Large model fine tuning method, device and equipment of power system, storage medium and program product

Abstract

The invention relates to a large model fine tuning method, device and equipment of a power system, a storage medium and a program product, and relates to the field of power systems. Comprising the following steps: acquiring multi-modal time sequence data of a power system, inputting the multi-modal time sequence data into a pre-trained multi-modal large model, and acquiring a prediction sequence of each signal output by the multi-modal large model in a time dimension; scoring the importance of each signal at each time point based on the volatility index of the prediction sequence; screening the signal time sequence correlation points with the importance scores exceeding a preset threshold value; for each signal time sequence correlation point, calculating a deviation between a predicted value and a historical steady-state mean value, fusing deviation correlation characteristics of different signals at the same time point, and constructing a state vector; inputting the state vector into a reinforcement learning strategy network to generate a fine tuning action; and performing parameter updating on the corresponding model substructure in the multi-modal large model according to the fine tuning action. According to the method, the fine tuning efficiency can be improved, and the response efficiency and reliability of the multi-modal large model in a complex power scene are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a large model fine-tuning method and device for a power system, computer equipment, a computer readable storage medium and a computer program product. BACKGROUND

[0002] With the continuous improvement of the intelligence and digitization level of the power system, the data generated in the operation of the power grid presents significant characteristics such as multi-modal, high dimension, high frequency and strong correlation. The power time series signal not only includes basic parameters such as voltage, current, frequency, phase, power factor, load, but also contains multi-source observation signals from the substation protection system, monitoring terminal, PMU (PMU: Phasor Measurement Unit, Synchronous Phasor Measurement Device) measurement point and edge node. These data have obvious periodicity and random disturbance in time, and there are actual problems such as inconsistent sampling frequency, modal missing and signal drift in modal.

[0003] With the gradual deepening of the application of multi-modal large models in power prediction, anomaly detection and dispatching optimization, the contradiction between the strong representation ability of the pre-training model and the complex dynamics of the power scene also gradually becomes prominent. The traditional scheme usually adopts full model fine-tuning or local fixed window fine-tuning mode, which fails to adaptively adjust according to the real-time state and modal changes of the power signal, resulting in problems such as excessive model update range, unbalanced fine-tuning frequency, and waste of computing resources. On the other hand, in the field operation of the power system, signal anomalies often have cross-modal coordination and regional concentration characteristics, and simple fine-tuning scheduling based on static rules cannot accurately identify the time series region that needs to be optimized. In addition, the traditional method lacks a dynamic state modeling mechanism combining historical working conditions and model prediction behavior, and when there is a device disturbance, external load fluctuation or partial measurement point data failure, the model cannot actively perceive the prediction deviation and perform targeted optimization, thereby causing the performance of the model to decay in long-period operation. SUMMARY

[0004] Therefore, it is necessary to provide a large model fine-tuning method, device, computer equipment, computer readable storage medium and computer program product for a power system, which can improve the fine-tuning efficiency and improve the response efficiency and reliability of multi-modal large models in complex power scenarios.

[0005] In a first aspect, the present application provides a large model fine-tuning method for a power system, comprising:

[0006] obtaining multi-modal time series data of the power system; the multi-modal time series data includes voltage signals, current signals and frequency signals;

[0007] inputting the multi-modal time-series data into a pre-trained multi-modal large model to obtain a prediction sequence of each signal in a time dimension output by the multi-modal large model;

[0008] scoring importance of each signal at each time point based on a volatility index of the prediction sequence to obtain a scoring result;

[0009] filtering out a signal time-series correlation point with an importance score exceeding a preset threshold according to the scoring result to form a candidate set;

[0010] calculating a deviation between a predicted value of the signal time-series correlation point and a historical steady-state mean value, and fusing deviation correlation features among different signals at the same time point to construct a structured state vector for each signal time-series correlation point in the candidate set;

[0011] inputting the structured state vector into a reinforcement learning policy network to generate a corresponding fine-tuning action; wherein the fine-tuning action includes whether to perform fine-tuning, fine-tuning intensity, and fine-tuning strategy;

[0012] updating a model substructure corresponding to the signal time-series correlation point in the multi-modal large model according to the fine-tuning action.

[0013] In one embodiment, the scoring result of the importance of each signal at each time point based on the volatility index of the prediction sequence includes:

[0014] calculating a standard deviation of each signal in the time dimension based on the prediction sequence; the standard deviation is used to represent the volatility index of the prediction sequence;

[0015] normalizing the standard deviation based on a normalized upper and lower bound to obtain the importance score; the normalized upper and lower bound is obtained based on historical running data of a preset time period.

[0016] In one embodiment, the deviation correlation feature is obtained by calculating a standard deviation of deviations of all signals at the same time point.

[0017] In one embodiment, the structured state vector includes state vector items of each signal at each time point; the method further includes:

[0018] When constructing the structured state vector, if the deviation correlation feature among different signals at the target time point is higher than a preset value, the numerical value of the state vector item of all signals at the target time point is increased.

[0019] In one embodiment, inputting the structured state vector into the reinforcement learning policy network to generate the corresponding fine-tuning action includes:

[0020] calculating a prediction accuracy improvement rate of the signal time-series correlation point after performing fine-tuning based on historical fine-tuning records.

[0021] In a case where the prediction accuracy improvement rate is lower than the preset effective threshold, a penalty weight is applied to an action selection probability corresponding to the signal time sequence correlation point in the current decision of the reinforcement learning strategy network.

[0022] In one of the embodiments, the fine-tuning strategy includes low-rank adaptation incremental update or residual channel sparse update, and the parameter of the model substructure corresponding to the signal time sequence correlation point in the multi-modal large model is updated according to the fine-tuning action, including:

[0023] If the fine-tuning strategy is low-rank adaptation incremental update, a low-rank adapter is inserted on the encoding path corresponding to the signal time sequence correlation point and incremental training is performed;

[0024] If the fine-tuning strategy is residual channel sparse update, all parameters of the multi-modal large model except the target substructure are frozen, and the target substructure is updated by gradient descent with a sparse regularization term; the target substructure is the model substructure corresponding to the signal time sequence correlation point in the multi-modal large model.

[0025] In a second aspect, the present application also provides a large model fine-tuning device for a power system, including:

[0026] The acquisition module is configured to acquire multi-modal time sequence data of the power system; the multi-modal time sequence data includes voltage signals, current signals and frequency signals;

[0027] The prediction module is configured to input the multi-modal time sequence data into the pre-trained multi-modal large model, and acquire a prediction sequence of each signal in the time dimension output by the multi-modal large model;

[0028] The scoring module is configured to score the importance of each signal at each time point based on the volatility index of the prediction sequence, and obtain a scoring result;

[0029] The screening module is configured to screen out signal time sequence correlation points with importance scores exceeding a preset threshold according to the scoring result, and form a candidate set;

[0030] The fine-tuning module is configured to calculate the deviation between the prediction value and the historical steady-state mean value of each signal time sequence correlation point in the candidate set, and fuse the deviation correlation characteristics between different signals at the same time point to construct a structured state vector; input the structured state vector into the reinforcement learning strategy network to generate a corresponding fine-tuning action; wherein the fine-tuning action includes whether to perform fine-tuning, fine-tuning intensity and fine-tuning strategy; and update the parameters of the model substructure corresponding to the signal time sequence correlation point in the multi-modal large model according to the fine-tuning action.

[0031] 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 of the method described in the first aspect above.

[0032] 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 of the method described in the first aspect above.

[0033] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect above.

[0034] The aforementioned large-scale power system fine-tuning method, apparatus, computer equipment, computer-readable storage medium, and computer program product acquire multimodal time-series data of the power system. This multimodal time-series data includes voltage, current, and frequency signals. The multimodal time-series data is input into a pre-trained multimodal large-scale model to obtain the predicted sequences of each signal output by the model in the time dimension. Based on the volatility index of the predicted sequences, the importance of each signal at each time point is scored to obtain a scoring result. Based on the scoring results, time-series correlation points of signals with importance scores exceeding a preset threshold are selected to form a candidate set. For each time-series correlation point in the candidate set, the deviation between the predicted value and the historical steady-state mean of the correlation point is calculated, and the deviation correlation features between different signals at the same time point are fused to construct a structured state vector. The structured state vector is input into a reinforcement learning policy network to generate corresponding fine-tuning actions. These actions include whether to perform fine-tuning, the fine-tuning intensity, and the fine-tuning strategy. The parameters of the model substructure corresponding to the time-series correlation points in the multimodal large-scale model are updated according to the fine-tuning actions. By utilizing the prediction sequences of the large model itself as the criterion, modes and time segments that significantly impact model performance are dynamically identified, and a candidate set is constructed. This transforms the fine-tuning process from a full-scale mode to a locally focused mode driven by model behavior, avoiding the computational burden of full model updates and improving fine-tuning efficiency. Deviations are determined by combining model output with historical steady-state mean values, and the correlation features between different signals at the same time point are introduced to construct a structured state vector that simultaneously reflects single-mode deviations and system-level coordinated anomalies, enhancing the large model's sensitivity to changes in the overall operating state of the power grid. A reinforcement learning policy network is introduced to generate an action set containing execution decisions, fine-tuning intensity, and fine-tuning strategies based on the structured state vector. This allows model updates to respond differently based on the type and degree of anomalies perceived in real time, improving the targeting and intelligence of the tuning. By updating the parameters of the model substructures corresponding to the selected signal time-series correlation points, local prediction deviations can be corrected while maximizing the protection of the model's learned overall knowledge structure and generalization ability, ensuring stable and reliable performance during long-term deployment. This method can significantly improve the response efficiency and reliability of multimodal large models in complex power scenarios. Attached Figure Description

[0035] 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.

[0036] Figure 1This is a flowchart illustrating a method for fine-tuning a large model of a power system in one embodiment.

[0037] Figure 2 This is a structural block diagram of a large-scale power system fine-tuning device in one embodiment;

[0038] Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0039] 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.

[0040] 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.

[0041] In one exemplary embodiment, such as Figure 1 As shown, a method for fine-tuning a large model of a power system is provided. The method is illustrated using a server as an example, and includes:

[0042] Step 102: Obtain multi-modal time-series data of the power system; the multi-modal time-series data includes voltage signals, current signals, and frequency signals.

[0043] The multimodal time-series data originates from multimodal time-series observation sequences within a real power system. For example, a SCADA (Supervisory Control and Data Acquisition) system and PMU (Phasor Measurement Unit) devices deployed in a substation provide multi-source synchronization signals. Voltage modes are acquired by the SCADA system at a sampling frequency of 1Hz; current modes are sampled in alignment with the voltage modes using the same SCADA system; frequency modes are acquired by the PMU devices at an initial sampling frequency of 50Hz, which is then synchronized to 1Hz after low-pass filtering and uniform interpolation. During data acquisition, these data are transmitted to an edge server via a protocol (such as IEC 61850) and uniformly normalized on the edge server to form a tensor structure. .in Indicates the first The mode in the th ... The observations at each time point. All values ​​are normalized floating-point values, and the data window size can be set to... (i.e., continuous sampling for 5 minutes).

[0044] As can be understood, a power system refers to an electrical energy production and consumption system comprised of power generation, transmission, transformation, distribution, and consumption. Its core function is to ensure the safe, reliable, high-quality, and economical operation of electrical energy. In the context of intelligentization, power systems are typically equipped with various sensing, communication, and control devices to form an observable and controllable energy internet. Multimodal time-series data refers to the collection of voltage, current, frequency, and other signals that are synchronously or correlatedly acquired within the power system and have time-aligned relationships.

[0045] Step 104: Input the multimodal time series data into the pre-trained multimodal large model to obtain the predicted sequences of each signal output by the multimodal large model in the time dimension.

[0046] Among them, pre-trained multimodal large models are used. For each mode, predict the input sequence within the sliding time window and obtain its corresponding output sequence. ,in This is the window length, which can be 10 or other values. Optionally, for multimodal large models... This is a multimodal temporal modeler based on the standard Transformer architecture, consisting of three modality embedding layers, a position encoder, and two stacked multi-head self-attention encoders. The modality embedding layers map data from different sources to a unified feature dimension. The encoders perform modality fusion and temporal modeling on this feature dimension, finally outputting a continuous sequence of predicted values ​​for each modality. For example, to simplify deployment, the pre-trained large multimodal model uses a frozen parameter structure when providing prediction services. In this mode, the model only performs input-to-output computation without triggering any learning processes that might change its own parameters.

[0047] For example, suppose the current analysis time window is arrive The model supports current modes The predicted output for this segment is:

[0048] .

[0049] Among them, a pre-trained multimodal large model refers to a basic machine learning model that has been pre-trained on a large amount of general or domain-related data. A prediction sequence refers to the sequence of predicted values ​​at multiple future time points generated by the model for each target signal. In essence, multimodal time-series data containing various power signals is fed into a large model with powerful basic capabilities capable of processing multi-source data; the large model then uses the knowledge embedded in its parameters to calculate the possible changing trends of each signal over a future period and provides specific numerical prediction results in the form of a time series.

[0050] Step 106: Based on the volatility index of the predicted sequence, score the importance of each signal at each time point to obtain the score result.

[0051] Among them, volatility index is used to characterize the degree of drastic change or uncertainty of the predicted sequence within a local time window. Importance is used to characterize the criticality of a signal's predicted value at a specific future time point for the overall predictive reliability of the model, the safety of system operation, or subsequent fine-tuning decisions. The scoring result can be represented as a quantified matrix, with each element corresponding to a signal-time combination importance value.

[0052] This embodiment does not limit the method of measuring volatility indicators or the method of scoring their importance. For example, regarding the aforementioned prediction output... Calculate the standard deviation of the prediction output for the current window of this modality. ,get:

[0053] ;

[0054] in, It is a model In modality ,time The predicted output values ​​have been normalized. This represents the predictive volatility index of the model at this time point in this mode. This volatility index reflects the degree of uncertainty in the model's predictions in this region, indirectly indicating that the model may fail at this point.

[0055] To facilitate lateral comparisons between different modalities, Normalization is performed:

[0056] ;

[0057] Normalized upper and lower bounds and The data is derived from offline statistics of the running window data over the past 24 hours and is updated daily. That is, mode In time The relative predictive instability score is a real number between 0 and 1, with a larger value indicating greater fluctuation in the model's predictions at that location.

[0058] Step 108: Based on the scoring results, select signal time-series correlation points whose importance scores exceed a preset threshold to form a candidate set.

[0059] In this process, key modes and time periods that may lead to a decline in the predictive performance of large multimodal models are identified from multimodal time-series data collected from the power system. Importance scores are assigned based on the volatility of the model's predictive output, and a dynamic threshold is used to filter high-risk candidate sets. This helps identify which data regions are more worthy of inclusion in the fine-tuning process, avoiding the resource waste associated with full-scale processing.

[0060] For example, after the scoring is completed, all score values ​​exceeding a preset threshold are... (e.g., the 95th percentile) position is added to the candidate set. :

[0061] ;

[0062] Each element in the set Indicates the modalities and time points that need to be analyzed and evaluated in the future, such as This indicates that the frequency mode at 122 seconds is a location where the model output is significantly anomalous, requiring further analysis of its deviation in subsequent steps. This step does not simply use data characteristics (such as anomalous amplitudes) for filtering, but rather utilizes the model... This method identifies potential uncertainties from a model perspective by analyzing the output behavior. It is compatible with large model inference frameworks and has good deployment feasibility. In actual power system operation, this method can be deployed on the edge gateway server of a substation or the front-end server of the dispatch master station, providing real-time analysis and dynamic output in a rolling window manner. .

[0063] Step 110: For each signal time series correlation point in the candidate set, calculate the deviation between the predicted value of the signal time series correlation point and the historical steady-state mean, and fuse the deviation correlation features between different signals at the same time point to construct a structured state vector.

[0064] Among them, the candidate set of high uncertainty signal-time correlation points identified in the aforementioned steps Further detailed analysis is conducted to construct state vectors that can be used by subsequent reinforcement learning policy modules. This state vector should not only be able to express the modeling error of the current model at a specific mode and time point, but also comprehensively reflect the historical stability of the power system operation, the current degree of anomaly, and the potential correlation between modes, thereby enhancing the judgment basis of the policy network and making fine-tuning behavior more targeted and adaptable to power scenarios.

[0065] For example, the input to step 110 is a set This set is derived from the screening of unstable regions in the large model predictions by the aforementioned steps. Indicates the mode number (such as voltage, current, frequency). This indicates the time step index, i.e., a specific time point. For example, when the model shows a significant prediction fluctuation in the frequency mode at the 120th second, Will be added On these candidate sets, combined with the current model output... A multidimensional deviation index is calculated using the historical steady-state mean, which is then used to construct the state vector in a structured manner. .

[0066] For each Calculate the absolute deviation term between the current predicted value and the historical steady-state mean:

[0067] ;

[0068] in, It is a model For modes At any moment The predicted value comes from the output head of the Transformer encoder structure, whose front end is a modal embedding layer and a self-attention mechanism; It is modal In time The historical steady-state average, derived from the substation's SCADA system or PMU historical cache, is constructed, for example, by averaging the values ​​at the same time over the past 7 days to form a time-averaged reference. This method avoids misjudgments caused by single-point anomalies and reflects the highly regular operating trends of the power system. Deviation value It characterizes the degree of modeling mismatch for this mode.

[0069] However, in power systems, many anomalies are not driven by a single mode. For example, a rapid drop in voltage is often accompanied by an increase in current, or frequency fluctuations may be caused by load changes. Therefore, considering only single-mode deviations may overlook systemic modeling anomalies. To address this issue, this embodiment introduces a mode cooperative deviation suppression term, which calculates the standard deviation of each mode deviation at the same time point. This is used to measure the consistency of cross-modal deviations at that point in time (i.e., deviation correlation characteristics):

[0070] ;

[0071] in, It is the total number of modes (e.g., 3 modes). It is time The average of all modal deviations. This term measures the dispersion of the deviation distribution. In power system scenarios, if the deviations are highly concentrated at a certain point in time, it usually indicates that the prediction error occurs consistently across multiple modes, suggesting the existence of system-level anomalies, such as voltage drops or line tripping; conversely, if the deviations are discrete and inconsistent, it often indicates local anomalies or measurement errors.

[0072] Based on the above two indicators, define the final state vector term. for:

[0073] ;

[0074] in, To balance the coefficient (suggested value: 0.5~1.0), this item introduces an anomaly consistency enhancement mechanism: when the overall system deviation is relatively consistent at a given moment, the importance score of this mode at that moment is increased, amplifying its impact on the policy network. This method is rarely used in traditional policy modeling and is a mechanism specifically designed for power system scenarios in this solution, possessing strong domain adaptability.

[0075] All The above calculation Merge to form a structured state vector:

[0076] ;

[0077] This state vector not only contains the model's prediction error for the current input ( Essentially a weighted bias, it also retains accurate modal and temporal index information. This is used for action localization in subsequent strategy modules, forming a closed loop between data, state, and action. The vector structure adapts to candidate sets of any size. This approach can generate data in real-time in dynamically updated scenarios. It introduces cross-modal deviation consistency as an auxiliary signal, which not only improves the system's ability to identify collaborative anomalies but also maintains computational efficiency and deployability through a simple structure. Compared to using only single-point deviation, it proactively amplifies the response sensitivity to globally anomalous regions, enabling the policy controller to prioritize regions that may cause systemic misjudgments during fine-tuning scheduling, thus reducing the overall prediction failure risk.

[0078] Step 112: Input the structured state vector into the reinforcement learning policy network to generate corresponding fine-tuning actions; wherein, the fine-tuning actions include whether to perform fine-tuning, the fine-tuning intensity, and the fine-tuning policy.

[0079] Among them, the structured state vector constructed in the previous step Generate large models for controlling multimodal modes The decision-making actions for fine-tuning behavior specifically include: whether to perform fine-tuning, the scope and intensity of fine-tuning, and the specific fine-tuning strategy to be adopted. Its core task is to utilize reinforcement learning policy networks. In the context of complex and dynamic power system operation, for each candidate mode-time position It outputs quantitatively executable optimization instructions, ultimately achieving locally accurate and cost-effective model parameter update operations.

[0080] Input state vector Each triplet Indicates the first Each mode in time The modeling bias intensity is determined by the state vector constructed in the previous step by modeling the difference between the model prediction results and the historical operating condition average, combined with the degree of inter-modal cooperative anomalies. This vector not only possesses high-resolution spatiotemporal structure localization capabilities but also includes a model bias expression for anomaly-consistent scheduling. In power systems, short-term anomalies in a particular mode are not always sufficient to trigger a scheduling response; rather, a combination of system-level biases and historical experience is needed to make reasonable parameter adjustments. This step introduces a strategy learning module for scheduling value optimization based on this understanding.

[0081] To ensure the policy network The output actions have a clearly defined structure and control capability; each output action... Designed as a triple containing three items:

[0082] ;

[0083] in, Indicate whether or not Perform fine-tuning; This indicates the fine-tuning intensity; a higher value indicates a larger update step size. This indicates the choice of fine-tuning method; for example, 1 represents applying low-rank adaptive incremental updates (such as LoRA structure injection), and 2 represents sparse updates of residual channels. During the policy generation process, each... Its corresponding state It is a one-to-one correspondence, realizing the structured output of fine-tuning decision-making actions.

[0084] For example, policy network It is constructed using a Transformer architecture based on position embedding and modality embedding, specifically including an input embedding layer (which embeds...). Mapped to a 64-dimensional vector), two Transformer encoder blocks (each containing a 4-head attention mechanism), and one output decoding layer (for each...). Output The triplet structure has good parallelism and scalability, and can run independently on the main server or regional edge device.

[0085] Output action set Each action has clear location, intensity adjustment, and strategy selection information, which can provide precise control instructions for fine-tuning the model parameters in the next step.

[0086] Step 114: Update the parameters of the model substructure corresponding to the signal time-series correlation points in the multimodal large model according to the fine-tuning action.

[0087] Understandably, unlike conventional large-scale model updates, the power scenario exhibits two typical characteristics: first, data sources are highly dispersed but the logical coupling between modes is tight; second, most of the time the model does not require an overall update, only fine-tuning of local structures. Therefore, this step, in its design, not only satisfies the controllability and directionality of model updates, but also needs to consider the actual deployment of multimodal power signals and possess high-granularity scheduling capabilities for the model structure.

[0088] The input is the set of actions output from the previous policy decision step. Each action consists of three parts: Indicates whether or not modality is affected. Time point Update; Indicates the update strength; This indicates the fine-tuning strategy employed. All actions are associated with structural positions in the state vector. That is, each fine-tuning operation is bound to a modality-time pair. The system uses a structure mapping table to... Mapping to a large model Specific sub-modules, such as frequency modes ,time The weight subset of the third channel in the fourth layer of the Transformer is mapped to the cross-modal attention module. This mapping table is automatically generated based on the architecture structure during model initialization and is registered in the edge deployment nodes.

[0089] After fine-tuning, register all updated parameter structures in the update record module, and simultaneously record each Corresponding update result error improvement Write the policy feedback cache pool for policy optimization in the next cycle. This is typically calculated as the difference in prediction error before and after fine-tuning of the sliding window, for example, the normalized decrease in MAE, with a typical value in... It fluctuates between these values.

[0090] The final output is the new model structure after fine-tuning. Its architecture remains consistent with the original model, only inserting adapters, modifying residual channels, or activating parameters in frozen regions within specific structural paths. The model as a whole can still be used for inference deployment without changing the interface and functionality, only improving its adaptability and accuracy under the current power system conditions. Through the above mechanism, the system achieves dynamic, local, and controllable fine-tuning of model capabilities while ensuring model structural stability and deployment compatibility, possessing high efficiency, high stability, and high security for practical engineering applications.

[0091] For example, the finely tuned large model can be used for power stability analysis, load forecasting, and real-time monitoring of power grid operation status.

[0092] In the aforementioned large-scale power system model fine-tuning method, multimodal time-series data of the power system is acquired. This data includes voltage, current, and frequency signals. The multimodal time-series data is input into a pre-trained multimodal large-scale model to obtain the predicted sequences of each signal output by the model in the time dimension. Based on the volatility index of the predicted sequences, the importance of each signal at each time point is scored, yielding a scoring result. Based on the scoring results, time-series correlation points with importance scores exceeding a preset threshold are selected to form a candidate set. For each time-series correlation point in the candidate set, the deviation between the predicted value and the historical steady-state mean is calculated, and the deviation correlation features between different signals at the same time point are fused to construct a structured state vector. The structured state vector is input into a reinforcement learning policy network to generate corresponding fine-tuning actions. These actions include whether to perform fine-tuning, the fine-tuning intensity, and the fine-tuning strategy. Based on the fine-tuning actions, the parameters of the model substructure corresponding to the time-series correlation points in the multimodal large-scale model are updated. By utilizing the prediction sequences of the large model itself as the criterion, modes and time segments that significantly impact model performance are dynamically identified, and a candidate set is constructed. This transforms the fine-tuning process from a full-scale mode to a locally focused mode driven by model behavior, avoiding the computational burden of full model updates and improving fine-tuning efficiency. Deviations are determined by combining model output with historical steady-state mean values, and the correlation features between different signals at the same time point are introduced to construct a structured state vector that simultaneously reflects single-mode deviations and system-level coordinated anomalies, enhancing the large model's sensitivity to changes in the overall operating state of the power grid. A reinforcement learning policy network is introduced to generate an action set containing execution decisions, fine-tuning intensity, and fine-tuning strategies based on the structured state vector. This allows model updates to respond differently based on the type and degree of anomalies perceived in real time, improving the targeting and intelligence of the tuning. By updating the parameters of the model substructures corresponding to the selected signal time-series correlation points, local prediction deviations can be corrected while maximizing the protection of the model's learned overall knowledge structure and generalization ability, ensuring stable and reliable performance during long-term deployment. This method can significantly improve the response efficiency and reliability of multimodal large models in complex power scenarios.

[0093] In an exemplary embodiment, step 106 includes: calculating the standard deviation of each signal in the time dimension based on the predicted sequence; the standard deviation is used to characterize the volatility index of the predicted sequence; the standard deviation is normalized based on the normalization upper and lower bounds to obtain an importance score; the normalization upper and lower bounds are statistically derived from historical operating data based on a preset time period.

[0094] For example, regarding the aforementioned prediction output Calculate the standard deviation of the prediction output for the current window of this modality. ,get:

[0095] ;

[0096] in, It is a model In modality ,time The predicted output values ​​have been normalized. This represents the predictive volatility index of the model at this time point in this mode. This volatility index reflects the degree of uncertainty in the model's predictions in this region, indirectly indicating that the model may fail at this point.

[0097] To facilitate lateral comparisons between different modalities, Normalization is performed:

[0098] ;

[0099] Normalized upper and lower bounds and The data is derived from offline statistics of the running window data over the past 24 hours and is updated daily. That is, mode In time The relative predictive instability score is a real number between 0 and 1, with a larger value indicating greater fluctuation in the model's predictions at that location.

[0100] In one exemplary embodiment, the deviation correlation feature is obtained by calculating the standard deviation of the deviations of all signals at the same time point.

[0101] This involves calculating the standard deviation of each modal deviation at the same time point. This is used to measure the consistency of cross-modal deviations at that point in time (i.e., deviation correlation characteristics):

[0102] ;

[0103] in, It is the total number of modes (e.g., 3 modes). It is time The average of all modal deviations. This term measures the dispersion of the deviation distribution. In power system scenarios, if the deviations are highly concentrated at a certain point in time, it usually indicates that the prediction error occurs consistently across multiple modes, suggesting the existence of system-level anomalies, such as voltage drops or line tripping; conversely, if the deviations are discrete and inconsistent, it often indicates local anomalies or measurement errors.

[0104] In an exemplary embodiment, the structured state vector includes state vector terms for each signal at each time point; the method further includes: when constructing the structured state vector, if the deviation correlation characteristics between different signals at the target time point are higher than a preset value, then the values ​​of the state vector terms for all signals at the target time point are increased.

[0105] The preset value is a threshold parameter set in advance. When constructing the structured state vector, if the consistency of deviations (i.e., deviation correlation features) of different signals at the same time point is higher than the preset threshold, the value of the state vector terms of all signals at that time point is increased to enhance the priority of the reinforcement learning policy network in responding to system-level anomalies.

[0106] In an exemplary embodiment, step 112 includes: calculating the prediction accuracy improvement rate after fine-tuning based on historical fine-tuning records; and, if the prediction accuracy improvement rate is lower than a preset effective threshold, applying a penalty weight to the action selection probability corresponding to the signal temporal correlation point in the current decision of the reinforcement learning policy network.

[0107] Specifically, when generating the fine-tuning action, a historical fine-tuning efficiency penalty mechanism is introduced; for signal time point combinations whose prediction error improvement after historical fine-tuning is lower than a preset standard, i.e., signal time-series correlation points whose prediction accuracy improvement rate is lower than a preset effective threshold, the probability of their selection is reduced when generating the fine-tuning action.

[0108] For example, to improve the long-term effectiveness and decision quality of the strategy, a reward estimation function with historical parameter tuning efficiency is introduced, and the following policy loss function is constructed:

[0109] ;

[0110] In this formula, Indicating in the Execute fine-tuned model The degree of improvement in prediction error at this position, i.e. the improvement rate of prediction accuracy, is specifically obtained by normalizing the difference in error before and after fine-tuning, and is the immediate reward signal in reinforcement learning. This indicates that the mode-time combination occurred in the past. The average return inverse indicator in the second fine-tuning is:

[0111] ;

[0112] in, For the first The degree of error improvement after fine-tuning this region. This section introduces a historical inefficiency penalty mechanism, which penalizes regions that have been fine-tuned multiple times historically but with little effect. By incorporating penalties into policy training, the probability of future policies selecting a particular policy is automatically reduced, thus avoiding resource waste. This regularization term is designed for structurally non-optimizable segments in power data that may exist for a long time (such as uncorrectable errors caused by sensor drift).

[0113] In this embodiment, the strategy behavior is constrained by a historical fine-tuning efficiency penalty term, enabling it to achieve long-term profit optimization.

[0114] In an exemplary embodiment, the fine-tuning strategy includes low-rank adaptive incremental update or residual channel sparse update. Step 114 includes: if the fine-tuning strategy is low-rank adaptive incremental update, then insert a low-rank adapter on the encoding path corresponding to the signal temporal correlation point and perform incremental training; if the fine-tuning strategy is residual channel sparse update, then freeze all parameters of the multimodal large model except for the target substructure, and update the target substructure using gradient descent with sparse regularization; the target substructure is the model substructure corresponding to the signal temporal correlation point in the multimodal large model.

[0115] Understandably, this step, in its structural logic, revolves around the fundamental policy modeling principle of determining actions based on states. For example, unlike policy generation in traditional general tasks, this step considers the dynamic multimodal structure, temporal stability, and local risk shifts of the power system, and designs a reinforcement learning fine-tuning policy controller with structural indexing, historical performance penalties, and modal dependency modeling capabilities.

[0116] To ensure the policy network The output actions have a clearly defined structure and control capability; each output action... Designed as a triple containing three items:

[0117] ;

[0118] in, Indicate whether or not Perform fine-tuning; This indicates the fine-tuning intensity; a higher value indicates a larger update step size. This indicates the choice of fine-tuning method; for example, 1 represents applying low-rank adaptive incremental updates (such as LoRA structure injection), and 2 represents sparse updates of residual channels. During the policy generation process, each... Its corresponding state It is a one-to-one correspondence, realizing the structured output of fine-tuning decision-making actions.

[0119] Taking current mode as an example Its original observation data Sampling points from the downstream of the distribution transformer are connected to the SCADA system after being transformed by the low-voltage side CT sensor (Current Transformer). The sampling message period is 100ms according to the IEC61850 standard, and the data is synchronized to the server cache. Corresponding predictive output. From the model The modal prediction head uses a fully connected layer after residual connections to achieve value output. The system will output the current time. The state is determined by the policy network to require fine-tuning, and the action is... This means that the current point needs to be fine-tuned with a step size of 0.02, using strategy 1, which is to perform incremental updates based on the low-rank adapter.

[0120] In practice, a low-rank adapter is inserted into the model structure. The initial value is all zeros, and the adapter is attached to the first... The encoding path for each modality is updated in parallel with the main model structure. Model task loss function. During the training phase, the time series prediction error, such as the mean absolute error (MAE), is used. During the deployment phase, it can be calculated based on the target prediction term. The parameter update method is as follows:

[0121] ;

[0122] in, It is modal The low-rank parameter matrix represents a structure as The weight approximation term ( By controlling the update dimension through two layers of linear mapping, the number of parameters and update latency are significantly reduced. The output from the policy determines the magnitude of the current fine-tuning, and its value is determined by the degree of state deviation. Obtained by mapping, can be set to The linear compression mapping value.

[0123] If the strategy output action is The system does not use the adapter method, but instead freezes the main model except for the current position. All parameters outside the mapped substructure are only allowed to be updated within that structure, and the updates are performed using gradient descent with sparse regularization.

[0124] ;

[0125] in, This is a subset of the residual weights under the current active path; To maintain a fixed sparsity coefficient, gradient updates on non-critical paths are suppressed, making the update behavior more concentrated. This method is suitable for scenarios with unstable abnormal distributions or small long-term deviations that do not converge. It is commonly used in frequency prediction problems when the grid load is at the peak-valley critical point and there are many intermittent disturbances with small amplitudes.

[0126] In this embodiment, a dual-path fine-tuning execution mechanism is designed at the model execution level, including two modes: low-rank adaptive incremental update and residual channel sparse update, which are adapted to the model adjustment requirements under different abnormal characteristics.

[0127] 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.

[0128] Based on the same inventive concept, this application also provides a large-scale power system fine-tuning device for implementing the large-scale power system fine-tuning method 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 embodiments of the large-scale power system fine-tuning device provided below can be found in the limitations of the large-scale power system fine-tuning method described above, and will not be repeated here.

[0129] In one exemplary embodiment, such as Figure 2 As shown, a large-scale model fine-tuning device for a power system is provided, comprising:

[0130] The acquisition module 202 is used to acquire multi-modal time-series data of the power system; the multi-modal time-series data includes voltage signals, current signals and frequency signals;

[0131] The prediction module 204 is used to input multimodal time series data into a pre-trained multimodal large model to obtain the predicted sequences of each signal output by the multimodal large model in the time dimension;

[0132] The scoring module 206 is used to score the importance of each signal at each time point based on the volatility index of the predicted sequence, and obtain the scoring result.

[0133] The filtering module 208 is used to filter out signal time-series correlation points whose importance scores exceed a preset threshold based on the scoring results, forming a candidate set;

[0134] The fine-tuning module 210 is used to calculate the deviation between the predicted value and the historical steady-state mean of each signal time-series correlation point in the candidate set, and to fuse the deviation correlation features between different signals at the same time point to construct a structured state vector; the structured state vector is input into the reinforcement learning policy network to generate the corresponding fine-tuning action; wherein, the fine-tuning action includes whether to perform fine-tuning, the fine-tuning intensity, and the fine-tuning policy; and the parameters of the model substructure corresponding to the signal time-series correlation points in the multimodal large model are updated according to the fine-tuning action.

[0135] In the aforementioned large-scale power system model fine-tuning device, the prediction sequence of the large model itself is used as the discrimination criterion to dynamically identify modes and time segments that significantly affect model performance, constructing a candidate set. This transforms the fine-tuning process from a full-scale mode to a model behavior-driven, locally focused mode, avoiding the computational burden of full model updates and improving fine-tuning efficiency. The deviation is determined by combining the model output with the historical steady-state mean, and the deviation correlation features between different signals at the same time point are introduced to construct a structured state vector that can simultaneously reflect single-mode deviations and system-level coordinated anomalies, improving the large model's sensitivity to changes in the overall operating state of the power grid. A reinforcement learning policy network is introduced to generate an action set containing execution decisions, fine-tuning intensity, and fine-tuning strategies based on the structured state vector. This allows model updates to respond differently based on the real-time perceived anomaly type and degree, improving the targeting and intelligence of the tuning. By updating the model substructures corresponding to the selected signal time-series correlation points with targeted parameters, local prediction deviations can be corrected while maximizing the protection of the model's learned overall knowledge structure and generalization ability, ensuring stable and reliable performance during long-term deployment. Using this device can significantly improve the response efficiency and reliability of multimodal large models in complex power scenarios.

[0136] In an exemplary embodiment, the scoring module 206 is further configured to calculate the standard deviation of each signal in the time dimension based on the predicted sequence; the standard deviation is used to characterize the volatility index of the predicted sequence; the standard deviation is normalized based on the normalization upper and lower bounds to obtain the importance score; the normalization upper and lower bounds are statistically derived from historical operating data based on a preset time period.

[0137] In one exemplary embodiment, the deviation correlation feature is obtained by calculating the standard deviation of the deviations of all signals at the same time point.

[0138] In an exemplary embodiment, the structured state vector includes state vector terms for each signal at each time point; the fine-tuning module 210 is further configured to, when constructing the structured state vector, increase the value of the state vector terms of all signals at the target time point if the deviation correlation characteristics between different signals at the target time point are higher than a preset value.

[0139] In an exemplary embodiment, the fine-tuning module 210 is further configured to calculate the prediction accuracy improvement rate after fine-tuning the signal temporal correlation points based on historical fine-tuning records; and when the prediction accuracy improvement rate is lower than a preset effective threshold, to apply a penalty weight to the action selection probability corresponding to the signal temporal correlation points in the current decision of the reinforcement learning policy network.

[0140] In an exemplary embodiment, the fine-tuning strategy includes low-rank adaptive incremental update or residual channel sparse update. The fine-tuning module 210 is further configured to, if the fine-tuning strategy is low-rank adaptive incremental update, insert a low-rank adapter on the encoding path corresponding to the signal temporal correlation point and perform incremental training; if the fine-tuning strategy is residual channel sparse update, freeze all parameters of the multimodal large model except for the target substructure, and update the target substructure using gradient descent with sparse regularization; the target substructure is the model substructure corresponding to the signal temporal correlation point in the multimodal large model.

[0141] Each module in the aforementioned large-scale power system fine-tuning device 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, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0142] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, this 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 and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for fine-tuning a large model of a power system.

[0143] Those skilled in the art will understand that Figure 3 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.

[0144] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the following steps: acquiring multimodal time-series data of a power system; the multimodal time-series data includes voltage signals, current signals, and frequency signals; inputting the multimodal time-series data into a pre-trained multimodal large-scale model to obtain the predicted sequences of each signal output by the multimodal large-scale model in the time dimension; scoring the importance of each signal at each time point based on the volatility index of the predicted sequences to obtain a scoring result; filtering out signal time-series correlation points whose importance scores exceed a preset threshold according to the scoring results to form a candidate set; for each signal time-series correlation point in the candidate set, calculating the deviation between the predicted value of the signal time-series correlation point and the historical steady-state mean, and fusing the deviation correlation features between different signals at the same time point to construct a structured state vector; inputting the structured state vector into a reinforcement learning policy network to generate corresponding fine-tuning actions; wherein, the fine-tuning actions include whether to perform fine-tuning, the fine-tuning intensity, and the fine-tuning strategy; updating the parameters of the model substructure corresponding to the signal time-series correlation points in the multimodal large-scale model according to the fine-tuning actions.

[0145] In one embodiment, when the processor executes the computer program, it further performs the following steps: calculating the standard deviation of each signal in the time dimension based on the predicted sequence; the standard deviation is used to characterize the volatility index of the predicted sequence; the standard deviation is normalized based on the upper and lower bounds of normalization to obtain an importance score; the upper and lower bounds of normalization are statistically derived from historical operating data based on a preset time period.

[0146] In one embodiment, when the processor executes the computer program, it further implements the following steps: when constructing a structured state vector, if the deviation correlation characteristics between different signals at the target time point are higher than a preset value, the values ​​of the state vector terms of all signals at the target time point are increased.

[0147] In one embodiment, when the processor executes the computer program, it further performs the following steps: calculating the prediction accuracy improvement rate after fine-tuning the signal time-series correlation points based on historical fine-tuning records; and when the prediction accuracy improvement rate is lower than a preset effective threshold, applying a penalty weight to the action selection probability corresponding to the signal time-series correlation points in the current decision of the reinforcement learning policy network.

[0148] In one embodiment, when the processor executes the computer program, it further implements the following steps: if the fine-tuning strategy is low-rank adaptation incremental update, then a low-rank adapter is inserted into the encoding path corresponding to the signal temporal correlation point and incremental training is performed; if the fine-tuning strategy is residual channel sparse update, then all parameters of the multimodal large model except for the target substructure are frozen, and the target substructure is updated using gradient descent with sparse regularization; the target substructure is the model substructure corresponding to the signal temporal correlation point in the multimodal large model.

[0149] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it performs the following steps: acquiring multimodal time-series data of a power system; the multimodal time-series data includes voltage signals, current signals, and frequency signals; inputting the multimodal time-series data into a pre-trained multimodal large-scale model to obtain the predicted sequences of each signal output by the multimodal large-scale model in the time dimension; scoring the importance of each signal at each time point based on the volatility index of the predicted sequences to obtain a scoring result; filtering out signal time-series correlation points whose importance scores exceed a preset threshold according to the scoring results to form a candidate set; for each signal time-series correlation point in the candidate set, calculating the deviation between the predicted value of the signal time-series correlation point and the historical steady-state mean, and fusing the deviation correlation features between different signals at the same time point to construct a structured state vector; inputting the structured state vector into a reinforcement learning policy network to generate corresponding fine-tuning actions; wherein, the fine-tuning actions include whether to perform fine-tuning, the fine-tuning intensity, and the fine-tuning strategy; updating the parameters of the model substructure corresponding to the signal time-series correlation points in the multimodal large-scale model according to the fine-tuning actions.

[0150] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: calculating the standard deviation of each signal in the time dimension based on the predicted sequence; the standard deviation is used to characterize the volatility index of the predicted sequence; the standard deviation is normalized based on the upper and lower bounds of normalization to obtain an importance score; the upper and lower bounds of normalization are statistically derived from historical operating data based on a preset time period.

[0151] In one embodiment, when the computer program is executed by the processor, it further implements the following steps: when constructing a structured state vector, if the deviation correlation characteristics between different signals at the target time point are higher than a preset value, the values ​​of the state vector terms of all signals at the target time point are increased.

[0152] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: calculating the prediction accuracy improvement rate after fine-tuning the signal temporal correlation points based on historical fine-tuning records; and when the prediction accuracy improvement rate is lower than a preset effective threshold, applying a penalty weight to the action selection probability corresponding to the signal temporal correlation points in the current decision of the reinforcement learning policy network.

[0153] In one embodiment, when the computer program is executed by the processor, it further implements the following steps: if the fine-tuning strategy is low-rank adaptation incremental update, then a low-rank adapter is inserted into the encoding path corresponding to the signal temporal correlation point and incremental training is performed; if the fine-tuning strategy is residual channel sparse update, then all parameters of the multimodal large model except for the target substructure are frozen, and the target substructure is updated using gradient descent with sparse regularization; the target substructure is the model substructure corresponding to the signal temporal correlation point in the multimodal large model.

[0154] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps: acquiring multimodal time-series data of a power system; the multimodal time-series data includes voltage signals, current signals, and frequency signals; inputting the multimodal time-series data into a pre-trained multimodal large-scale model to obtain the predicted sequences of each signal output by the multimodal large-scale model in the time dimension; scoring the importance of each signal at each time point based on the volatility index of the predicted sequences to obtain a scoring result; filtering out signal time-series correlation points whose importance scores exceed a preset threshold based on the scoring results to form a candidate set; for each signal time-series correlation point in the candidate set, calculating the deviation between the predicted value of the signal time-series correlation point and the historical steady-state mean, and fusing the deviation correlation features between different signals at the same time point to construct a structured state vector; inputting the structured state vector into a reinforcement learning policy network to generate corresponding fine-tuning actions; wherein, the fine-tuning actions include whether to perform fine-tuning, the fine-tuning intensity, and the fine-tuning strategy; updating the parameters of the model substructure corresponding to the signal time-series correlation points in the multimodal large-scale model according to the fine-tuning actions.

[0155] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: calculating the standard deviation of each signal in the time dimension based on the predicted sequence; the standard deviation is used to characterize the volatility index of the predicted sequence; the standard deviation is normalized based on the upper and lower bounds of normalization to obtain an importance score; the upper and lower bounds of normalization are statistically derived from historical operating data based on a preset time period.

[0156] In one embodiment, when the computer program is executed by the processor, it further implements the following steps: when constructing a structured state vector, if the deviation correlation characteristics between different signals at the target time point are higher than a preset value, the values ​​of the state vector terms of all signals at the target time point are increased.

[0157] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: calculating the prediction accuracy improvement rate after fine-tuning the signal temporal correlation points based on historical fine-tuning records; and when the prediction accuracy improvement rate is lower than a preset effective threshold, applying a penalty weight to the action selection probability corresponding to the signal temporal correlation points in the current decision of the reinforcement learning policy network.

[0158] In one embodiment, when the computer program is executed by the processor, it further implements the following steps: if the fine-tuning strategy is low-rank adaptation incremental update, then a low-rank adapter is inserted into the encoding path corresponding to the signal temporal correlation point and incremental training is performed; if the fine-tuning strategy is residual channel sparse update, then all parameters of the multimodal large model except for the target substructure are frozen, and the target substructure is updated using gradient descent with sparse regularization; the target substructure is the model substructure corresponding to the signal temporal correlation point in the multimodal large model.

[0159] 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.

[0160] 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.

[0161] 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.

[0162] 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 method for fine-tuning a large model of a power system, characterized in that, The method includes: Acquire multimodal time-series data of a power system; the multimodal time-series data includes voltage signals, current signals, and frequency signals; The multimodal time series data is input into a pre-trained multimodal large model to obtain the predicted sequences of each signal output by the multimodal large model in the time dimension. Based on the volatility index of the predicted sequence, the importance of each signal at each time point is scored to obtain the scoring results; Based on the scoring results, signal time-series correlation points with importance scores exceeding a preset threshold are selected to form a candidate set; For each signal time-series correlation point in the candidate set, the deviation between the predicted value of the signal time-series correlation point and the historical steady-state mean is calculated, and the deviation correlation features between different signals at the same time point are fused to construct a structured state vector; The structured state vector is input into the reinforcement learning policy network to generate corresponding fine-tuning actions; wherein, the fine-tuning actions include whether to perform fine-tuning, the fine-tuning intensity, and the fine-tuning policy; The parameters of the model substructure corresponding to the signal timing correlation point in the multimodal large model are updated according to the fine-tuning action.

2. The method according to claim 1, characterized in that, The volatility index based on the predicted sequence scores the importance of each signal at each time point, yielding a scoring result, including: Based on the predicted sequence, the standard deviation of each signal in the time dimension is calculated; the standard deviation is used to characterize the volatility index of the predicted sequence. The standard deviation is normalized based on the normalized upper and lower bounds to obtain the importance score; the normalized upper and lower bounds are statistically derived from historical operating data over a preset time period.

3. The method according to claim 1, characterized in that, The deviation correlation feature is obtained by calculating the standard deviation of the deviations of all signals at the same time point.

4. The method according to claim 3, characterized in that, The structured state vector includes state vector terms for each signal at each time point; the method further includes: When constructing the structured state vector, if the deviation correlation characteristics between different signals at the target time point are higher than a preset value, the value of the state vector term of all signals at the target time point is increased.

5. The method according to any one of claims 1 to 4, characterized in that, The step of inputting the structured state vector into the reinforcement learning policy network to generate corresponding fine-tuning actions includes: The prediction accuracy improvement rate after fine-tuning is calculated based on the historical fine-tuning records of the signal time-series correlation points; If the improvement rate of prediction accuracy is lower than a preset effective threshold, a penalty weight is applied to the action selection probability corresponding to the temporal correlation point of the signal in the current decision of the reinforcement learning policy network.

6. The method according to any one of claims 1 to 4, characterized in that, The fine-tuning strategy includes low-rank adaptive incremental update or residual channel sparse update. The step of updating the parameters of the model substructure corresponding to the signal time-series correlation points in the multimodal large model according to the fine-tuning action includes: If the fine-tuning strategy is low-rank adaptation incremental update, then a low-rank adapter is inserted into the encoding path corresponding to the signal timing correlation point and incremental training is performed. If the fine-tuning strategy is residual channel sparse update, then all parameters of the multimodal large model except for the target substructure are frozen, and the target substructure is updated using gradient descent with sparse regularization terms; the target substructure is the model substructure corresponding to the signal time-series correlation point in the multimodal large model.

7. A large-scale model fine-tuning device for a power system, characterized in that, The device includes: The acquisition module is used to acquire multimodal time-series data of the power system; the multimodal time-series data includes voltage signals, current signals, and frequency signals; The prediction module is used to input the multimodal time series data into a pre-trained multimodal large model to obtain the predicted sequences of each signal output by the multimodal large model in the time dimension. The scoring module is used to score the importance of each signal at each time point based on the volatility index of the predicted sequence, and obtain the scoring result. The filtering module is used to filter out signal time-series correlation points whose importance scores exceed a preset threshold based on the scoring results, forming a candidate set; The fine-tuning module is used to calculate the deviation between the predicted value and the historical steady-state mean of each signal time-series correlation point in the candidate set, and to fuse the deviation correlation features between different signals at the same time point to construct a structured state vector; the structured state vector is input into the reinforcement learning policy network to generate a corresponding fine-tuning action; wherein, the fine-tuning action includes whether to perform fine-tuning, the fine-tuning intensity, and the fine-tuning policy; and the parameters of the model substructure corresponding to the signal time-series correlation point in the multimodal large model are updated according to the fine-tuning action.

8. 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 6.

9. 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 6.

10. A computer program product, comprising a computer program, 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 6.

Images