the corresponding scenario need to be input, and the adaptive prediction process can be automatically built and deployed within 5 minutes.

[0145] As an example, after completing the entire process of a prediction cycle (7 days), the core scheduling large language model reads all execution data from the shared memory module and performs a multi-dimensional review analysis: In terms of execution time, data governance took an average of 12 minutes (threshold 15 minutes), feature engineering took an average of 8 minutes (threshold 10 minutes), and model training took an average of 28 minutes (threshold 30 minutes). It was identified that the model training phase was a bottleneck due to the large fluctuation in time consumption (standard deviation of 6 minutes) during the morning peak hours (9:00-11:00) caused by GPU resource competition; in terms of data quality, 15% of the extreme weather warning texts in the unstructured data had inconsistent timestamp formats, affecting the accuracy of semantic feature extraction; in terms of model accuracy, the average MAPE over 7 days was 2.3%, and the maximum relative error was 6.8%; in terms of error distribution, the errors were mainly concentrated during holidays (Dragon Boat Festival) due to sudden changes in residential electricity consumption patterns; in terms of agent collaboration efficiency, serial tasks accounted for 40%, parallel tasks accounted for 60%, and the average token waiting time was 1.2 minutes.

[0146] Based on the review results, the following process optimization strategies were generated: To address model training bottlenecks, the task scheduling strategy was adjusted, prioritizing the allocation of model training subtasks to A100 nodes in the GPU resource pool, and postponing training tasks during the morning peak hours to midday; to address data quality issues, the data governance rule base was updated, adding rules for automatic timestamp format conversion; to address holiday prediction bias, a "holiday identifier" feature dimension (including holiday type, number of days until the holiday, etc.) was added during the feature engineering stage. Iterative update triggering: A fixed cycle of 15 days was set for a regular iterative update; simultaneously, if the relative error exceeds 5% for three consecutive time steps (e.g., three consecutive hours), a triggered update is immediately initiated, and the core scheduling large language model automatically adjusts the error threshold to 8% and re-executes steps 201-205 of this application embodiment.

[0147] Standardized template solidification verification: A residential load forecasting process achieved an average MAPE of 2.1% (≤3%) and a maximum relative error of 5.2% (≤8%) over 10 consecutive forecasting periods (10 weeks), meeting the solidification conditions. The core scheduling big language model solidified this process into a "Residential Day-ahead Load Forecasting Standard Template." The template includes pre-configured data governance rules (residential electricity consumption characteristic cleaning rules), model parameters (optimal hyperparameter combination of the power sector fine-tuning time-series big language model), interaction rules (data governance → feature engineering → model training serial, forecast execution → error verification → security verification parallel), and security verification thresholds (physical boundary ±10%). Subsequently, for similar residential load forecasting scenarios, dispatchers only need to input the target area code and forecast period, and the system can automatically build and deploy the forecasting process within 3 minutes.

[0148] For residential load forecasting scenarios, the core scheduling big data model calls the "Residential Day-a-Day Load Forecasting Standard Template," inputting aggregated residential electricity consumption data and corresponding meteorological data for a certain city. The system completes the process setup within 4 minutes, forecasting the load for the next 24 hours with an average MAPE of 2.1%. For commercial load forecasting scenarios, the "Commercial Ultra-Short-Term Load Forecasting Template" is called, inputting electricity consumption data of commercial complexes, business hours characteristics, and pedestrian density data. The system completes the setup within 5 minutes, forecasting the load for the next 4 hours with an average MAPE of 1.8%. For industrial load forecasting scenarios, the "Industrial Day-a-Day Load Forecasting Template" is called, inputting electricity consumption data of industrial parks, production plan text, and equipment maintenance records. The system completes the process setup within 4 minutes, forecasting the load for the next 4 hours with an average MAPE of 1.8%. The system can be set up within 5 minutes, predicting the load for the next 24 hours with an average MAPE of 2.5%. For the comprehensive load forecasting scenario in the industrial park, the "Multi-Energy Collaborative Forecasting Template for the Industrial Park" is invoked. By inputting mixed electricity consumption data for residents, businesses, and industries within the park, as well as distributed photovoltaic output data, the system can be set up within 5 minutes, predicting the net load for the next 24 hours with an average MAPE of 2.8%. For the virtual power plant aggregated load forecasting scenario, the "Virtual Power Plant Aggregated Forecasting Template" is invoked. By inputting real-time status data of electric vehicle charging piles, energy storage systems, and controllable loads within the aggregation range, as well as user response behavior text, the system can be set up within 5 minutes, predicting the adjustable load capacity for the next 4 hours with an average MAPE of 3.2%. Each scenario template automatically matches the corresponding dedicated functional intelligent agent combination and interaction rules through the core scheduling large language model, achieving rapid adaptation and deployment of the forecasting process.

[0149] In the above implementation process, the core scheduling big language model performs multi-dimensional review and analysis of the entire process execution data, supports an iterative update mechanism that combines fixed period and trigger-based updates, and solidifies the verified process into a standardized template, realizing the continuous evolution of prediction capabilities and rapid reuse across scenarios.

[0150] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0151] Based on the same inventive concept, this application also provides a load forecasting device that integrates large language models and intelligent agent technology for implementing the load forecasting method integrating large language models and intelligent agent technology described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more load forecasting device embodiments integrating large language models and intelligent agent technology provided below can be found in the limitations of the load forecasting method integrating large language models and intelligent agent technology described above, and will not be repeated here.

[0152] In one exemplary embodiment, Figure 4 This is a schematic diagram of a load forecasting device integrating large language model and intelligent agent technology, provided in an embodiment of this application. This device can be applied to a multi-agent power load forecasting system, which includes a large language model and multiple intelligent agents, such as... Figure 4 As shown, the device may include:

[0153] The data processing module 401 is used to call the large language model to generate the data to be predicted based on the historical structured data and historical unstructured data in the power system. The historical structured data is the data formed by monitoring the operation process of the power system based on historical time series, and the historical unstructured data is the data that uses natural language as a carrier to constrain the historical operation process of the power system.

[0154] The task allocation module 402 is used to call the large language model, divide the prediction task into multiple sub-tasks based on the load prediction demand of the power system, and allocate the multiple sub-tasks to multiple intelligent agents; the multiple intelligent agents include at least a model selection intelligent agent, a hyperparameter optimization intelligent agent, and a prediction execution intelligent agent;

[0155] The model selection module 403 is used to call the model selection agent to determine the fit of multiple preset basic prediction models based on the data to be predicted and the load prediction requirements, and to determine the candidate prediction model from the multiple preset basic prediction models based on the fit.

[0156] The parameter optimization module 404 is used to call the hyperparameter optimization agent to optimize the model parameters of the candidate prediction model and obtain the target prediction model.

[0157] The prediction module 405 is used to call the prediction execution agent to perform power load prediction on the data to be predicted based on the target prediction model and prediction requirements, and obtain the power load prediction result.

[0158] In one embodiment, the model selection module 403 is specifically used for:

[0159] Determine the historical prediction accuracy and prior fit of the target basic prediction model; the target basic prediction model is any one of multiple preset basic prediction models; the historical prediction accuracy is the prediction accuracy of the data to be predicted based on the target basic prediction model and load prediction requirements in the historical prediction data; the prior fit is the fit of the large language model determined based on domain prior knowledge.

[0160] The fitness of the target basic prediction model is determined based on historical prediction accuracy, prior fitness, and computational complexity of the target basic prediction model.

[0161] Based on the fit of multiple preset basic prediction models, at least two candidate prediction models are determined from the multiple preset basic prediction models.

[0162] In one embodiment, the parameter optimization module 404 is specifically used for:

[0163] Based on domain prior knowledge, the range of model parameters for each candidate prediction model is determined;

[0164] Based on the range of model parameters, the corresponding candidate prediction models are iteratively trained until the corresponding candidate prediction models converge, thus obtaining the converged prediction model corresponding to each candidate prediction model.

[0165] Among all converged prediction models, the one with the smallest model error is selected as the target prediction model.

[0166] In one embodiment, the plurality of agents further includes an error tracing agent; the prediction module 405 is specifically used for:

[0167] The prediction execution agent is invoked to perform power load prediction based on the target prediction model and load prediction requirements, and the initial prediction results are obtained.

[0168] The error tracing agent is invoked to determine the prediction error based on the initial prediction result and the corresponding measured results in historical data. Then, based on the prediction error, a large language model is invoked to analyze the source of the error and obtain the error type. Based on the error type, the initial prediction result is corrected to obtain the power load prediction result.

[0169] In one embodiment, the plurality of agents further includes a security verification agent; the prediction module 405 is also configured to:

[0170] The error tracing agent is invoked to correct the initial prediction result based on the error type, thus obtaining the corrected prediction result.

[0171] The safety verification agent is invoked to verify the corrected prediction result. If the verification result is that the verification fails, the steps of invoking the error tracing agent to correct the corrected prediction result based on the measured result and invoking the safety verification agent to verify the corrected prediction result are executed repeatedly until the verification result is that the verification passes. The corrected prediction result that passes the verification is then determined as the power load prediction result.

[0172] In one embodiment, the data processing module 401 is specifically used for:

[0173] Semantic features are extracted from historical unstructured data to obtain multiple semantic feature vectors;

[0174] Multiple semantic feature vectors are weighted and fused to obtain fused semantic features;

[0175] The semantic features and historical structured data are integrated, and then time-series aligned and normalized to obtain the data to be predicted.

[0176] In one embodiment, the data processing module 401 is further configured to:

[0177] The operational status of multiple agents is monitored, and the monitoring results of each agent are obtained;

[0178] Optimize tasks in agents whose monitoring results are abnormal;

[0179] Based on preset update conditions, the operating parameters of each agent are updated;

[0180] If the prediction results meet the preset conditions in multiple prediction periods, the operating parameters of each agent are determined as standard operating parameters, and standardized power load prediction is performed in the power load prediction system based on the standard operating parameters.

[0181] The various modules in the aforementioned load forecasting device integrating large language models and intelligent agent technology can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0182] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, Figure 5 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application, such as... Figure 5 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores historical data of the power system. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a load forecasting method that integrates large language models and intelligent agent technology.

[0183] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0184] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the load prediction method that integrates large language models and agent technology in any of the above embodiments.

[0185] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the load prediction method that integrates large language models and agent technology in any of the above embodiments.

[0186] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0187] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0188] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0189] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A load prediction method integrating large language models and intelligent agent technology, characterized in that, The method is applied to a multi-agent power load forecasting system, which includes a large language model and multiple agents. The method includes: The large language model is invoked to generate data to be predicted based on historical structured data and historical unstructured data in the power system; wherein, the historical structured data is data formed by monitoring the operation process of the power system based on historical time series, and the historical unstructured data is data constrained by the historical operation process of the power system using natural language as a carrier; The large language model is invoked, and the prediction task is divided into multiple sub-tasks based on the load prediction requirements of the power system. The multiple sub-tasks are then assigned to multiple intelligent agents. The multiple intelligent agents include at least a model selection agent, a hyperparameter optimization agent, and a prediction execution agent. The model selection agent is invoked to determine the fit of multiple preset basic prediction models based on the data to be predicted and the load prediction requirements, and to determine candidate prediction models from the multiple preset basic prediction models based on the fit. The hyperparameter optimization agent is invoked to optimize the model parameters of the candidate prediction model, thereby obtaining the target prediction model; The prediction execution agent is invoked to perform power load prediction on the data to be predicted based on the target prediction model and the load prediction requirements, thereby obtaining the power load prediction result.

2. The method according to claim 1, characterized in that, The step of determining the fit of multiple preset basic prediction models based on the data to be predicted and the load prediction demand, and determining candidate prediction models from the multiple preset basic prediction models based on the fit, includes: Determine the historical prediction accuracy and prior fit of the target basic prediction model; the target basic prediction model is any one of the multiple preset basic prediction models; the historical prediction accuracy is the prediction accuracy of the data to be predicted based on the target basic prediction model and the load prediction demand in historical prediction data; the prior fit is the fit of the large language model determined based on domain prior knowledge. The fitness of the target basic prediction model is determined based on the historical prediction accuracy, the prior fitness, and the computational complexity of the target basic prediction model. Based on the fit of the multiple preset basic prediction models, at least two candidate prediction models are determined from the multiple preset basic prediction models.

3. The method according to claim 2, characterized in that, The optimization of the model parameters of the candidate prediction model to obtain the target prediction model includes: Based on domain prior knowledge, the range of model parameters for each candidate prediction model is determined; Based on the range of model parameters, the corresponding candidate prediction models are iteratively trained until the corresponding candidate prediction models converge, thereby obtaining the converged prediction model corresponding to each candidate prediction model. Among the converged prediction models, the converged prediction model with the smallest model error is determined as the target prediction model.

4. The method according to claim 1, characterized in that, The plurality of said intelligent agents also includes an error tracing intelligent agent; the method further includes: The prediction execution agent is invoked to perform power load prediction on the data to be predicted based on the target prediction model and the load prediction demand, and an initial prediction result is obtained. The error tracing agent is invoked to determine the prediction error based on the initial prediction result and the measured results corresponding to the data to be predicted in historical data; the large language model is then invoked to analyze the source of the error based on the prediction error to obtain the error type; the initial prediction result is then corrected based on the error type to obtain the power load prediction result.

5. The method according to claim 4, characterized in that, The plurality of said intelligent agents also includes a security verification intelligent agent; the method further includes: The error tracing agent is invoked to correct the initial prediction result based on the error type, thereby obtaining a corrected prediction result; The security verification agent is invoked to verify the corrected prediction result. If the verification result is that the verification fails, the steps of invoking the error tracing agent to correct the corrected prediction result based on the measured result and invoking the security verification agent to verify the corrected prediction result are executed repeatedly until the verification result is that the verification passes. The corrected prediction result that passes the verification is then determined as the power load prediction result.

6. The method according to claim 1, characterized in that, The data to be predicted is generated based on historical structured and unstructured data from the power system, including: Semantic features are extracted from the historical unstructured data to obtain multiple semantic feature vectors; The multiple semantic feature vectors are weighted and fused to obtain the fused semantic features; The fused semantic features and the historical structured data are time-series aligned and normalized to obtain the data to be predicted.

7. The method according to any one of claims 1 to 5, characterized in that, The method also includes invoking a large language model to perform the following steps: The operating status of the multiple intelligent agents is monitored to obtain the monitoring results of each intelligent agent; Optimize tasks in agents whose monitoring results are abnormal; Based on preset update conditions, the operating parameters of each of the intelligent agents are updated; If the prediction results meet the preset conditions in multiple prediction periods, the operating parameters of each intelligent agent are determined as standard operating parameters, and power load standardization prediction is performed in the power load prediction system based on the standard operating parameters.

8. A load prediction device integrating large language model and intelligent agent technology, characterized in that, The device is applied to a multi-agent power load forecasting system, which includes a large language model and multiple agents. The device includes: The data processing module is used to call the large language model and generate data to be predicted based on historical structured data and historical unstructured data in the power system; wherein, the historical structured data is data formed by monitoring the operation process of the power system based on historical time series, and the historical unstructured data is data constrained by the historical operation process of the power system using natural language as a carrier; The task allocation module is used to call the large language model, divide the prediction task into multiple sub-tasks based on the load prediction requirements of the power system, and allocate the multiple sub-tasks to multiple intelligent agents; the multiple intelligent agents include at least a model selection intelligent agent, a hyperparameter optimization intelligent agent, and a prediction execution intelligent agent; The model selection module is used to call the model selection agent to determine the fit of multiple preset basic prediction models based on the data to be predicted and the load prediction requirements, and to determine the candidate prediction model from the multiple preset basic prediction models based on the fit. The parameter optimization module is used to call the hyperparameter optimization agent to optimize the model parameters of the candidate prediction model to obtain the target prediction model; The prediction module is used to call the prediction execution agent to perform power load prediction on the data to be predicted based on the target prediction model and the prediction requirements, and obtain the power load prediction result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

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 implements the steps of the method according to any one of claims 1 to 7.

Images