Multi-modal power data restoration method based on local-global correlation learning

Abstract

The invention relates to a local-global association learning-based multi-modal power data restoration method, which comprises the steps of obtaining initial data and performing preprocessing to obtain a power data abnormal value, the initial data comprising power operation data, external association data and multi-modal data; based on the preprocessed power operation data and the power data abnormal value, local association learning is carried out, local data association is mined, and a local restoration scheme of the power data abnormal value is obtained; performing global association learning based on the preprocessed initial data and the abnormal value of the power data, mining a cross-time and cross-source global association rule, and obtaining a reasonable correction range of the abnormal value of the power data; and generating a power operation data restoration result in combination with the local restoration scheme and the reasonable correction range. Compared with the prior art, the method has the advantages that the data exception can be accurately recognized and repaired, and the data repair accuracy is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of power data restoration, and in particular to a multimodal power data restoration method based on local-global association learning. Background Technology

[0002] With the continuous and rapid advancement of the construction of new power systems, the rapid and accurate forecasting of electricity sales has become an important basis for the operation and decision-making of power companies. However, in the process of conducting data forecasting, quality issues such as missing, abnormal, and inconsistent data have seriously restricted the improvement of the effectiveness and scientific nature of electricity forecasting models. There is an urgent need to carry out quality repair of multi-source and multi-modal power data to improve the scientific nature and accuracy of power forecasting analysis.

[0003] Power data quality restoration involves data anomaly identification and repair processing. Data anomaly identification aims to detect problems such as missing data, noise, and outliers in the data through algorithms and models, while repair processing corrects and completes these problems to restore the integrity and accuracy of the data.

[0004] Chinese patent CN113485986B discloses a power data restoration method, which is based on the SOM neural network and Pearson correlation coefficient theory. Considering the complex nonlinearity of power data, it utilizes the strong learning ability of neural networks to handle nonlinear problems, thus achieving power data restoration and effectively improving classification efficiency and accuracy. However, this application relies only on historical power data and a few external factors for data restoration, failing to comprehensively capture the complex operating states of the power system. Furthermore, the use of SOM neural network classification and a fixed imputation strategy lacks dynamic threshold adjustment, making it difficult to handle complex and abnormal scenarios. The LSTM prediction used in this application is mainly used to predict short-term time-series features; its restoration effect decreases further when dealing with long-term data, making it difficult to handle high-dimensional, nonlinear, multi-source data correlation problems.

[0005] Existing data anomaly identification and repair technologies often employ statistical methods, such as threshold setting, deletion of missing values, mean, median or linear interpolation, moving average, and simple prediction imputation. However, when faced with increasingly complex and multi-sourced large amounts of data anomalies, these methods not only lack a high degree of scientific rigor but also easily lead to excessive data deletion, which is detrimental to efficient power forecasting modeling. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a multimodal power data repair method based on local-global correlation learning.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A multimodal power data restoration method based on local-global correlation learning, the method comprising:

[0009] Acquire initial data and perform preprocessing to obtain abnormal values ​​in power data. The initial data includes power operation data, external correlation data, and multimodal data.

[0010] Based on the preprocessed power operation data and power data anomalies, local correlation learning is performed to explore local data correlations and obtain local repair solutions for power data anomalies.

[0011] Based on the preprocessed initial data and power data outliers, global correlation learning is performed to explore global correlation patterns across time and sources, and to obtain a reasonable correction range for power data outliers.

[0012] Based on the aforementioned local repair scheme and reasonable correction range, power operation data repair results are generated.

[0013] Furthermore, the power operation data includes electricity consumption type, electricity consumption, electricity price, and load; the external correlation data includes meteorological data and regional economic indicators; and the multimodal data includes equipment fault descriptions, maintenance operation records, and equipment information.

[0014] Furthermore, the process of local association learning includes:

[0015] Based on the preprocessed power operation data and power data outliers, time series prediction results are obtained using a pre-trained LSTM+Transformer model and graph neural network.

[0016] The dynamic residuals are calculated based on the time series prediction results, and an initial threshold is set by combining sliding window kernel density estimation and extreme value theory.

[0017] The dynamic residual is set to exceed the initial threshold by a generalized Pareto distribution, and the excess is fitted to calculate the dynamic threshold.

[0018] Time-series prediction results that exceed dynamic thresholds are marked as anomalies, and multi-level anomaly labels are generated based on the degree of exceedance. Based on local spatiotemporal patterns, a local repair scheme for abnormal values ​​in power data is output.

[0019] Furthermore, the calculation expression for the dynamic residual is as follows:

[0020]

[0021] Among them, y t This refers to the actual value, i.e., the abnormal value in the power data. r represents the time series prediction result. t For dynamic residuals;

[0022] The calculation expression for the dynamic threshold is:

[0023]

[0024]

[0025] Where u is the dynamic threshold, u0 is the initial threshold, δ and β are fitting parameters, α is the significance level, P is the exceedance probability of the dynamic threshold, x is the amount by which the dynamic residual exceeds the initial threshold, and F(x; δ, β) is the generalized Pareto distribution.

[0026] Furthermore, the global association learning process includes:

[0027] Establish temporal and spatial relationships between the preprocessed power operation data, external correlation data, and multimodal data, and generate frequent itemsets through association rules;

[0028] Calculate the occurrence frequency of each frequent itemset, and filter out frequent itemsets whose occurrence frequency is greater than a preset first threshold;

[0029] Based on the frequent itemsets, generate all possible association rules, calculate the confidence of the association rules, and filter out association rules with a confidence greater than a preset second threshold.

[0030] Based on the association rules, a global time series model is established, and based on the global time series model, a reasonable correction range is generated to limit the reasonable correction range of abnormal values ​​in power data.

[0031] Furthermore, the expression for the global time series model is:

[0032] y(t)=g(t)+s(t)+h(t)+∈(t)

[0033] Where g(t) represents the trend term, i.e. the change trend of the time series on non-periodic terms, s(t) represents the periodic term, i.e. the change trend of the time series on periodic terms, h(t) represents the holiday influence trend, i.e. the potential impact of non-fixed periodic holidays in the time series on the predicted value, and ∈(t) represents the error term that follows a Gaussian distribution, i.e. the fluctuations that the model did not predict.

[0034] Furthermore, the preprocessing includes:

[0035] Wavelet transform is used to remove high-frequency noise interference from the initial data;

[0036] Abnormal values ​​in the initial data are identified using the 3σ principle and sliding window detection.

[0037] Align the time and spatial scales of the power operation data, external correlation data, and multimodal data.

[0038] Furthermore, the process of combining the aforementioned local repair scheme with a reasonable correction range includes:

[0039] The confidence level of each initial data point is defined based on the sensor accuracy used to acquire the initial data, and the severity of the anomalies in the power data outliers is calculated.

[0040] The weights of local repair schemes and reasonable correction ranges are calculated based on the confidence level of each initial data and the severity of the abnormality of the power data anomalies.

[0041] Based on the weights of local repair schemes and reasonable correction ranges, an objective function is constructed with the goal of minimizing repair error and maximizing data smoothness.

[0042] The objective function is solved using gradient descent and a genetic algorithm to obtain the optimal repair value.

[0043] Furthermore, the expression for the objective function is:

[0044]

[0045] Where α is a weighting parameter used to balance the relative importance between minimizing repair error and maximizing data smoothness, E error To correct the error, E smooth Here, n represents the data smoothing term, and n is the total number of power data points. This is the repair value for the i-th data. w represents the actual observed value of the i-th data point. l,i w represents the weight of the local repair scheme. g,i To reasonably adjust the weights within the range;

[0046] w l,i =βC i +(1-β)S i

[0047] w g,i =γS i +(1-γ)C i

[0048] Where β and γ are weighting parameters for the relative importance between the confidence level and the severity of the anomaly, C i S represents the confidence level. i This indicates an abnormally severe level.

[0049] Furthermore, if the local repair plan conflicts with the reasonable correction range, a second learning process is triggered. The power data is judged and corrected by adjusting the parameters of local correlation learning and global correlation learning or by introducing manual verification.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. This invention divides data repair into two stages: local and global. First, it performs local learning analysis on the power operation data that needs to be repaired to obtain a local repair plan for power data anomalies. Then, it combines external correlation data and multimodal data, which are less prone to errors, to obtain a reasonable correction range for power data anomalies. It analyzes the causes of data anomalies from multiple dimensions such as equipment status, operation records, and spatial distribution, avoiding the one-sidedness of single numerical analysis, ensuring that the repair results conform to the global trend, and improving the accuracy of power data repair.

[0052] 2. This invention utilizes the LSTM+Transformer model for local learning, which can take into account both short-term temporal dependencies and long-term periodic patterns, accurately extract local features, and combine graph neural networks to identify load correlations between nodes, accurately identify outliers in a single data source or local time period, and solve the problem that traditional methods are insensitive to short-term mutations or spatial correlations.

[0053] 3. This invention generates frequent itemsets through association rules, mines abnormal association patterns between different source data, and then establishes a global time series model that can automatically capture seasonal and trend characteristics in the data, and consider the influence of external factors such as weather and holidays, further improving the ability to analyze the long-term trend of power data and the influence of external factors, improving the limitations of local repair, and ensuring that the repair results can conform to the global trend.

[0054] 4. Based on sensor accuracy and the severity of anomalies, this invention assigns weights to local repair schemes and reasonable global correction ranges. High-confidence data has higher local weights, so the repair values ​​given by the local repair schemes are adopted first. Severely abnormal data has higher global weights, so the repair values ​​are forced to conform to logic through cross-source correlation rules in the reasonable correction range. The local-global weight allocation allows the final repair result to find the optimal balance between local details and global rules.

[0055] 5. After allocating weights to local repair schemes and reasonable global correction ranges, this invention also establishes an objective function aimed at minimizing repair errors and maximizing data smoothness, thereby further balancing the accuracy and continuity of the repaired data. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the present invention;

[0057] Figure 2 This is a flowchart of the preprocessing process of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0059] Example 1

[0060] This embodiment discloses a multimodal power data repair method based on local-global association learning, the method as follows: Figure 1 As shown, the specific steps are as follows:

[0061] Step S1: Obtain initial data and perform preprocessing to identify abnormal values ​​in the power data;

[0062] Step S2: Based on the preprocessed power operation data and power data anomalies, perform local correlation learning to mine local data correlations and obtain local repair solutions for power data anomalies.

[0063] Step S3: Based on the preprocessed initial data and power data anomalies, perform global correlation learning to explore global correlation patterns across time and sources, and obtain a reasonable correction range for power data anomalies.

[0064] Step S4: Combine the local repair plan and the reasonable correction range to generate the power operation data repair result.

[0065] In step S1, the initial data includes power operation data, external correlation data, and multimodal data.

[0066] Electricity operation data: electricity consumption type, electricity consumption, electricity price, and load;

[0067] External related data: meteorological data and regional economic indicators;

[0068] Multimodal data: equipment fault descriptions, maintenance operation records, and equipment information.

[0069] Of the three types of data, power operation data and external correlation data are conventional data that can be obtained directly, while multimodal data requires certain processing before it can be obtained.

[0070] Routine data extraction: First, real-time electricity data such as electricity type, electricity consumption, electricity price, and load are collected using IoT devices such as smart sensors and smart meters. Second, data is exchanged with external institutions such as meteorological departments and economic statistics departments through API interfaces to obtain externally related data such as meteorological data and regional economic indicators. Subsequently, historical electricity operation and holiday data are directly extracted and matched from the enterprise's electricity database to ensure data accuracy and completeness.

[0071] Multimodal data parsing and extraction: Based on multimodal data such as the geographical distribution of electricity meters, text logs of electricity inspections, abnormal photos, and abnormal videos, unstructured information about power operation is extracted, such as equipment fault descriptions, maintenance operation records, and equipment information, providing additional clues for data anomaly analysis.

[0072] Missing values ​​in power data may be caused by sensor malfunctions, communication interruptions, or other reasons. Direct deletion or simple addition often disrupts the temporal continuity of the data, affecting subsequent analysis and prediction. Therefore, this application first marks the missing time periods, and then uses a local-global approach to fill in the missing data based on data characteristics and other relevant information.

[0073] In step S1, the preprocessing is as follows: Figure 2 The following are included:

[0074] Wavelet transform is used to remove high-frequency noise interference from the initial data;

[0075] Anomalies in the initial power data are identified using the 3σ principle and sliding window detection.

[0076] Align the temporal and spatial scales of power operation data, external correlation data, and multimodal data.

[0077] Wavelet transform is a commonly used signal processing method that can decompose a signal into components of different frequencies, thereby effectively removing high-frequency noise interference.

[0078] The 3σ principle states that the probability of a data point falling outside the mean plus or minus three standard deviations is extremely small, thus these points can be initially identified as outliers. Sliding window detection, on the other hand, analyzes data within a specific time window to detect trends and anomalies. When analyzing power load data, a one-hour sliding window is set, and the mean and standard deviation of the data within the window are calculated. When the data at a certain moment exceeds three standard deviations, it is marked as an outlier.

[0079] In this embodiment, the specific process of identifying outliers using the 3σ principle includes:

[0080] 1) Assume the power data sequence is x1, x2, ... x n ;

[0081] 2) First, calculate the mean.

[0082] 3) Then calculate the standard deviation.

[0083] 4) If data point x j satisfy or It is then initially identified as an outlier.

[0084] In this embodiment, the specific process of detecting outliers using a sliding window includes:

[0085] 1) Assume the power load data are y1, y2, ... y m The time interval is 1 hour;

[0086] 2) For each 1-hour sliding window, let the data within the window be y. k y k+1 , ......y k+h-1 ;

[0087] 3) Calculate the mean of the data within the window.

[0088] 4) Calculate the standard deviation of the data within the window.

[0089] 5) When y k+h (The data point that enters after the window is slid) satisfies or At that time, y k+h Mark as an outlier.

[0090] In this embodiment, when aligning the time and spatial scales of power operation data, external correlation data, and multimodal data, interpolation is used for time alignment, interpolating data from different data sources to a unified time granularity. Spatial alignment is achieved through regional grid matching, mapping data from different regions to a unified spatial grid to facilitate spatial alignment analysis and correlation mining.

[0091] The process of using interpolation includes:

[0092] (1) Let the original data time series be t1, t2, ... t n The corresponding data values ​​are v1, v2, ... v n And the time intervals are uneven;

[0093] (2) The target time series are T1, T2, ... T m The time interval is 15 minutes (assuming T1 is the start time, and T is the start time). m End time);

[0094] (3) For the target time point T j If T j Between t i and t i+1 Between them, using linear interpolation, the interpolated values ​​are:

[0095] In addition, the timestamps, units (e.g., kW→MW) and data encoding formats of power-related data have been standardized, with all timestamps unified to the ISO8601 format and power units unified to MW, to facilitate data merging and analysis.

[0096] In step S2, the process of local association learning includes:

[0097] Based on the preprocessed power operation data and power data outliers, time series prediction results are obtained using a pre-trained LSTM+Transformer model and graph neural network.

[0098] The dynamic residuals are calculated based on the time series prediction results, and an initial threshold is set by combining sliding window kernel density estimation and extreme value theory.

[0099] The dynamic residual is set to exceed the initial threshold by a generalized Pareto distribution, and the excess is fitted to calculate the dynamic threshold.

[0100] Time-series prediction results that exceed dynamic thresholds are marked as anomalies, and multi-level anomaly labels are generated based on the degree of exceedance. Based on local spatiotemporal patterns, a local repair scheme for abnormal values ​​in power data is output.

[0101] In the LSTM+Transformer model of this embodiment, the LSTM layer is responsible for capturing local temporal dependencies, and the Transformer layer models global sequence relationships through a multi-head self-attention mechanism. In addition, position encoding and timestamp embedding are introduced to enhance the model's ability to perceive periodic (such as daily / weekly / seasonal cycles) and trend (such as load increase / abrupt changes) features.

[0102] Graph Neural Networks (GNNs) model the topology of power grids based on data from power grid GIS and SCADA systems. They represent nodes (such as substations and user nodes) and edges (such as transmission lines) in the power grid as graph structures. Through information transmission and aggregation between nodes, they learn the load correlation and spatiotemporal propagation patterns between nodes.

[0103] The expression for calculating dynamic residuals is:

[0104]

[0105] Among them, y t This refers to the actual value, i.e., the abnormal value in the power data. r represents the time series prediction result. t For dynamic residuals.

[0106] Within the sliding window [t-ω,t] (where ω is the window size), the kernel density estimate of the residuals is:

[0107]

[0108] Where n is the number of samples in the window, K is the kernel function, and h is the bandwidth.

[0109] A dynamic threshold is set based on extreme value theory (EVT). The initial threshold u0 can be determined through historical data analysis or preset. It is assumed that the excess amount x = ru after the residual exceeds the threshold u follows a generalized Pareto distribution (GPD):

[0110]

[0111] Where δ and β are fitting parameters, α is the significance level, x is the amount by which the dynamic residual exceeds the initial threshold, and F(x; δ, β) is the generalized Pareto distribution.

[0112] By fitting the excess, the parameters δ and β can be estimated, and then the dynamic threshold can be calculated.

[0113] The expression for calculating the dynamic threshold is:

[0114]

[0115] Where u is the dynamic threshold, u0 is the initial threshold, δ and β are fitting parameters, α is the significance level, P is the exceedance probability of the dynamic threshold, x is the amount by which the dynamic residual exceeds the initial threshold, and F(x; δ, β) is the generalized Pareto distribution.

[0116] By combining topological correlation analysis, for local outliers, alternative value suggestions (local repair solutions) are given based on the changing trends and correlation patterns of the surrounding data, using methods such as correlation learning and transfer learning.

[0117] In step S3, the global association learning process includes:

[0118] Establish temporal and spatial relationships between preprocessed power operation data, external correlation data, and multimodal data, and generate frequent itemsets through association rules;

[0119] Calculate the occurrence frequency of each frequent itemset, and filter out frequent itemsets whose occurrence frequency is greater than a preset first threshold;

[0120] Based on frequent itemsets, generate all possible association rules, calculate the confidence of the association rules, and filter out association rules with a confidence greater than a preset second threshold.

[0121] Based on the association rules, a global time series model is established, and based on the global time series model, a reasonable correction range is generated to limit the reasonable correction range of abnormal values ​​in power data.

[0122] By mining association rules (using the Apriori algorithm), abnormal association patterns can be discovered in multi-source data. For example, in hot weather, electricity consumption increases significantly due to measures such as air conditioning. In hot weather, a region may frequently experience sudden increases in industrial load and power line losses, which may indicate that power equipment is overloaded or malfunctioning. Therefore, it can be inferred that power equipment overload or malfunction is more likely to occur in that region during hot weather.

[0123] The expression for the global time series model is:

[0124] y(t)=g(t)+s(t)+h(t)+∈(t)

[0125] Where g(t) represents the trend term, that is, the change trend of the time series in non-periodic terms; s(t) represents the periodic term, that is, the change trend of the time series in periodic terms (generally in units of weeks or years); h(t) represents the holiday influence trend, that is, the potential impact of non-fixed periodic holidays in the time series on the predicted value; and ∈(t) represents the error term that follows a Gaussian distribution, that is, the fluctuations that the model did not predict.

[0126] Establishing a global time series model can automatically capture seasonal and trend characteristics in the data, and take into account the impact of external factors such as weather and holidays, thereby improving the ability to analyze the long-term trends of power data and the impact of external factors.

[0127] Step S4, the process of determining the local repair plan and the reasonable scope of correction includes:

[0128] The confidence level of each initial data point is defined based on the sensor accuracy used to acquire the initial data, and the severity of the anomalies in the power data outliers is calculated.

[0129] The weights of local repair schemes and reasonable correction ranges are calculated based on the confidence level of each initial data and the severity of the abnormality of the power data anomalies.

[0130] Based on the weights of local repair schemes and reasonable correction ranges, an objective function is constructed with the goal of minimizing repair error and maximizing data smoothness.

[0131] The objective function is solved using gradient descent and a genetic algorithm to obtain the optimal repair value.

[0132] Genetic algorithms, on the other hand, simulate biological evolution processes, such as selection, crossover, and mutation, to find the optimal repair value, minimize repair errors, and ensure data smoothness.

[0133] The expression for the objective function is:

[0134]

[0135] Where α is a weighting parameter used to balance the relative importance between minimizing repair error and maximizing data smoothness, E error To correct the error, E smooth Here, n represents the data smoothing term, and n is the total number of power data points. This is the repair value for the i-th data. w represents the actual observed value of the i-th data point. l,i w represents the weight of the local repair scheme. g,i To reasonably adjust the weights of the range.

[0136] A more detailed process for weighting local repair schemes and reasonable correction ranges includes:

[0137] (1) Define data confidence related variables

[0138] Suppose there are n data points in total. For the i-th data point, let the sensor accuracy be denoted as a. i , here a i ∈[0,1], a larger value indicates higher sensor accuracy. Define the data confidence level C. i With sensor accuracy a i Proportional, i.e., C i =k1a i , where k1 is a proportionality constant.

[0139] (2) Define variables related to the severity of anomalies

[0140] Let S be the severity of the anomaly at the i-th data point. i Similarly, S i ∈[0,1], a larger value indicates a more severe anomaly. The severity of the anomaly is determined by the ratio of the deviation to the threshold. Assume the actual value is... Predicted value Deviation is If a base threshold T is set, then (when At that time, S i It is proportional to the ratio of deviation to threshold; when At that time, S i Take the maximum value of 1).

[0141] (3) Calculate the weight w of the local repair scheme l,i

[0142] Weight w of local repair scheme l,i Taking into account the data confidence level C i and severity of abnormality S i , can be defined as:

[0143] w l,i =βCi +(1-β)S i

[0144] Where β∈[0,1] is a weighting parameter used to balance the influence of data confidence and anomaly severity on the weights of local repair schemes. When β=1, the weights of local repair schemes are determined only by data confidence; when β=0, they are determined only by anomaly severity.

[0145] (4) Calculate the weight w of the reasonable correction range g,i

[0146] Reasonable adjustment of range weight w g,i This approach also considers both data confidence and anomaly severity, but with different weights than local remediation schemes to reflect the differences between the two. It can be defined as:

[0147] w g,i =γS i +(1-γ)C i

[0148] The weighting parameter γ here works in conjunction with β when calculating the weights of local repair schemes to achieve different weighting strategies. For example, when it is desirable to place more emphasis on data confidence (with a larger β) in the local repair scheme, the weight of the appropriate correction range will place relatively more emphasis on the severity of the anomaly.

[0149] In another embodiment, if the local repair scheme conflicts with the reasonable correction range, a second learning is triggered, and the power data is judged and corrected by adjusting the parameters of local correlation learning and global correlation learning or by introducing manual verification.

[0150] In another embodiment, the repaired data is also input into a historical model trained based on historical data, and the deviation between the predicted results and the actual values ​​is compared. The repair effect is evaluated by calculating indicators such as root mean square error (RMSE) and mean absolute error (MAE).

[0151] By dynamically updating model parameters based on new data, the adaptability of data repair can be improved. Over time, the power system will continuously generate new data. Through incremental learning, new data can be incorporated into the model training, enabling the model to adapt to changes in the power system in a timely manner.

[0152] Example 2

[0153] Based on Embodiment 1, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the multimodal power data repair method as described above for local-global association learning.

[0154] At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the aforementioned multimodal power data restoration method based on local-global correlation learning. Of course, besides software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0155] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0156] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0157] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multimodal power data repair method based on local-global correlation learning, characterized in that, The method includes: Acquire initial data and perform preprocessing to obtain abnormal values ​​in power data. The initial data includes power operation data, external correlation data, and multimodal data. Based on the preprocessed power operation data and power data anomalies, local correlation learning is performed to explore local data correlations and obtain local repair solutions for power data anomalies. Based on the preprocessed initial data and power data outliers, global correlation learning is performed to explore global correlation patterns across time and sources, and to obtain a reasonable correction range for power data outliers. Based on the aforementioned local repair scheme and reasonable correction range, power operation data repair results are generated.

2. The multimodal power data repair method based on local-global correlation learning according to claim 1, characterized in that, The power operation data includes electricity consumption type, electricity consumption, electricity price, and load; the external correlation data includes meteorological data and regional economic indicators; the multimodal data includes equipment fault descriptions, maintenance operation records, and equipment information.

3. The multimodal power data repair method based on local-global correlation learning according to claim 1, characterized in that, The process of local association learning includes: Based on the preprocessed power operation data and power data outliers, time series prediction results are obtained using a pre-trained LSTM+Transformer model and graph neural network. The dynamic residuals are calculated based on the time series prediction results, and an initial threshold is set by combining sliding window kernel density estimation and extreme value theory. The dynamic residual is set to exceed the initial threshold by a generalized Pareto distribution, and the excess is fitted to calculate the dynamic threshold. Time-series prediction results that exceed dynamic thresholds are marked as anomalies, and multi-level anomaly labels are generated based on the degree of exceedance. Based on local spatiotemporal patterns, a local repair scheme for abnormal values ​​in power data is output.

4. The multimodal power data repair method based on local-global correlation learning according to claim 3, characterized in that, The expression for calculating the dynamic residual is: Among them, y t This refers to the actual value, i.e., the abnormal value in the power data. r represents the time series prediction result. t For dynamic residuals; The calculation expression for the dynamic threshold is: Where u is the dynamic threshold, u0 is the initial threshold, δ and β are fitting parameters, α is the significance level, P is the exceedance probability of the dynamic threshold, x is the amount by which the dynamic residual exceeds the initial threshold, and F(x; δ, β) is the generalized Pareto distribution.

5. The multimodal power data repair method based on local-global correlation learning according to claim 1, characterized in that, The process of global association learning includes: Establish temporal and spatial relationships between the preprocessed power operation data, external correlation data, and multimodal data, and generate frequent itemsets through association rules; Calculate the occurrence frequency of each frequent itemset, and filter out frequent itemsets whose occurrence frequency is greater than a preset first threshold; Based on the frequent itemsets, generate all possible association rules, calculate the confidence of the association rules, and filter out association rules with a confidence greater than a preset second threshold. Based on the association rules, a global time series model is established, and based on the global time series model, a reasonable correction range is generated to limit the reasonable correction range of abnormal values ​​in power data.

6. The multimodal power data repair method based on local-global correlation learning according to claim 5, characterized in that, The expression for the global time series model is: y(t)=g(t)+s(t)+h(t)+∈(t) Where g(t) represents the trend term, i.e. the change trend of the time series on non-periodic terms, s(t) represents the periodic term, i.e. the change trend of the time series on periodic terms, h(t) represents the holiday influence trend, i.e. the potential impact of non-fixed periodic holidays in the time series on the predicted value, and ∈(t) represents the error term that follows a Gaussian distribution, i.e. the fluctuations that the model did not predict.

7. The multimodal power data repair method based on local-global correlation learning according to claim 1, characterized in that, The preprocessing includes: Wavelet transform is used to remove high-frequency noise interference from the initial data; Abnormal values ​​in the initial data are identified using the 3σ principle and sliding window detection. Align the time and spatial scales of the power operation data, external correlation data, and multimodal data.

8. The multimodal power data repair method based on local-global correlation learning according to claim 1, characterized in that, The process of combining the aforementioned local repair scheme with a reasonable correction range includes: The confidence level of each initial data point is defined based on the sensor accuracy used to acquire the initial data, and the severity of the anomalies in the power data outliers is calculated. The weights of local repair schemes and reasonable correction ranges are calculated based on the confidence level of each initial data and the severity of the abnormality of the power data anomalies. Based on the weights of local repair schemes and reasonable correction ranges, an objective function is constructed with the goal of minimizing repair error and maximizing data smoothness. The objective function is solved using gradient descent and a genetic algorithm to obtain the optimal repair value.

9. The multimodal power data repair method based on local-global correlation learning according to claim 8, characterized in that, The expression for the objective function is: Where α is a weighting parameter used to balance the relative importance between minimizing repair error and maximizing data smoothness, E error To correct the error, E smooth Here, n represents the data smoothing term, and n is the total number of power data points. This is the repair value for the i-th data. w represents the actual observed value of the i-th data point. l,i w represents the weight of the local repair scheme. g,i To reasonably adjust the weights within the range; w l,i =βC i +(1-β)S i w g,i =γS i +(1-γ)C i Where β and γ are weighting parameters for the relative importance between the confidence level and the severity of the anomaly, C i S represents the confidence level. i This indicates an abnormally severe level.

10. The multimodal power data repair method based on local-global association learning according to claim 1, characterized in that, If the local repair plan conflicts with the reasonable correction range, a second learning process is triggered. The power data is judged and corrected by adjusting the parameters of local correlation learning and global correlation learning or by introducing manual verification.

Images