Data perception adaptive backup task chain construction method oriented to electric power information system

Abstract

The invention provides a data perception adaptive backup task chain construction method for an electric power information system, and relates to the field of electric power information data processing. The method comprises three steps of data collection and analysis, resource evaluation and strategy generation, and task execution and monitoring. In the data collection and analysis step, various data of an electric power information system are preprocessed, and a high-precision service load prediction model is constructed; in the resource evaluation and strategy generation step, a multi-dimensional resource consumption model is established in combination with power information system resource characteristics, and an optimal backup strategy adapting to power business requirements is dynamically generated by adopting reinforcement learning and an intelligent optimization algorithm; the task execution and monitoring step depends on the infrastructure of the power cloud platform, the execution state of a backup task is monitored, and when an abnormality is detected, a hierarchical adjustment mechanism is automatically triggered according to the priority of the power business. According to the method, through a data perception self-adaptive mechanism, the capacity of the power information system for coping with a complex data environment is remarkably enhanced, and efficient operation of power business is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of power information data processing, and in particular to a method for constructing a data-aware adaptive backup task chain for power information systems. Background Technology

[0002] In power information systems, data integrity and availability are paramount. With the deepening of smart grid construction, power information system data is becoming increasingly massive and diverse in type. Traditional data backup methods have revealed numerous insurmountable problems when dealing with these changes.

[0003] On the one hand, traditional data backup methods often adopt a fixed time interval and uniform strategy. This mode is usually executed based on a pre-set time interval and backup method. However, during peak electricity consumption periods, continuing to start backup tasks according to the fixed backup strategy will consume a lot of computing resources, network bandwidth and storage I / O, resulting in transaction delays or poor data transmission for device monitoring. During idle periods, the fixed backup strategy cannot be flexibly adjusted, resulting in resource waste and difficulty in adapting to the dynamic characteristics of power business.

[0004] On the other hand, the workload of power information systems has significant time-period fluctuations. During peak electricity consumption periods, performing full backups according to a fixed strategy will consume a large amount of computing resources and network bandwidth, resulting in monitoring data delays or transaction response timeouts. During off-peak periods, the fixed backup strategy cannot make full use of idle resources, resulting in a waste of storage and computing resources. Traditional backup methods lack the ability to finely perceive the characteristics of power data and are difficult to adapt to the dynamic characteristics of power business.

[0005] Considering the aforementioned limitations, current technologies only optimize single aspects such as data compression and incremental backup, but still lack a comprehensive consideration of the dynamic changes in business load within the power information system, fluctuations in cloud platform resources, and the diversity of data characteristics. They fail to construct a backup task chain from the overall perspective of the power information system. Therefore, it is necessary to design a backup task chain construction method based on data-aware adaptation. This method involves conducting a comprehensive evaluation and analysis of cloud platform resources using a large-scale model, and combining this with backup task requirements to design a highly adaptive backup strategy, thereby ensuring the efficient operation of power services. Summary of the Invention

[0006] In view of the shortcomings of the above-mentioned technologies, the purpose of this invention is to solve the problems of low efficiency and resource waste of traditional backup methods in scenarios with fluctuating power business load, dynamic changes in cloud platform resources, and diverse characteristics of power data.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: a data-aware adaptive backup task chain construction method for power information systems, comprising data collection and analysis steps, resource assessment and strategy generation steps, and task execution and monitoring steps. The data collection and analysis step preprocesses various types of operational data from the power information system to construct a load forecasting model that conforms to the patterns of power business. The resource assessment and strategy generation step establishes a multi-dimensional resource consumption model based on the resource characteristics of the power cloud platform and searches for the optimal backup strategy adapted to the peak-valley characteristics of power business. The task execution and monitoring step initiates backup tasks based on the power cloud platform, uses a large model to monitor the task execution status, and triggers a tiered adjustment mechanism when anomalies occur to reassess and adjust the strategy to adapt to changes in the operating status of the power system.

[0008] Optionally, the data collection and analysis steps specifically include:

[0009] Collect data from power information system business load, cloud platform resources, backup task history, and power information attributes.

[0010] The business load data includes the operating load values ​​and fluctuations of power equipment at different times;

[0011] Cloud platform resource data includes the throughput of the power cloud platform network bandwidth, the remaining capacity of storage devices, and the CPU utilization and memory usage of computing nodes;

[0012] The backup task historical data includes the start time, completion time, backup data volume, and backup method used for historical backup operations of the power information system.

[0013] Power information attribute data includes power orchestration data, the sensitivity level of user electricity consumption data, update frequency, and the business module to which it belongs.

[0014] Furthermore, the collected power information data is cleaned using a pre-trained language model based on the Transformer architecture.

[0015] During the cleaning process, for unstructured text data such as power equipment operation and maintenance records, natural language processing technology is used to remove redundant expressions, erroneous characters, and non-standard formatting.

[0016] For structured data such as power load curves, large models combined with statistical methods are used to identify and remove outliers;

[0017] For business load data sequences By calculating the mean and standard deviation The deviation from the mean will exceed Data points are considered outliers and are removed.

[0018] in, The length of the data sequence. The threshold coefficient is set according to the needs of the power business, and is generally taken as 2 or 3.

[0019] Furthermore, feature extraction is performed on the cleaned data.

[0020] For time-series business load data, a Long Short-Term Memory (LSTM) network is used to extract features along the time dimension, such as peak and off-peak electricity consumption. The LSTM hidden layer state update formula is as follows:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] in, , , These are the activation values ​​for the input gate, forget gate, and output gate, respectively. Candidate memory units, For memory units, This is the hidden layer state. It is the sigmoid activation function. For element-wise multiplication, This is the weight matrix. It is the bias vector;

[0027] For structured data of cloud platform resource attributes, a convolutional neural network (CNN) is used to extract the correlation features between different resource indicators of the power cloud platform by sliding the convolution kernel on the data matrix.

[0028] Optionally, the resource assessment and strategy generation steps specifically include:

[0029] The Apriori algorithm is used, combined with features extracted from a large model, to mine potential correlations between power data. Association rules are obtained by calculating support and confidence. The support calculation formula is as follows:

[0030]

[0031] in, and It comes from two different datasets. Indicates association rules, For inclusion and The number of transactions, It is the total number of transactions;

[0032] The confidence level is calculated using the following formula:

[0033] Among them and Same support calculation formula.

[0034] Furthermore, based on the extracted features and mined correlations, a power business load prediction model is constructed using the random forest regression algorithm.

[0035] In random forest regression, each decision tree is trained on a subset of the original power business training dataset with replacement, and at each node, the optimal feature is selected from the feature subset for splitting.

[0036] For those Business data for each sample, each sample containing multiple influencing factors. and the corresponding business load value ,

[0037] For each decision tree During training, the tree structure is constructed based on selected features and splitting criteria. During prediction, a new set of influencing factors is given. Each decision tree will give a predicted value. The final predicted value is the average of all the predictions from the decision trees.

[0038] Furthermore, the network, storage, and computing resources of the power cloud platform are evaluated using deep neural networks (DNNs).

[0039] Construct an input feature vector based on the preprocessed power data. These feature vectors are input into a DNN, and after calculations through multiple hidden layers, the performance prediction of the resource at a certain future time is finally obtained. ,in For the output layer function,

[0040] For computing resources, assess the remaining available capacity and performance bottlenecks to ensure the efficient operation of the power orchestration algorithm;

[0041] For storage resources, assess storage capacity usage and data read / write speed to ensure reliable storage of historical power data;

[0042] For network resources, assess network bandwidth utilization and network latency to meet the transmission requirements of power monitoring data;

[0043] Furthermore, the types, data volumes, and data importance of backup tasks in the power information system are analyzed.

[0044] Regarding the data volume, based on the historical data volume records generated by different power information system modules and combined with the development trend of power business, the ARIMA model is used to predict the data volume of future backup tasks. The ARIMA model expression is:

[0045]

[0046] in, It is a shift operator. and These are model parameters. It is the difference order. It is the original time series. It is a white noise sequence;

[0047] For data importance levels, the importance weights of the data are determined using the Analytic Hierarchy Process (AHP) based on the criticality of the power business module to which the data belongs and the degree of impact of data loss on the power business.

[0048] Next, a judgment matrix is ​​constructed based on the factors influencing the importance of the data. By calculating the eigenvectors and eigenvalues ​​of the matrix, the weights of each factor are obtained, thereby determining the importance level of the power data.

[0049] Furthermore, based on the resource assessment results of the power cloud platform and the backup task requirements, an adaptive backup strategy is designed.

[0050] A reinforcement learning framework is adopted, with the resource status of the power cloud platform as the environment state. The backup policy operation is treated as an action. A reward function is constructed based on resource usage costs, backup task completion time, and data integrity. The agent continuously performs actions in the environment, receives feedback based on the reward function, and iteratively optimizes the backup strategy using the Q-learning algorithm. The algorithm expression is:

[0051]

[0052] in, It's the learning rate. It is a discount factor.

[0053] The Particle Swarm Optimization (PSO) algorithm is used to dynamically adjust the backup strategy parameters, which constitute the position vectors of the particles. The velocity vector of each particle is ,

[0054] In each iteration, the particle is determined based on its optimal position. and the global optimal position Update speed and position, the update formula is:

[0055]

[0056]

[0057] in, It is inertial weight. and It is a learning factor. and yes Random numbers between It's time.

[0058] In the resource assessment and analysis of the power cloud platform, in addition to using numerical historical data, images of the operating status of power equipment and audio recordings of equipment anomalies are also integrated.

[0059] Using the MT multimodal fusion model, image features are extracted from image data through a convolutional neural network. ,

[0060] For audio data, audio features are extracted using Mel-frequency cepstral coefficients. At the same time, combined with the characteristics of numerical power data Fusion is performed using a cross-modal attention mechanism:

[0061]

[0062] The fused features are input into a deep neural network for resource assessment, improving the accuracy and comprehensiveness of resource assessment on the power cloud platform.

[0063] Furthermore, in the backup task requirements analysis, the weight of the importance level of power data is dynamically adjusted according to the power business status.

[0064] An attention mechanism is introduced, and an attention weight matrix is ​​constructed. Based on the key indicators of the current power business, the attention weights of the factors influencing the importance of each data point are calculated. ,Right now:

[0065]

[0066] in, The scores of each factor are calculated using an attention function, and the importance level of the data is reassessed based on dynamically adjusted weights.

[0067] Furthermore, in the design of adaptive backup strategies, when the business scenarios of the power information system or the cloud platform architecture undergo significant changes, transfer learning is used to optimize the reinforcement learning framework.

[0068] The model parameters pre-trained in similar power business scenarios or related tasks will be transferred to the current backup strategy design task as initial parameters. In the new power business environment, they will be fine-tuned and trained with a small amount of sample data to quickly adapt to the new scenario.

[0069] When dynamically adjusting backup strategy parameters using particle swarm optimization, hybrid optimization is performed by combining it with simulated annealing (SA).

[0070] During the iteration process of the PSO algorithm, the SA algorithm is introduced every certain number of iterations.

[0071] In the SA algorithm, based on the current solution Generate new solutions Calculate the change in the objective function value ,

[0072] if Then accept the new solution.

[0073] if Then, based on probability Receive new solution,

[0074] in, It is a temperature parameter, which gradually decreases with iteration.

[0075] Optionally, in the task execution and monitoring step, based on the aforementioned generated optimal backup strategy, task execution is triggered via the power cloud platform API interface.

[0076] Based on the time parameter in the strategy By dynamically generating task triggering rules using cron expressions, the impact on power services can be avoided during peak electricity consumption periods.

[0077] According to the resource ratio in the strategy Allocate computing resources and network bandwidth according to the mode parameters in the policy. The formula for calculating the allocated resources and network bandwidth when calling the corresponding backup engine module is as follows:

[0078]

[0079]

[0080] in, Total number of CPU cores Total network bandwidth

[0081] Furthermore, a monitoring and analysis engine is built using the Transformer large model to collect and analyze three types of key indicators, including:

[0082] The first category is the task execution status, using an LSTM model trained on historical data, with the input being the elapsed time. and remaining data volume Output the predicted remaining time :

[0083]

[0084] Furthermore, based on the amount of backed-up data... and total data volume Calculate the completion rate:

[0085]

[0086] The second category is resource consumption, which calculates resource utilization rate. and memory utilization The thresholds were set to 80% and 70% respectively.

[0087] Calculate the current network bandwidth rate , Then the flow shaping mechanism is activated to ensure the normal transmission of power monitoring data;

[0088] The third category is data consistency, which involves performing hash verification on backup data blocks, comparing the hash values ​​of the source data and the backup data, and setting the inconsistency rate threshold to 0.01% to ensure the integrity of critical power data.

[0089] When the monitoring engine detects an anomaly, it triggers a three-level response mechanism:

[0090] Level 1 adjustment, backup completion rate If the time taken exceeds 1.5 times the estimated time, the current task will be automatically retried, with the number of retries limited. ;

[0091] Level 2 adjustment: If resource utilization exceeds the threshold for 10 consecutive minutes, the PSO algorithm is invoked to regenerate the policy parameters. These represent the startup time, resource allocation ratio, and backup mode of the regenerated backup, respectively.

[0092]

[0093] in, To adjust the step size sgn is the sign function;

[0094] Then the flow shaping mechanism will be activated;

[0095] Level 3 adjustment, data inconsistency rate If three consecutive retries fail, a cross-domain backup switch is triggered, migrating the task to a remote data center and using a hybrid cloud resource frequency conversion model for retransmission to ensure reliable backup of power data.

[0096] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned data-aware adaptive backup task chain construction method for power information systems.

[0097] A computer-readable storage medium storing computer instructions that, when executed by a processor, implement the aforementioned data-aware adaptive backup task chain construction method for power information systems.

[0098] Compared with the prior art, the advantages of the present invention are as follows:

[0099] By conducting a comprehensive evaluation and analysis of cloud platform resources through a large model and combining backup task requirements, a highly adaptive backup strategy is designed to realize the transformation of power information data backup from a fixed mode to an intelligent adaptive mode. This improves backup efficiency by about 30%, increases resource utilization from an average of 50% to about 75%, and shortens fault recovery time to 1 / 5 of the traditional method. It effectively solves the inherent defects of traditional backup methods and meets the intelligent data backup requirements of power information systems. Attached Figure Description

[0100] Figure 1 A flowchart illustrating the solution for constructing a data-aware adaptive backup task chain for power information systems.

[0101] Figure 2 This is a diagram of the core structure of LSTM.

[0102] Figure 3 This is a flowchart of the random forest regression algorithm.

[0103] Figure 4 This is a diagram of a deep neural network (DNN) structure. Detailed Implementation

[0104] The invention will be further described below with reference to the accompanying drawings.

[0105] Example: In Figure 1 The overall process shown first involves collecting power system data from the power information system's business load, cloud platform resources, backup task history, and power information attributes.

[0106] Specifically, power equipment load data is collected from the SCADA system to form a time series; the network bandwidth, storage capacity, and computing resources of the power cloud platform are obtained through the Prometheus monitoring tool; the start time, duration, data volume, and backup method of historical tasks are extracted from backup logs; and the sensitivity level, update frequency, and business module to which the power orchestration data and user electricity consumption information belong are marked through the metadata database.

[0107] Furthermore, the collected power grid data is cleaned using a pre-trained language model based on the Transformer architecture.

[0108] Specifically, for unstructured text-based data such as power equipment operation and maintenance records and fault reports, natural language processing techniques are used to remove redundant expressions, erroneous characters, and non-standard formatting. For structured data such as power load curves and equipment operating parameters, large-scale models combined with statistical methods are used to identify and remove outliers. For business load data sequences... By calculating the mean and standard deviation The deviation from the mean will exceed Data points are considered outliers and are removed. The length of the data sequence. The threshold coefficient is set according to business needs, and is generally 2 or 3;

[0109] Furthermore, feature extraction is performed on the cleaned data.

[0110] For time-series service load data, a 3-layer LSTM network is constructed to process the daily load time series, such as... Figure 2 As shown, the hidden layer state update formula of LSTM is:

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117] in, , , These are the activation values ​​for the input gate, forget gate, and output gate, respectively. Candidate memory units, For memory units, This is the hidden layer state. It is the sigmoid activation function. For element-wise multiplication, This is the weight matrix. It is the bias vector;

[0118] For structured data of cloud platform resource attributes, a convolutional neural network (CNN) is used to construct a data matrix of power cloud platform resource data, and spatial correlation features are extracted by sliding convolution kernels.

[0119] Furthermore, the Apriori algorithm is employed, combined with features extracted from a large model, to uncover potential correlations between power data. Association rules are obtained by calculating support and confidence. The support calculation formula is as follows:

[0120]

[0121] in, and Derived from two different sets of power data, Indicates association rules, For inclusion and The number of transactions, It is the total number of transactions; the confidence score is calculated using the following formula:

[0122]

[0123] Furthermore, based on the extracted features and the mined relationships, such as Figure 3 As shown, a business load prediction model is constructed using the random forest regression algorithm.

[0124] In random forest regression, each decision tree is trained on a subset of the original power business training dataset with replacement. At each node, the optimal feature is selected from the feature subset for splitting. Business data for each sample, each sample containing multiple influencing factors. and the corresponding business load value ,

[0125] For each decision tree During training, the tree structure is constructed based on selected features and splitting criteria. During prediction, a new set of influencing factors is given. Each decision tree will give a predicted value. The final prediction is the average of all decision tree predictions, which is used to guide backup tasks to avoid peak electricity usage periods.

[0126] like Figure 4 The diagram shows the construction of a 5-layer deep neural network (DNN), with input feature vectors. The hidden layer states are calculated using the ReLU activation function. After calculations through three hidden layers, the performance prediction of the resource at a future time is finally obtained. ,in This is the output layer function.

[0127] The obtained performance predictions can be used to assess remaining available capacity and performance bottlenecks for computing resources; to assess storage capacity utilization and data read / write speeds for storage resources; and to assess network bandwidth utilization and network latency for network resources.

[0128] Furthermore, the types of backup tasks, data volume, and data importance are analyzed.

[0129] Regarding the data volume, based on historical backup data of different power information system modules and combined with business development trends, through... The non-stationary sequence is transformed into a stationary sequence, and an ARIMA model is fitted to predict the data volume of future backup tasks. The ARIMA model expression is as follows:

[0130]

[0131] in, It is a shift operator. and These are model parameters. It is the difference order. It is the original time series. It is a white noise sequence;

[0132] For data importance levels, the importance weights of the data are determined using the Analytic Hierarchy Process (AHP) based on the criticality of the power business module to which the data belongs and the impact of data loss on the business.

[0133] Construct a data importance judgment matrix, using factors such as "the degree of impact of data loss on business", "update frequency" and "criticality of the module to which it belongs" as factors, calculate the eigenvector and eigenvalue of the matrix, obtain the weight of each factor, and then determine the importance level of the data;

[0134] Furthermore, based on the resource assessment results of the power cloud platform and the backup task requirements, an adaptive backup strategy is designed.

[0135] A reinforcement learning framework is adopted, with the cloud platform resource state as the environment state. The backup policy operation is treated as an action. A reward function is constructed based on resource usage costs, backup task completion time, and data integrity. The agent continuously performs actions in the environment, receives feedback based on the reward function, and iteratively updates the backup strategy using the Q-learning algorithm to adapt to changes in cloud platform resources and backup task requirements. The algorithm expression is:

[0136]

[0137] in, It's the learning rate. It is a discount factor.

[0138] The Particle Swarm Optimization (PSO) algorithm is used to dynamically adjust the backup strategy parameters, which constitute the position vectors of the particles. The velocity vector of each particle is In each iteration, the particle determines its optimal position. and the global optimal position Update speed and position, the update formula is:

[0139] ;

[0140]

[0141] in, It is inertial weight. and It is a learning factor. and yes Random numbers between It's time.

[0142] In the resource assessment and analysis of the power cloud platform, in addition to using numerical historical data, it also integrates infrared thermal imaging images of substation equipment, operating sounds of transformers, and abnormal sounds and audio of these devices.

[0143] Using the MT multimodal fusion model, for image data, taking device thermal imaging as an example, temperature distribution and component status are extracted as image features through a convolutional neural network. For audio data, taking transformer noise as an example, the sound frequency and amplitude are extracted as audio features using Mel-frequency cepstral coefficients. At the same time, combined with the characteristics of numerical data Fusion is performed using a cross-modal attention mechanism:

[0144]

[0145] The fused features are input into a deep neural network for resource evaluation, improving the ability to predict potential faults in power equipment.

[0146] In the backup task requirements analysis, the weight of the importance level of power data is dynamically adjusted according to the power business status.

[0147] An attention mechanism is introduced, and an attention weight matrix is ​​constructed. Based on the key indicators of the current power business, the attention weights of the factors influencing the importance of each data point are calculated. ,Right now:

[0148]

[0149] in, The scores of each factor are calculated using an attention function, and the importance level of the data is reassessed based on dynamically adjusted weights.

[0150] In adaptive backup strategy design, when the business scenarios of the power information system or the cloud platform architecture undergo significant changes, transfer learning is used to optimize the reinforcement learning framework.

[0151] The model parameters pre-trained in power business scenarios similar to smart substations and distribution automation will be transferred to the current backup strategy design task as initial parameters. In the new environment, they will be fine-tuned and trained with a small amount of sample data to quickly adapt to the new scenario.

[0152] When dynamically adjusting backup strategy parameters using the Particle Swarm Optimization (PSO) algorithm, a hybrid optimization approach combining it with Simulated Annealing (SA) is employed. During the PSO iteration process, SA is introduced after every 50 iterations. In SA, the current solution is used to adjust the backup strategy parameters. Generate new solutions Calculate the increment of the objective function value ,if Receive new solutions directly.

[0153] if Then, based on probability Accepting new solutions and avoiding getting trapped in local optima, among which, It is a temperature parameter, which gradually decreases with iteration.

[0154] Based on the aforementioned optimal backup strategy, task execution is triggered via the power cloud platform API interface, according to the time parameters in the strategy. Task triggering rules are dynamically generated using cron expressions, based on the resource ratio specified in the strategy. Allocate computing resources and network bandwidth according to the mode parameters in the policy. The formula for calculating the allocated resources and network bandwidth when starting the corresponding incremental backup engine is as follows:

[0155]

[0156]

[0157] in, Total number of CPU cores Total network bandwidth

[0158] A monitoring and analysis engine is built using the Transformer large model to collect and analyze three types of key metrics, including:

[0159] The first category is the task execution status, using an LSTM model trained on historical data, with the input being the elapsed time. and remaining data volume Output the predicted remaining time :

[0160]

[0161] Then based on the amount of backed-up data and total data volume Calculate the completion rate:

[0162]

[0163] The second category is resource consumption, which calculates resource utilization rate. and memory utilization The thresholds are set to 80% and 70% respectively, and the current network bandwidth rate is calculated. ;

[0164] The third category is data consistency, which involves performing hash verification on backup data blocks, comparing the hash values ​​of the source data and the backup data, and setting the inconsistency rate threshold to 0.01%.

[0165] When the monitoring engine detects an anomaly, it triggers a three-level response mechanism:

[0166] Level 1 adjustment, backup completion rate If the time taken exceeds 1.5 times the estimated time, the current task will be automatically retried, with the number of retries limited. ;

[0167] Level 2 adjustment: If resource utilization exceeds the threshold for 10 consecutive minutes, the PSO algorithm is invoked to regenerate the policy parameters. These represent the startup time, resource allocation ratio, and backup mode of the regenerated backup, respectively.

[0168]

[0169] in, To adjust the step size sgn is the sign function;

[0170] Then the flow shaping mechanism is activated to prioritize the transmission of relay protection data;

[0171] Level 3 adjustment, data inconsistency rate If three consecutive retries fail, a cross-domain backup switch is triggered, migrating the task to a remote power disaster recovery center and using a hybrid cloud resource orchestration model for retransmission to ensure the integrity of core power data.

[0172] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A method for constructing a data-aware adaptive backup task chain for power information systems, characterized in that, The method includes the following steps: Data collection and analysis, tailored to the characteristics of power information systems, involves collecting multi-data elements from these systems, including business load data, cloud platform resource data, backup task historical data, and power information attribute data. Large-scale models are used for data cleaning and feature extraction to deeply explore potential relationships between data and construct business load prediction models. Resource assessment and strategy generation: Based on the large model, the resources of the power information system cloud platform are assessed and analyzed. Combined with the backup task requirements, an adaptive backup strategy is designed, and the strategy parameters are dynamically adjusted through intelligent algorithms. Task execution and monitoring: Based on the generated backup strategy, the backup task of the power information system is initiated. The large model is used to monitor the entire process of task execution. Once an anomaly is detected, an emergency mechanism is triggered to optimize and adjust the backup plan in a timely manner.

2. The data-aware adaptive backup task chain construction method for power information systems according to claim 1, characterized in that, Data collection and analysis require processing diverse data from the power information system, including business load data, cloud platform resource data, backup task historical data, and power information attribute data. Business load data includes the operating load values ​​and fluctuations of power equipment at different times; cloud platform resource data includes network bandwidth throughput, remaining storage capacity, CPU utilization of computing nodes, and memory usage; backup task historical data involves the start time, completion time, backup data volume, and backup method used for historical backup operations; and power information attribute data covers the data's sensitivity level, update frequency, and the business module to which it belongs. The collected data is cleaned using a pre-trained language model based on the Transformer architecture: for unstructured text data, natural language processing techniques are used to remove redundant expressions, erroneous characters, and non-standard formatting; for structured data, statistical methods are combined to identify and remove outliers. For the business load data sequence {x1,x2,…,x n }, by calculating the mean and standard deviation Data points deviating from the mean by more than a preset threshold kσ are considered outliers and removed, where n is the length of the data sequence and k is a threshold coefficient set according to business requirements, taking the value of 2 or 3. Feature extraction is performed on the cleaned data. For time-series business load data, a Long Short-Term Memory (LSTM) network is used to extract features along the time dimension. The LSTM hidden layer state update formula is as follows: i t =σ(W xi x t +W hi h t-1 +b i ) f t =σ(W xf x t +W hf h t-1 +b f ) o t =σ(W xo x t +W ho h t-1 +b o ) h t =o t ⊙tanh(c t ) Among them, i t f t o t These are the activation values ​​for the input gate, forget gate, and output gate, respectively. As a candidate memory unit, c t h is a memory unit. t Let σ be the hidden layer state, σ be the sigmoid activation function, ⊙ be the element-wise multiplication, W be the weight matrix, and b be the bias vector. For structured data of cloud platform resource attributes, a convolutional neural network (CNN) is used to extract the correlation features between different resource indicators by sliding the convolution kernel on the data matrix.

3. The data-aware adaptive backup task chain construction method for power information systems according to claim 2, characterized in that, The Apriori algorithm is used, combined with features extracted from a large model, to uncover potential associations between data. Association rules are obtained by calculating support and confidence. The support calculation formula is as follows: Here, X and Y come from two different datasets, X→Y represents the association rule, |X∪Y| is the number of transactions containing X and Y, and |D| is the total number of transactions; the confidence score is calculated using the following formula: Based on the extracted features and the mined correlations, a business load prediction model is constructed using the random forest regression algorithm. In random forest regression, each decision tree is trained on a subset of the original training dataset with replacement. At each node split, the optimal feature is selected from the feature subset for splitting. For business data with n samples, each sample contains multiple influencing factors x1, x2, ..., x... m And the corresponding business load value y, For each decision tree T i During training, the tree structure is constructed based on selected features and splitting criteria. During prediction, a new set of influencing factors x is given. new1 ,x new2 ,…,x newm Each decision tree will give a predicted value. The final prediction is the average of all the predictions from the decision trees.

4. The data-aware adaptive backup task chain construction method for power information systems according to claim 2, characterized in that, In the resource assessment and strategy generation steps, a deep neural network (DNN) is used to assess the network, storage, and computing resources of the power information system cloud platform, and an input feature vector X is constructed based on the preprocessed data. i =[x i1 ,x i2 ,…,x in These feature vectors are input into a DNN, and after calculations through multiple hidden layers, the final performance prediction value of the resource at a certain future time is obtained as y = g(h). last ), where g is the output layer function, which evaluates the remaining available capacity and performance bottlenecks for computing resources; evaluates the storage capacity usage and data read / write speed for storage resources; and evaluates the network bandwidth utilization and network latency for network resources. The analysis considers the type of backup task, data volume, and data importance. Regarding data volume, based on historical data volume records generated by different power information system modules and combined with business development trends, the ARIMA model is used to predict the data volume of future backup tasks. The ARIMA model expression is: Where B is the shift operator, and θ i These are the model parameters, where d is the difference order and X is the model parameter. t It is the original time series, Z t It is a white noise sequence; For data importance levels, the importance weights of the data are determined by the Analytic Hierarchy Process (AHP) based on the criticality of the business module to which the data belongs and the impact of data loss on the business. The factors influencing data importance are constructed into a judgment matrix. By calculating the eigenvectors and eigenvalues ​​of the matrix, the weights of each factor are obtained, and thus the importance level of the data is determined. In response to the backup requirements of the power information system, an adaptive backup strategy is designed using a reinforcement learning framework. The cloud platform resource state is taken as the environment state S, and the backup strategy operations are taken as actions A. A reward function R is constructed based on resource usage cost, task completion time, and data integrity. The agent continuously executes actions in the environment and receives feedback based on the reward function. The backup strategy is iteratively optimized using a Q-learning algorithm. The algorithm expression is: Where α is the learning rate and γ is the discount factor. The Particle Swarm Optimization (PSO) algorithm is used to dynamically adjust the backup strategy parameters, which constitute the particle position vector X. i =[x i1 ,x i2 ,…,x in The velocity vector of each particle is V. i =[v i1 ,v i2 ,…,v in In each iteration, the particle determines its optimal position pBest. i The update formula for the global optimal position gBest is: v ij (t+1)=ωv ij (t)+c1r 1j (t)[pBest ij (t)-x ij (t)]+c2r 2j (t)[gBest j (t)-x ij (t)] x ij (t+1)=x ij (t)+v ij (t+1) Where ω is the inertia weight, c1 and c2 are learning factors, and r 1j and r 2j It is a random number between [0,1], and t is time.

5. The data-aware adaptive backup task chain construction method for power information systems according to claim 4, characterized in that, In cloud platform resource assessment and analysis, in addition to using numerical historical data, images of equipment operating status and audio recordings of equipment anomalies are also integrated; utilizing the MT multimodal fusion model, for image data, image features F are extracted through a convolutional neural network. image For audio data, audio features F are extracted using Mel-frequency cepstral coefficients. audio At the same time, combined with the numerical data feature F numerical Fusion is performed using a cross-modal attention mechanism: F fusion =CrossAttention([F image ;F audio ;F numerical ]) The fused features are then input into a deep neural network for resource evaluation. In backup task requirements analysis, the weights of data importance levels are dynamically adjusted based on business status. An attention mechanism is introduced, and an attention weight matrix is ​​constructed. The attention weight α of each data importance influencing factor is calculated based on the key indicators of the current business. i ,Right now: Among them, e i The scores of each factor are calculated using an attention function, and the importance level of the data is reassessed based on dynamically adjusted weights. In adaptive backup strategy design, when the business scenario or cloud platform architecture of the power information system undergoes significant changes, the reinforcement learning framework is optimized by using transfer learning technology. The model parameters pre-trained in similar or related tasks are transferred to the current backup strategy design task as initial parameters. In the new environment, fine-tuning training is performed using a small amount of sample data to quickly adapt to the new scenario. When dynamically adjusting backup strategy parameters using the Particle Swarm Optimization (PSO) algorithm, a hybrid optimization approach combining it with Simulated Annealing (SA) is employed. During the PSO iteration process, SA is introduced at regular intervals. In SA, the current solution x is used to adjust the backup strategy parameters. current Generate a new solution x new Calculate the change in the objective function value ΔE = f(x) new )-f(x current If ΔE < 0, then accept the new solution; if ΔE > 0, then accept the solution with probability. Receive a new solution, where T is a temperature parameter that gradually decreases with each iteration.

6. The data-aware adaptive backup task chain construction method for power information systems according to claim 5, characterized in that, In the task execution and monitoring steps, based on the generated optimal backup strategy, task execution is triggered through the power information system cloud platform API interface. According to the time parameter 't' in the strategy, task triggering rules are dynamically generated using cron expressions. Computational resources and network bandwidth are dynamically allocated according to the resource ratio 'r' in the strategy. The corresponding backup engine module is called according to the mode parameter 'm' in the strategy. The formulas for calculating the allocated resources and network bandwidth are as follows: Among them, C total B represents the total number of CPU cores. total Total network bandwidth A monitoring and analysis engine is built using the Transformer large model to collect and analyze three types of key metrics, including: The first category is the task execution status, using an LSTM model trained on historical data, with the elapsed time as input. T elapsed and remaining data volume D remaining Output the predicted remaining time T est : T est =f LSTM (T elapsed ,D remaining ) Then, based on the amount of backed-up data D processed and total data volume D total Calculate the completion rate: The second category is resource consumption, calculating resource utilization rate U. cpu and memory utilization U mem The thresholds are set to 80% and 70% respectively, and the current network bandwidth rate B is calculated. current ; The third category is data consistency, which involves performing hash verification on backup data blocks, comparing the hash values ​​of the source data and the backup data, and setting the inconsistency rate threshold to 0.01%. When the monitoring engine detects an anomaly, it triggers a three-level response mechanism: Level 1 adjustment: If the backup completion rate (CR) is less than 30% and the backup time exceeds the estimated time by more than 1.5 times, the current task is automatically retried, with a retry count of ≤3. Level 2 adjustment: If resource utilization exceeds the threshold for 10 consecutive minutes, the PSO algorithm is invoked to regenerate the policy parameters t′, r′, and m′, representing the regenerated backup start time, resource allocation ratio, and backup mode, respectively. r′=r+Δr·sgn(U cpu -80%) Δr is the adjustment step size (5%-10%), and sgn is the sign function; B current >B alloc If the value is ×1.2, the flow shaping mechanism will be activated; Level 3 adjustment: If the data inconsistency rate is >0.1% or three consecutive retries fail, a cross-domain backup switch is triggered, the task is migrated to a remote data center, and the hybrid cloud resource scheduling model is invoked for retransmission.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the data-aware adaptive backup task chain construction method for power information systems as described in any one of claims 1 to 6 above.

8. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, the computer instructions implement the data-aware adaptive backup task chain construction method for power information systems as described in any one of claims 1-6.

Images