Electric energy metering error correction method and system based on data fusion

Abstract

The invention provides an electric energy metering error correction method and system based on data fusion. Belongs to the technical field of electric energy metering. The method comprises the following steps: collecting multi-dimensional data, and preprocessing the collected multi-dimensional data; performing feature extraction on the operation data by using a convolutional neural network, and mining spatial features in the data; a recurrent neural network is adopted to extract the time sequence characteristics of the secondary loop and environmental data, and time sequence association between the data is established; through convolution kernels (local, middle and global) of different sizes and a depth separable convolution technology, the calculation complexity is reduced, spatial feature multi-level extraction of equipment operation data and environment data is realized, and the comprehensiveness and efficiency of feature expression are improved.

Description

Technical Field

[0001] The present invention provides a method and system for correcting power metering errors based on data fusion, belonging to the technical field of power metering. Background Art

[0002] In the power system, power metering is an important basis for power trading, cost accounting, and power management. However, there are many problems with traditional power metering methods, resulting in large metering errors. On the one hand, factors such as mechanical wear, temperature influence, and electromagnetic interference in the electricity meter itself cause the metering accuracy of the electricity meter to decline. For example, during the long-term operation of the electricity meter, the frictional torque between the turntable and the bearing changes, resulting in a non-proportional relationship between the rotation speed and power, thus generating errors. On the other hand, the performance of the transformer also affects the accuracy of power metering. Current transformers will cause certain data deviations due to current changes and voltage changes inside the line. The accuracy of early transformer equipment was relatively low and could not meet advanced detection standards. In addition, load changes in the secondary circuit system will also cause voltage drops, further increasing metering errors.

[0003] Currently, although there are some methods for correcting power metering errors, most of them are limited to correcting single factors and lack consideration of the comprehensive influence of multiple factors. Moreover, these methods often rely on specific equipment or parameters, with poor generality and difficulty in adapting to power metering requirements under different environments and conditions. Therefore, a more comprehensive, accurate, and general power metering error correction method is needed. Summary of the Invention

[0004] The present invention provides a method and system for correcting power metering errors based on data fusion to solve the problems mentioned in the above background art: The method for correcting power metering errors based on data fusion proposed by the present invention includes: S1. Collect multi-dimensional data and preprocess the collected multi-dimensional data; S2. Use a convolutional neural network to extract features from the operation data and mine the spatial features in the data; use a recurrent neural network to extract the temporal features of the secondary circuit and environmental data and establish the temporal correlation between the data; S3. Fusion the extracted spatial features and temporal features; deeply mine and analyze the multi-dimensional data; S4. Train the built-in error correction model; evaluate and verify the trained error correction model; feedback the evaluation results to form a closed-loop feedback mechanism.

[0005] The power metering error correction system based on data fusion proposed by the present invention includes a memory, a processor, and a computer program stored on the memory and operable on the memory. The processor executes the program to implement the power metering error correction method based on data fusion as described in any one of the above.

[0006] Advantages of the present invention: Through convolution kernels of different sizes and depthwise separable convolution technology, while reducing the computational complexity, multi-level extraction of spatial features of device operation data and environmental data is achieved, improving the comprehensiveness and efficiency of feature expression; based on the temperature change rate, the time window is dynamically adjusted, combined with the physical quantity-window size joint optimization model, enabling the method to flexibly cope with environmental mutations, significantly improving the robustness and stability of the model in complex environments. The cross-modal feature interaction layer fuses the spatial features of secondary circuit data and environmental data, combined with the capture of temporal dependencies by the recurrent unit, to achieve joint modeling of spatio-temporal features and improve the accuracy of error correction. The channel attention mechanism dynamically assigns weights to strengthen key channel features; depthwise separable convolution reduces the number of parameters, combined with gradient monitoring and truncation mechanisms, effectively alleviating the problem of gradient vanishing / explosion and improving the training efficiency and model convergence speed. The sliding window and the autocorrelation coefficient decay rate dynamically adjust the window size, combined with the self-attention mechanism to focus on key time points, taking into account short-term fluctuations and long-term trends, significantly improving the processing ability for irregular time series data. The environmental data is mapped as the initial parameters of the convolution kernel and jointly input into the recurrent unit with the secondary circuit data to construct a physical quantity-time series coupling model, fully exploring the correlation between environmental factors and electrical quantities and improving the physical rationality of error correction. Based on reinforcement learning for dynamic sequence length optimization, the input sequence length is automatically adjusted according to the model performance, balancing the computational cost and correction accuracy, and being applicable to diverse practical application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 It is a flowchart of the method described in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0008] The following describes the preferred embodiments of the present invention with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not intended to limit the present invention.

[0009] An embodiment of the present invention, as Figure 1 shown, is a power metering error correction method based on data fusion, and the method includes: S1. Collect multi-dimensional data through a multi-dimensional data acquisition system, and preprocess the collected multi-dimensional data; S2. Use a convolutional neural network to extract features from the operation data of devices such as electric energy meters and instrument transformers, and mine the spatial features in the data; use a recurrent neural network to extract the temporal features of secondary circuit and environmental data, and establish the temporal correlation between the data; S3. Integrate the extracted spatial features and temporal features, and use a deep learning model to establish a complex mapping relationship between the data; through data fusion, deeply mine and analyze multi-dimensional data; S4. According to the results of deep data fusion, use historical data to train the built-in error correction model, and use an optimization algorithm to determine the parameters of the model; introduce a dynamic adaptive mechanism into the error correction model to adjust the model parameters in real time according to the actual operation data; use cross-validation to evaluate and verify the trained error correction model, evaluate the performance indicators of the model, input the multi-dimensional data collected in real time into the trained dynamic adaptive error correction model to correct the electric energy measurement data in real time; further optimize and adjust the error correction model according to the evaluation results of the correction results; and feedback the evaluation results to form a closed-loop feedback mechanism.

[0010] The working principle of the above technical solution is: collect and preprocess multiple data sources related to electric energy measurement through a multi-dimensional data acquisition system; the collected data includes: Electric energy meter data: reflects the actual measurement situation of electric energy; Instrument transformer data: records signals in the power system such as current and voltage; Secondary circuit data: provides real-time status information for equipment control in the power system; Environmental data: environmental parameters such as temperature and humidity will affect the operation of the equipment; Extract spatial features from the operation data of devices such as electric energy meters and instrument transformers through a convolutional neural network (CNN) to capture the spatial relationship between the device state and electric energy metering; extract the temporal features in the secondary circuit and environmental data through a recurrent neural network (RNN) to identify the change trends of these data in the time dimension and their correlation with the device operation state; after feature extraction, the spatial features and temporal features will be fused. At this time, the deep learning model will establish a complex mapping relationship between multi-dimensional data. The data fusion process can fully explore the mutual influence between different data sources, reveal deep-seated laws, and thus improve the accuracy of electric energy metering data; based on the fused deep data, train the error correction model using historical data. By adopting optimization algorithms such as the backpropagation algorithm and the stochastic gradient descent method, continuously adjust and optimize the model parameters to reduce the error in electric energy metering. A dynamic adaptive mechanism is introduced to enable the model to automatically adjust its parameters according to real-time operation data; after the model training is completed, evaluate the performance of the error correction model through methods such as cross-validation. Evaluation metrics such as mean square error and mean absolute error will be used to measure the effect of the model. If the evaluation result does not meet the expectation, optimize and adjust the model until the performance meets the requirements; in actual operation, input the multi-dimensional data collected in real time into the trained error correction model for online correction of electric energy metering data to ensure the real-time and accuracy of the correction result. At the same time, evaluate the correction effect by analyzing the change in the error of the electric energy metering data before and after correction; during the electric energy metering correction process, the effect of the model needs to be continuously monitored and fed back. Through the evaluation of the correction result, further optimize and adjust the error correction model to form a closed-loop feedback mechanism.

[0011] The effects of the above technical solution are as follows: Through the collection of multi-dimensional data and in-depth data fusion, the errors in power metering can be effectively corrected, thus significantly improving the accuracy of power metering and reducing the errors caused by traditional error correction methods; By adopting a dynamic adaptive mechanism and combining real-time data to adjust model parameters, it can quickly respond during the operation of the power system, improve the real-time performance of error correction, reduce the correction delay, and ensure the timeliness of power metering; The dynamic adaptive mechanism enables the error correction model to automatically adjust according to real-time operation data, enhancing the adaptability of the system under different working conditions and adapting to the continuous changes and complex operating environment of the power system; By combining the deep learning model with historical data, rich information can be extracted from multi-dimensional data, improving the robustness of the error correction model against various interference factors and noises, so that power metering can still maintain high accuracy under different working scenarios; Through the automation of the model training and optimization process, the need for manual intervention is reduced, the impact of human errors is lowered, the automation level of the power metering system is improved, and the costs of manual maintenance and adjustment are reduced; By fusing and analyzing multi-dimensional data through deep learning, the operating state of the power system can be comprehensively understood from multiple perspectives, the data analysis ability of the power system is improved, and a more accurate basis is provided for subsequent decision-making support; Through evaluation means such as cross-validation, the efficiency and accuracy of the error correction model during training and application are ensured. Continuously optimizing and adjusting the model improves the efficiency of error correction and reduces the impact of unqualified models on the system; Through the comparative analysis of the data before and after correction, the performance of the power metering system can be continuously improved, ensuring that the quality of the corrected data continues to improve, which is helpful for more accurate power billing and load forecasting; The implementation of this technical solution makes the power metering correction process of the power system more intelligent and can adaptively adjust according to different environments and system states, further promoting the intelligent development of power operation management; By adopting an automated error correction method, manual intervention and equipment adjustment are reduced, the maintenance cost of the power metering system is lowered, and the potential losses caused by errors are reduced.

[0012] In one embodiment of the present invention, the S1 includes: S11. Set a timed acquisition task to collect multi-dimensional data at a preset time interval, and at the same time open a real-time acquisition channel to obtain data in a timely manner when an abnormal situation occurs; S12. During the data acquisition process, check the integrity of the acquired data to determine whether there is data loss or damage; if incomplete data is found, re-collect or use interpolation to repair the data; S13. Preprocess the acquired data; and label the preprocessed data to mark normal data and abnormal data.

[0013] The working principle of the above technical solution is as follows: By setting the acquisition interval (such as every minute, every hour, etc.) through a scheduled task, multi-dimensional data is acquired regularly to ensure the acquisition of a continuous and periodic data stream. At the same time, a real-time data acquisition channel is established. When an abnormal situation (such as a sudden change in current) occurs in the electricity meter or the system, the system can immediately trigger the real-time acquisition mechanism and quickly record the relevant data. This combination of scheduled and real-time acquisition can ensure the comprehensiveness and timeliness of data acquisition and capture abnormal situations. During the data acquisition process, the system will perform an integrity check on each batch of data. During the data acquisition process, the system will verify whether the data is lost or damaged. For example, if some data is not acquired due to communication failures or hardware problems, the system will be able to detect this problem in a timely manner and take measures to re-acquire or repair the data. Re-acquisition can be achieved by requesting the data again or using interpolation methods for repair to ensure the integrity and continuity of the entire data set. After the data is acquired and the integrity check is completed, the system will preprocess the data. While preprocessing the data, the system will also label the data to clarify which data belongs to the normal range and which belongs to abnormal data. These labeled data can provide key reference information for subsequent analysis and modeling, especially playing an important role in identifying and processing abnormal data.

[0014] The effects of the above technical solution are as follows: By combining the scheduled acquisition task and the real-time acquisition channel, it can ensure that data is collected periodically under normal circumstances and respond in a timely manner when abnormal situations such as sudden changes in the current of the electricity meter occur, quickly obtaining the relevant data, avoiding data loss or delay caused by accidental events, and ensuring the efficient operation of the system and the timeliness of data. Adding an integrity check mechanism during the data acquisition process can effectively detect and repair data loss or damage. Through re-acquisition or interpolation repair, the continuity and accuracy of the acquired data are guaranteed, thus reducing subsequent analysis deviations caused by data loss or errors. Marking the difference between normal data and abnormal data helps improve the accuracy of the abnormal detection model, enabling the system to identify and respond to potential problems more quickly and accurately. Through the real-time data acquisition and abnormal handling mechanism, even if an abnormality occurs during the acquisition process, the system can still quickly remedy it, avoiding system errors or performance degradation caused by data loss or damage, enhancing the stability and robustness of the system, and enabling the electricity metering system to operate stably in various environments. Preprocessing and labeling the data during the data acquisition and processing stage ensure the accuracy and availability of the data, providing a reliable basis for subsequent data mining, analysis, and decision support. The labeled data helps with retrospective analysis of historical data and provides data support for possible future abnormal situations, thus effectively improving the intelligent level of the system.

[0015] In one embodiment of the present invention, the S2 includes: S21. Slice the operation data of the device according to a certain time window, and perform convolution operations on the sliced data through multiple convolutional layers using convolutional kernels of different sizes to extract the spatial features in the data; S22. Add a pooling layer after the convolutional layer to reduce the dimensionality of the convolutional features while retaining important feature information; S23. Construct the secondary circuit and environmental data into sequence data according to the time order, and process the sequence data through a recurrent unit; S24. Encode the features output by the recurrent unit to convert the temporal features into a vector representation of a fixed length.

[0016] The working principle of the above technical solution is as follows: Slice the operation data of devices such as electricity meters and transformers according to a certain time window. For example, the data per hour can be divided into multiple 10-minute time windows. This slicing helps to capture more fine-grained local features and provides richer data for subsequent processing. Next, use the convolutional layer in the convolutional neural network to process, and perform convolution operations on the sliced data using convolutional kernels of different sizes (such as 3×3, 5×5, etc.). These convolutional kernels can extract spatial features of different scales from the data and identify change patterns of different granularities; after the convolution operation, add a pooling layer (such as a max pooling layer or an average pooling layer), whose function is to reduce the dimensionality of the convolutional features. In addition to the device operation data, other environmental data related to the device operation (such as temperature, humidity, etc.) also need to be processed. These data form sequence data in time order. For example, the temperature and humidity data per hour can form a sequence, representing the environmental changes within a certain time period. By introducing recurrent units such as recurrent neural networks or long short-term memory networks, process these time series data. These recurrent units can capture the temporal dependencies in the data, learn the long-term and short-term dependence features in the data, enabling the model to identify and process the dynamic change patterns in the time series; after being processed by the recurrent unit, encode the output temporal features into a vector representation of a fixed length. Through this encoding process, complex time series data can be compressed into a more representative vector, facilitating subsequent analysis, prediction, or other tasks.

[0017] The effects of the above technical solutions are as follows: By slicing the device operation data according to time windows and performing multi-layer convolution operations using convolution kernels of different sizes, multi-level and multi-scale spatial features can be extracted from the data. This refined feature extraction method can effectively capture the details and patterns in the device operation process, providing rich feature information for subsequent analysis and prediction; Adding a pooling layer after the convolution layer for dimensionality reduction can reduce the data dimension, thereby reducing the computational amount and storage requirements. The pooling operation can not only reduce the complexity of the model but also retain the important feature information in the data, avoiding information loss, and helping the model maintain good performance with fewer computational resources; By using a recurrent neural network or a long short-term memory network to process the secondary circuit and environmental data, the temporal dependence relationship in the data can be effectively captured. This method can learn the long-term and short-term dependence features in the time series, which is very helpful for analyzing and predicting the long-term behavior and instantaneous changes of the device. Especially in a dynamically changing environment, it can adjust the prediction strategy in a timely manner; Encoding the features output by the recurrent unit and converting them into a fixed-length vector representation enables complex temporal data to be transformed into a more concise and structured feature representation. This not only improves the efficiency of data processing but also makes subsequent classification, regression, or other machine learning tasks simpler and easier to execute, and can ensure efficient feature utilization; By comprehensively extracting spatial and temporal features, the dynamic information of device operation and environmental changes can be captured more comprehensively. By combining the advantages of convolutional neural networks and recurrent neural networks, this technical solution can improve the prediction accuracy of device states and can more accurately respond to operation changes in different scenarios, enhancing the robustness and adaptability of the system; Through the dimensionality reduction processing of the pooling layer and the fixed-length vector representation after encoding, the model can not only improve the computational efficiency but also maintain high-precision prediction ability while reducing the consumption of computational and storage resources.

[0018] In an embodiment of the present invention, the S21 includes: Align the operation data of the device and the environmental data according to a unified timestamp format; and use the timestamp interpolation method to perform linear interpolation or physical model-based interpolation on the missing timestamps; Calculate the autocorrelation coefficient of each device operation data; at the same time, divide multiple time windows to capture features of different time granularities; Calculate the information entropy of the data in each time window, introduce the environmental data as an auxiliary variable, and dynamically reduce the time window when the temperature change rate exceeds the threshold; Use convolution kernels of different sizes to extract local, mesoscopic, and global spatial features respectively; introduce depthwise separable convolution in the convolution layer to decompose the standard convolution into depth convolution and pointwise convolution; Embed the channel attention mechanism in the convolutional layer to assign different weights to the features of different channels and strengthen the extraction of key features; map the environmental data to the initial parameters of the convolutional kernel; Add a cross-modal feature interaction layer after the convolutional layer to fuse the convolutional features of the device operation data and the environmental data; introduce the change rate of the environmental data as a sensitive index for window selection, and automatically shrink the time window when the temperature gradient exceeds the threshold; Construct a physical quantity-window size joint optimization model, take the environmental data and the window size as joint variables, and determine the optimal window size through an optimization algorithm.

[0019] The working principle of the above technical solution is as follows: The device operation data (such as current, voltage, etc.) and environmental data (such as temperature, humidity, etc.) are first aligned according to a unified timestamp. Through the timestamp interpolation method, the missing data points are filled to ensure data continuity. The interpolation methods include linear interpolation and interpolation based on physical models (such as temperature interpolation based on thermodynamic equations) to ensure that all data points have consistent time stamps; By calculating the autocorrelation coefficients of each device operation data (such as current, voltage, etc.), the time dependence range of the data is determined. For example, when the autocorrelation of the current data decays to a set threshold within 10 minutes, 10 minutes is selected as the initial window size; At the same time, multiple time windows of different sizes (such as 5 minutes, 10 minutes, 15 minutes, etc.) are divided to capture features at different time granularities; By calculating the information entropy of the data within each time window, the effectiveness of the window is evaluated. Select a window with moderate information entropy to avoid information redundancy caused by too large a window or feature loss caused by too small a window. At the same time, the window size is dynamically adjusted, and environmental data (such as the temperature change rate) is introduced as an auxiliary variable. When the temperature change rate exceeds the threshold, the time window is automatically reduced to improve the response speed to environmental changes; Convolution kernels of different sizes (such as 3×3, 7×7) are used to extract features of different scales. Small convolution kernels (such as 3×3) capture local features, while large convolution kernels (such as 7×7) capture global trends; The convolution operation is decomposed into depth convolution and pointwise convolution through depthwise separable convolution, reducing the computational amount while maintaining the feature extraction ability. For example, a 5×5 convolution kernel can be decomposed into a 5×1 depth convolution and a 1×5 pointwise convolution, thereby reducing the computational amount by about 75%; A channel attention mechanism is embedded in the convolution layer, and different weights are assigned to different channels according to their importance. Through this mechanism, the extraction of key features can be strengthened. Especially when the harmonic components increase, the weight of the corresponding channel is automatically increased, thereby enhancing the ability to capture harmonic features; After the convolution layer, a cross-modal feature interaction layer is introduced to fuse the convolution features of the device operation data and the environmental data. Methods such as Bilinear Pooling are used to generate a joint feature representation through an outer product operation. The change rate of the environmental data (such as the temperature gradient) is introduced as a sensitive indicator for window size selection. For example, when the temperature gradient exceeds a set threshold (such as 5℃ / h), the time window is automatically reduced to improve the response speed to environmental changes. This method of dynamically adjusting the window size can optimize data processing in real time according to actual environmental changes; A joint optimization model of physical quantity - window size is constructed, combining environmental data (such as temperature, humidity) and window size, and an optimization algorithm (such as a genetic algorithm) is used to find the optimal window size. For example, under the conditions of a temperature of 30℃ and a humidity of 80%, the optimization algorithm will select 8 minutes as the optimal window size.

[0020] The effects of the above technical solution are as follows: By aligning the device operation data and environmental data through a unified timestamp format and applying linear interpolation or interpolation techniques based on physical models, the continuity of data from different sources in the time dimension can be ensured. This helps to solve the problems of data missing or inconsistency and improves the accuracy of subsequent analysis; By calculating the autocorrelation coefficient of the device operation data, the time dependence range of the data can be effectively determined, thereby determining a reasonable time window, avoiding the subjectivity of manually selecting the window size, and by setting different time windows, the characteristics of different time granularities can be captured to ensure multi-dimensional extraction of information; The concept of information entropy is introduced to select an appropriate time window, avoiding information redundancy caused by an overly large window or feature loss caused by an overly small window. In addition, by dynamically adjusting the size of the time window in combination with environmental data (such as temperature), the response speed of the model to environmental changes is increased, and the adaptability of data processing is enhanced; Convolution kernels of different sizes are used to extract local, mesoscopic, and global features, and depthwise separable convolution is used to reduce the computational amount while maintaining the feature extraction ability. This method optimizes the computational efficiency and effectively captures different levels of features of device operation. Combining the channel attention mechanism can dynamically adjust the feature weights of each channel and improve the ability to extract key features; The device operation data and environmental data are subjected to convolutional feature fusion, and the bilinear pooling method is introduced for feature interaction, effectively fusing the feature information of the two types of data and enhancing the model's perception ability of complex systems. In addition, the change rate of environmental data, as a sensitive indicator for window selection, can quickly respond to environmental changes and improve the flexibility and timeliness of the model; Through the joint optimization of physical quantities (such as temperature and humidity) and window size, the time window can be adaptively adjusted, optimizing the accuracy and efficiency of data processing. The introduction of optimization methods such as genetic algorithms provides a more intelligent window selection strategy and can be dynamically adjusted according to actual environmental changes. By introducing environmental data, not only the division and dynamic adjustment ability of the time window are optimized, but also the model's expressiveness and adaptability are enhanced in multiple aspects, improving the model's performance in complex and changing environments and further promoting the intelligent analysis and prediction ability based on time series data.

[0021] In one embodiment of the present invention, S23 includes: Align the secondary circuit data and environmental data according to the timestamp, and use linear interpolation to complete the missing data; Normalize the data with different dimensions and map it to a unified interval; Divide the time series into windows with a fixed length for extracting local time series features; On the basis of the fixed window, introduce a sliding window mechanism to dynamically adjust the window step size; Capture the dependence relationships of different time granularities; Calculate the decay rate of the autocorrelation coefficient of the time series. When the autocorrelation decays to the threshold, determine the effective length of the current sequence and dynamically adjust the window size; Introduce information entropy as an auxiliary index for window length selection, and select a window with moderate information entropy to balance feature richness and computational complexity; Adopt a hybrid structure of long short-term memory network LSTM and GRU to process short-term and long-term dependencies respectively; Introduce a self-attention mechanism in the recurrent unit to assign higher weights to key time points in the time series and enhance the response ability to abnormal events; Use environmental data as an external input and jointly input it into the recurrent unit with the secondary loop data; construct a physical quantity-time series coupling model to map physical quantities such as temperature and humidity to the same feature space as current and voltage data; During the training process, dynamically monitor the gradient changes of the recurrent unit. When the gradient vanishes or explodes, automatically truncate overly long sequences; based on the reinforcement learning method, dynamically adjust the sequence length according to the model performance.

[0022] The working principle of the above technical solution is as follows: Align the secondary circuit data (such as current and voltage) and environmental data through timestamps to ensure the time consistency between different data sources; for missing data, use linear interpolation to complete it to ensure the integrity of time series data; normalize data with different dimensions (for example, map the current unit A and temperature unit °C to the interval [0, 1]), and apply Z-score normalization to eliminate the skewness of the data distribution and improve the efficiency of subsequent model training; divide the time series data into windows of fixed length (for example, hourly data is divided into 6 10-minute windows); introduce a sliding window mechanism to dynamically adjust the window step size (for example, set it to 5 minutes); calculate the decay rate of the autocorrelation coefficient of the time series and set a threshold (such as 0.1) to determine the effective length of the sequence. Introduce information entropy as an auxiliary index for window length selection to balance the richness of features and computational complexity. Windows with higher information entropy represent more information, while lower entropy represents less redundant information; adopt a hybrid structure combining LSTM and GRU, which are used to handle short-term and long-term time series dependencies respectively. The LSTM layer mainly captures hourly time series dependencies, while the GRU layer is used to handle minute-level time series dependencies; perform multi-scale time series modeling through a hierarchical structure to capture both short-term and long-term dependencies simultaneously; introduce a self-attention mechanism into the recurrent unit to enable the model to automatically focus on key time points in the time series and assign higher weights to these time points. For situations such as current mutations, the self-attention mechanism can help the model focus on the time series information before and after the mutation and improve the response ability to abnormal events; use environmental data as an external input and jointly input it into the recurrent unit together with the secondary circuit data. The model can adjust the sensitivity to current data in combination with changes in environmental conditions (for example, the change in current when the temperature rises); construct a physical quantity-time series coupling model, combine a thermodynamic model to associate temperature and current data, predict the impact of temperature on current fluctuations, and use this information as an auxiliary input to the recurrent unit; during the training process, dynamically monitor the gradient changes of the recurrent unit to prevent the phenomenon of gradient disappearance or explosion. When gradient disappearance or explosion occurs, adopt the method of truncating long sequences to avoid the model being unable to learn effective features; based on the reinforcement learning method, dynamically adjust the sequence length according to the prediction error of the model. When the prediction error is large, increase the sequence length to capture more historical information; when the error is small, shorten the sequence length to reduce the computational amount and improve the efficiency.

[0023] The effects of the above technical solutions are as follows: By aligning data from different sources according to timestamps, data synchronization is ensured, and data deviation caused by time misalignment is avoided; linear interpolation is used to complete missing data, which helps to retain the continuity and stability of the data and reduce prediction errors caused by data loss; normalization processing ensures that data with different dimensions can be compared and processed on the same scale, avoiding the impact of dimensional differences on model training; Z-score standardization eliminates the skewness of the data distribution, enabling the model to better handle outliers in the data and improving the convergence speed of training; by dividing the time series into windows of fixed length and introducing a sliding window mechanism, local time series features can be effectively extracted, and the window size can be dynamically adjusted to adapt to dependency relationships at different time scales; the introduction of the decay rate of the autocorrelation coefficient helps to dynamically adjust the window length, capture the effective part of the sequence, and avoid unnecessary computational complexity introduced by overly long sequences; the introduction of information entropy as an auxiliary indicator effectively balances the richness of features and the computational complexity, ensuring that the selected window can contain sufficient information; using a hybrid structure of LSTM and GRU can capture both short-term and long-term time series dependencies, improving the model's time series modeling ability, especially suitable for processing complex power system data; the introduction of the self-attention mechanism enables the model to automatically identify key time points and improves the sensitivity of anomaly detection; taking environmental data as an external input and jointly inputting it with power data helps the model dynamically adjust its sensitivity to different factors and improve its adaptability to complex environmental changes. For example, when the temperature rises, the model can automatically adjust the prediction of current fluctuations; by constructing a coupling model between physical quantities and time series, the working state of the device under different environments can be simulated, enhancing the accuracy and robustness of the prediction; dynamically monitoring the gradient change and automatically truncating overly long sequences avoid the problems of gradient disappearance or explosion and improve the stability of the training process; dynamically adjusting the sequence length based on reinforcement learning enables the model to adaptively select the most suitable sequence length, thereby improving the prediction accuracy of the model and reducing the computational cost.

[0024] In one embodiment of the present invention, step S3 includes: S31. Concatenate the spatial features extracted by CNN and the temporal features extracted by RNN to form a comprehensive feature vector; S32. On the basis of feature concatenation, introduce an attention mechanism to weight the importance of different features and highlight key features; S33. Use a deep learning model as the model architecture for data fusion; design the number of network layers and neurons according to the characteristics of the data and the task requirements; S34. Use the fused data to train the deep learning model, and continuously adjust the parameters of the model using the backpropagation algorithm so that the model can learn the complex mapping relationship between the data. S35. During the training process, the model is regularly evaluated by using cross-validation. According to the evaluation results, the structure and parameters of the model are optimized.

[0025] The working principle of the above technical solution is as follows: Through a convolutional neural network, spatial features are extracted to capture the spatial structure information in the input data (such as images, spatial data, etc.). At the same time, a recurrent neural network is used to extract temporal features to capture the temporal dynamic changes in the data. Then, these two types of features (spatial features and temporal features) are concatenated to form a comprehensive feature vector. This concatenated feature vector can contain both spatial information and temporal information, providing more comprehensive input data for subsequent model learning. On the basis of feature concatenation, an attention mechanism is introduced. The purpose of the attention mechanism is to perform weighted processing on different features according to their importance, thereby highlighting key features and suppressing unimportant features. By calculating the weights of each feature, the model can automatically identify which features are more important for the prediction of the task and correspondingly enhance their influence on the model output. This helps to improve the performance of the model, especially when dealing with complex data, it can avoid overfitting and improve the generalization ability. Using the fused feature data as input, a deep learning model architecture is designed to process this data. The deep learning model can be a multi-layer neural network, and the specific number of network layers and neurons will be adjusted according to the characteristics of the data and the requirements of the task. At this time, the fused data (including spatial features, temporal features, and features weighted by the attention mechanism) will be used as the input of the model, and the model will process the data and learn features through multiple layers of neurons. The deep learning model is trained using the fused data. During the training process, the backpropagation algorithm is adopted to continuously adjust the model parameters to minimize the loss function, so that the model can learn the complex mapping relationship between the data. Through backpropagation, the model continuously adjusts its parameters according to the error feedback, so that the model can gradually improve the prediction accuracy. During the training process, the cross-validation method is used to regularly evaluate the model. Cross-validation can effectively detect whether the model has overfitted and at the same time provide evaluation metrics to help us judge the generalization ability of the model. According to the evaluation results of cross-validation, the structure and parameters of the model are further optimized. This includes adjusting hyperparameters such as the number of network layers, the number of neurons in each layer, and the learning rate, so as to ensure that the model can exhibit optimal performance in practical applications.

[0026] The effects of the above technical solutions are as follows: By concatenating the spatial features extracted by CNN and the temporal features extracted by RNN, it is possible to capture both the spatial and temporal information in the data simultaneously, enabling the model to understand the data more comprehensively. This feature fusion method effectively improves the model's ability to represent complex data structures; Introducing the attention mechanism can weight different features, highlight key features, and automatically identify and strengthen the features that are more important for task prediction. This helps the model avoid focusing on irrelevant features, enhances its sensitivity to key information, and thus improves the prediction accuracy; Through the attention mechanism, the model can dynamically adjust the importance of different features according to the characteristics of the input data without the need for manual specification of which features are more important. This adaptive feature selection mechanism can find effective features in complex data and improve the generalization ability of the model; Designing the number of network layers and neurons of the deep learning model according to the characteristics of the data and the task requirements can enable the model to better adapt to the needs of specific tasks. The flexibility of the deep learning architecture makes this solution highly adaptable and can be applied to different types of tasks and data; Training the model through the backpropagation algorithm and combining cross-validation for regular evaluation can effectively avoid overfitting and ensure the accuracy and generalization ability of the model in practical applications. Cross-validation provides accurate evaluation metrics to help quickly identify and adjust potential problems of the model, thereby continuously improving the performance of the model; Combining the methods of backpropagation and cross-validation can continuously optimize the parameters of the model during the training process, enabling the model to more efficiently learn the complex relationships in the data and improving the training effect of the model.

[0027] In one embodiment of the present invention, S31 includes: S311. Extract spatial features by extracting the geometric or position attributes of the original data; obtain temporal features by processing the temporal correlation of time series data; preprocess the obtained spatial data and temporal data to obtain a spatial feature set and a temporal feature set; S312. Expand the dimension of the spatial feature set to obtain an enhanced vector of spatial features, where the enhanced vector of spatial features includes coordinate information, spatial distance, and direction features; further extract the temporal feature set to obtain an enhanced vector of temporal features; the enhanced vector of temporal features includes timestamps, time intervals, and rates of change; S313. Concatenate the enhanced vector of spatial features and the enhanced vector of temporal features respectively based on different concatenation strategies to form a preliminary comprehensive feature vector; S314. Normalize the formed preliminary comprehensive feature vector to obtain a final comprehensive feature vector; perform further feature analysis and processing based on the final comprehensive feature vector to obtain a final multi-modal feature set; S315. Based on the final multi-modal feature set, model and infer the spatio-temporal dependence relationship to obtain a spatio-temporal dependence relationship graph; based on the spatio-temporal dependence relationship graph, conduct relationship inference and optimization to obtain the final relationship inference score.

[0028] The working principle of the above technical solution is as follows: By processing the geometric or position attributes of the original data, extract the space-related features, including spatial information such as the coordinate position, shape, and size of the object; process the time series data, analyze the time correlation therein, and extract the time series features. The time series features are usually related to time changes, such as periodicity, trend, fluctuation, etc.; preprocess the extracted spatial data and time series data to finally form a spatial feature set and a time series feature set; enhance the vector representation of the spatial features by expanding the dimensions of the spatial features, including coordinate information, spatial distance, and direction features; similarly, further extract the time series feature set to obtain an enhanced vector of the time series features. The enhanced vector of the time series features includes timestamps, time intervals, and change rates; adopt different splicing strategies to splice the enhanced spatial feature vector and the time series feature vector to form a preliminary comprehensive feature vector. Normalize the formed preliminary comprehensive feature vector; enable different features to be compared on the same scale, and the finally formed feature vector after normalization forms a multi-modal feature set that contains information in both spatial and time series dimensions. Based on the final multi-modal feature set, model and infer the spatio-temporal dependence relationship; based on the spatio-temporal dependence relationship graph, conduct relationship inference and optimization, thereby obtaining the final relationship inference score.

[0029] The effects of the above technical solution are as follows: By separately extracting spatial features and temporal features and enhancing them, the understanding and analysis capabilities of the model for complex spatio-temporal data are effectively improved, and the spatial and temporal dependence relationships in the data can be captured more accurately; by normalizing spatial features and temporal features, the scale differences between features in different dimensions are effectively avoided, the subsequent feature processing process of the model is simplified, and the training efficiency of the model is improved; in the electric energy metering of the power system, multiple different models are used, and different models have different requirements for input features. By performing different processing on different data with different attributes, feature sets suitable for different models can be generated, improving the performance and generalization ability of the model. Through the dimensionality expansion and enhancement of spatial and temporal features, multi-dimensional spatial enhancement vectors including coordinate information, spatial distance, direction features, etc., and temporal enhancement vectors including timestamps, time intervals, change rates, etc. are generated, further enriching the feature expression of the data; through the inference and optimization based on the spatio-temporal dependence graph, the recognition and prediction capabilities of the model for potential laws in spatio-temporal data are enhanced, the relationship inference score is improved, and the accuracy is increased; through the comprehensive application of multi-modal feature sets, the model can better adapt to different types of data scenarios and enhance its generalization ability in different fields; the normalization processing and feature splicing strategy reduce the interference of data noise and improve the stability and robustness of the model when processing unbalanced or missing data; through the effective enhancement and fusion of spatio-temporal features, the model can better distinguish important spatial and temporal features, reduce the influence of redundant information and noise, thereby reducing the prediction error and improving the accuracy of the model; by combining the spatio-temporal dependence graph with optimized inference, the model can deeply mine the potential complex relationships in spatio-temporal data, enhance the depth and breadth of spatio-temporal dependence relationship modeling, and improve the accuracy and credibility of the inference results.

[0030] In one embodiment of the present invention, the S313 includes: The time information is represented clockwise through time dependence, and the timestamp is represented counterclockwise; based on the relationship between position and time, the clockwise represented coordinate information and the counterclockwise represented timestamp are first spliced to obtain the first splicing. The spatial distance is divided into X equal small segments through a reference point, where X>2. The odd segments are represented in the positive direction, and the even segments are represented in the reverse direction. The spatially distant segments with different representations are spliced to obtain the first spatial distance; the time interval is divided into N equal parts of equal length through a gradient transition function, where N>2. The odd equal parts and the even equal parts are respectively spliced, and then the splicing results are spliced again to obtain the first time interval; based on the relative relationship between positions and the time change, the first spatial distance and the first time interval are second spliced to obtain the second splicing. Perform a third splicing of the direction feature and the change rate based on the direction and change speed of the object's movement to obtain a third-layer splicing; Extract the first 30% of the dimensions of the first-layer splicing and splice them with the second-layer splicing; extract the last 70% of the dimensions of the first-layer splicing and splice them with the third-layer splicing; splice the two spliced parts again to form a preliminary comprehensive feature vector.

[0031] The working principle of the above technical solution is as follows: First, represent the time information in the clockwise direction and represent the time stamp in the counterclockwise direction; then splice the clockwise represented coordinate information and the counterclockwise represented time stamp to form the first splicing. In the power system, the position (coordinate information) of the electric energy metering device is closely related to the time information. For example, the electricity consumption patterns of electrical equipment in different geographical locations vary in different time periods. The clockwise and counterclockwise representation methods can fuse the position and time information in a unique way, making the spliced features better reflect this spatio-temporal correlation, which helps the model better understand the impact of spatio-temporal factors on electric energy metering in the power system; the clockwise and counterclockwise representation methods introduce directional differences to the time information, making the features of different position and time combinations more distinguishable in the spliced features. In tasks such as power load forecasting, this distinguishability can help the model more accurately identify the electricity consumption characteristics under different spatio-temporal conditions and improve the forecasting accuracy; through this splicing method, the originally independent position and time information are organically combined, increasing the information dimension of the features; divide the spatial distance into multiple equal small segments through a reference point, where the odd segments are represented in the positive direction and the even segments are represented in the negative direction. For example, if the spatial distance is divided into three segments and represented as [1, 2, 3, 4, 5, 6], then the representation after splicing becomes [1, 2, 4, 3, 5, 6], thus changing the representation order and directionality of the spatial features. In the power system, the spatial distance between devices is relative. The positive and negative directions represent the spatial distances of different segments, which can reflect different aspects of the spatial relationship between devices. For example, there are mutual influences between some devices, and this influence varies with the positive and negative directions of the distance. By representing the spatial distance in this way, the spatial interaction between devices can be more accurately characterized. Then divide the time interval into multiple equal shares through a gradual transition function, and splice the odd and even equal shares separately; for example, if the time interval is divided into three segments and represented as [1, 2, 3, 4, 5, 6], then after splicing, it forms [1, 2, 5, 6, 3, 4], representing and fusing the time interval in different orders; the gradual transition function divides the time interval, which can analyze the changes within the time interval more carefully. After splicing the odd and even equal shares separately and then splicing them as a whole, the change characteristics of different stages within the time interval can be highlighted, enabling the model to better capture the dynamic change law within the time interval. In the power system, such as the change of power load within a day is not uniform, and through this method, the rhythm and pattern of load change can be more accurately described; compared with directly using the original time interval, the first time interval after division and splicing can provide more information about time change. Perform the second splicing on the already processed spatial distance and time interval according to the relative change relationship between the position and time; the second splicing is to splice the first spatial distance and the first time interval based on the relative relationship between the positions and the time change.In the power system, spatial distance and time variation are interrelated. For example, during the power transmission process, the farther the spatial distance, the longer the transmission time and the greater the loss. Through the second splicing, these two types of information can be integrated together, enabling the model to comprehensively consider the impact of spatio-temporal distance factors on electric energy metering; based on the motion direction and change speed of the object, these dynamic features are spliced with other features for the third time; the first 30% of the dimensions are extracted from the features of the first splicing and spliced with the features after the second splicing; the last 70% of the dimensions are extracted from the features of the first splicing and spliced with the features after the third splicing; the first splicing contains position and time information, the second splicing integrates spatio-temporal distance information, and the third splicing combines direction features and change rate. Splicing in the ratio of 30%:70% can balance the weights of different types of feature information in the preliminary comprehensive feature vector. In the power system, different types of features have different degrees of influence on electric energy metering, and this ratio allocation can enable the model to more reasonably utilize various feature information; through this ratio splicing method, the preliminary comprehensive feature vector can contain both the basic correlation information of position and time (the first splicing part) and highlight important features such as spatio-temporal distance and direction change (the second and third splicing parts); in this way, the model can more precisely combine the features at each level to form a preliminary comprehensive feature vector.

[0032] The effects of the above technical solutions are as follows: By representing time information in different directions, different dimensions of time changes can be captured more precisely, effectively identifying the patterns of time changes, improving the prediction accuracy, and enhancing the recognition ability of time series patterns; By dividing the spatial distance into multiple small segments and combining forward and reverse representations, the expression accuracy of spatial features is improved; By reasonably dividing and splicing the spatial distance and time interval, the generation of redundant features is reduced, and at the same time, the complexity in the calculation process is optimized, avoiding the excessive feature accumulation in traditional feature engineering and improving the calculation efficiency; By selecting the dimensions of the first-layer splicing, while maintaining high-efficiency calculation, the accuracy of features can be ensured, avoiding the risk of overfitting easily caused by simple splicing in the prior art; By combining and splicing the moving direction and changing speed of the object, the recognition ability of the model for the dynamic changes of the object's behavior is enhanced, improving the accuracy; The equal-length division and splicing method of the time interval can help the model understand the change rules in the time series, improving the performance of the model in dynamic change tasks. By gradually fusing the features of the first-layer splicing, second-layer splicing, and third-layer splicing, the complex relationships between space and time can be captured at different levels, thereby enhancing the model's understanding of spatio-temporal data; The splicing of space and time fully considers the relative relationships between positions and the changes in time, thereby improving the model's adaptability to spatio-temporal dynamic changes in complex environments; Through multi-level and multi-dimensional splicing, the model can analyze data from multiple perspectives, helping to reduce the dependence on specific patterns, thereby enhancing the generalization ability of the model; Through a reasonable feature selection and splicing strategy, the introduction of excessive redundant information is effectively avoided, reducing the possibility of overfitting and improving the performance of the model on new data; By combining the splicing results of different levels, a comprehensive feature vector is finally formed, enhancing the expression ability of the model; Through multi-level optimization of the features, the complexity required for each calculation is reduced, improving the processing speed and response time of the system when processing real-time data.

[0033] In one embodiment of the present invention, S4 includes: S41. Train the built-in error correction model according to the deep data fusion result; S42. Evaluate and verify the trained error correction model; S43. Feedback the evaluation result to form a closed-loop feedback mechanism.

[0034] In one embodiment of the present invention, S41 includes: S411. Build an error correction model according to the result of deep data fusion; Use the fused features as the input and the power metering error as the output to establish a mapping relationship between the input and the output; S422. Initialize the parameters of the error correction model. Use the method of random initialization or pre-training model parameter initialization to provide an initial state for the model training. S433. Collect a large amount of historical power metering data, where the historical power metering data includes multi-dimensional data and corresponding actual error values, as the data set for model training. S444. Train the error correction model and set hyperparameters. The hyperparameters include the learning rate and the number of iterations to control the training process. S455. Based on the online learning algorithm or incremental learning algorithm, enable the model to adjust parameters in real time according to the actual operation data; and based on the preset parameter update strategy, dynamically adjust the model parameters according to the difference between the actual operation data and the model prediction result.

[0035] The working principle of the above technical solution is as follows: Different sources of data (such as spatial features, temporal features, etc.) are fused through a deep learning algorithm to obtain a comprehensive feature set. Then, these fused features are used to construct an error correction model. The input of this model is the fused features, and the output is the power metering error. By establishing the mapping relationship between the input and the output, the error correction model can automatically predict and correct the power metering error; Initialize the parameters of the error correction model. Common methods include random initialization (assigning initial values to model parameters through random values) or using pre-trained model parameters for initialization. The purpose of initialization is to provide a reasonable starting point for model training, reduce the convergence time during training, and improve training efficiency; Collect a large amount of historical power metering data, including data in multiple dimensions (such as current, voltage, temperature, etc.), and the corresponding actual error values of these data. These data will be used as the training set for the model to learn the relationship between input features and power metering errors. The quality and diversity of historical data are crucial for the training effect and generalization ability of the model; Use the batch gradient descent method or mini-batch gradient descent method to train the error correction model. The batch gradient descent method adjusts model parameters by making a complete update for all training samples at once, while the mini-batch gradient descent method updates using a part of the data (mini-batch) in each iteration, with higher computational efficiency. During this process, set appropriate hyperparameters, such as the learning rate and the number of iterations, to control the training process of the model and ensure that the model can converge to the optimal solution; In order to enable the model to adapt to the actual data changes during operation in real time, introduce the online learning algorithm or incremental learning algorithm. Through these algorithms, the model can adjust parameters immediately after receiving new data to ensure that it can continuously update and improve according to the new input data. Through the preset parameter update strategy, the model can dynamically adjust its own parameters according to the difference between the real-time operation data and the prediction result, maintaining an efficient power metering correction ability, especially under different environmental or load conditions.

[0036] The effects of the above technical solution are as follows: Through the training of deep data fusion and error correction models, errors in power metering can be effectively corrected, thereby improving the accuracy of power metering. The model can dynamically adjust the correction parameters of power metering according to data characteristics in different dimensions, thus reducing the influence of external environment and equipment factors on the metering results; By adopting online learning and incremental learning algorithms, the error correction model can update parameters in real time according to actual operation data, ensuring that the model can adapt to the changing environment and operating conditions in the power system; By using batch gradient descent or mini-batch gradient descent methods for training, the model can efficiently process a large amount of historical power metering data and converge to the optimal solution in a relatively short time. The reasonable setting of hyperparameters can further accelerate the training process of the model and reduce unnecessary computational burdens; Deep data fusion comprehensively processes data characteristics in multiple dimensions, which helps to improve the generalization ability of the model and avoid the overfitting problem caused by relying only on a single feature. The diversity and richness of historical power metering data provide sufficient samples for the model, enhancing its prediction ability for unknown data; The strategy of adjusting model parameters based on real-time data can ensure that the error correction model is optimized at any time. Through this dynamic adjustment mechanism, the model can continuously learn from the experience of new data, optimize the correction strategy, and improve the stability and reliability of power metering in long-term operation; Since this solution can perform real-time correction and optimization on the basis of automation, it reduces the need for manual intervention and the cost of manual adjustment and maintenance. The system can learn and correct autonomously, improving the long-term stability and reliability of power metering and reducing errors caused by human operation.

[0037] In one embodiment of the present invention, S42 includes: S421. Use the cross-validation method to evaluate and verify the trained error correction model. Divide the data set into multiple subsets, and take each subset as the test set in turn, with the remaining subsets as the training set, and calculate the average performance index of the model; S422. Evaluate the performance index of the model to judge the accuracy and reliability of the model; Input the multi-dimensional data collected in real time into the trained dynamic adaptive error correction model to perform real-time correction on the power metering data; S423. Output the corrected power metering data to the monitoring system or metering equipment; and evaluate and analyze the corrected power metering data, and calculate the error change before and after correction.

[0038] The working principle of the above technical solution is: the error correction model is evaluated and verified by the cross-validation method. The performance of the model is evaluated by dividing the data set into multiple subsets, and each subset is used as a test set in turn, and the remaining subsets are used as training sets. This method can avoid overfitting and ensure the generalization ability of the model, thereby improving the performance of the model on different data sets. Evaluate multiple performance indicators of the model. These indicators are used to measure the accuracy of the model prediction, the size of the error, and the degree of fit of the model to the data changes; Once the model is evaluated and obtains good performance indicators, the system will input the multi-dimensional data collected in real time (such as voltage, current, load, etc.) into the trained error correction model, and the model will perform real-time power metering correction based on these input data. The corrected power metering data will be output to the monitoring system or metering equipment for further analysis and monitoring. The corrected data can more accurately reflect the actual power consumption and help operation and maintenance personnel to perform more accurate power management. In addition, the corrected power metering data needs to be further evaluated and analyzed. Specifically, by calculating the error changes before and after correction (such as the absolute value change of the error, the relative error change, etc.), the effect of error correction can be intuitively evaluated. For example, if the error before correction is large, but the error after correction is significantly reduced, then the effect of the correction model is considered successful. This analysis process also helps to continuously optimize the error correction model.

[0039] The effect of the above technical solution is: by training and verifying the error correction model, the error in electric energy metering can be significantly reduced and more accurate data can be provided. This makes the power system more accurate in metering, monitoring and management, thereby improving energy utilization efficiency and reducing unnecessary losses; the dynamic adaptive error correction model can be used to make real-time corrections based on multi-dimensional data (such as voltage, current, load, etc.) collected in real time. By evaluating and verifying the performance of the model through methods such as cross-validation, the overfitting phenomenon of the model can be effectively avoided and the generalization ability of the model can be improved. The model is tested on different subsets to ensure its stability and reliability in a variety of different power system conditions; the corrected electric energy metering data will be output to the monitoring system or metering equipment for real-time monitoring and analysis by operation and maintenance personnel. This not only helps to improve the efficiency of daily management, but also provides accurate data support for decision-making, ensuring that the operation of the power system is more efficient and safe; evaluating and analyzing the corrected electric energy metering data can provide valuable feedback for further optimization of the model. By calculating the error changes before and after the correction, the correction effect can be understood in real time, and the model can be further improved based on these feedbacks to improve the accuracy of the overall power metering. By reducing the error in power metering, unnecessary energy waste and electricity fee calculation errors can be avoided, which can effectively reduce the company's operating costs. In addition, accurate power metering helps to achieve better load management and further improve energy conservation.

[0040] In one embodiment of the present invention, S421 includes: Introduce environmental data as a stratification basis, divide the dataset into multiple environment-related layers; dynamically adjust the stratification threshold according to the change range of environmental physical quantities; In time series data, use the sliding window method to divide subsets, each subset containing consecutive time points; dynamically adjust the sliding window length according to the time resolution of the power metering data; Perform clustering analysis on the geographical location information in the multi-dimensional data, and divide the data points with similar geographical locations into the same subset; Assign weights to each feature in the multi-dimensional data. During the model training process, dynamically adjust the weights according to the change in the importance of the features; use environmental data as auxiliary information for outlier detection; mark the corresponding power metering data as potential outliers; Test the similarity of feature distributions between different subsets; when the distribution similarity between subsets is too high, readjust the stratification or clustering parameters; extract time series features from the environmental data and use them as auxiliary inputs for cross-validation; embed the time series features into the subset division process to enhance the time dependence of subset division; Calculate the influence degree of the change in environmental physical quantities on the subset division result. According to the sensitivity analysis result, set a sensitivity threshold. When the change in environmental physical quantities exceeds the threshold, trigger the re-optimization of subset division; During the model training process, monitor the change in environmental physical quantities in real time and dynamically adjust the parameters of cross-validation; automatically select the optimal cross-validation strategy according to the change pattern of environmental physical quantities; Verify whether the distribution of environmental physical quantities between different subsets is consistent, and quantify the consistency of the distribution of environmental physical quantities between different subsets.

[0041] The working principle of the above technical solution is as follows: Taking environmental data (such as temperature volatility, humidity gradient, etc.) as the basis, the original data set is stratified. By dynamically adjusting the stratification threshold (such as increasing the number of layers when the temperature volatility exceeds ±2°C), it is ensured that each subset has consistency in environmental characteristics; the specific operation is to determine the partitioning method of different data based on the changes in environmental data, further ensuring that the characteristics of each data subset are more uniform and consistent; the sliding window method is used to partition time series data to ensure that each subset contains continuous time points. This method adapts to the time resolution of power metering data (such as recording a data point every 15 minutes) and dynamically adjusts the window size according to the actual situation to avoid mixing data across days or seasons and ensure the time consistency of data subsets; cluster analysis is performed on the geographical location information in the data, and points with similar geographical locations are divided into the same subset. According to the electricity consumption characteristics of different regions (such as the differences in electricity use between cities and rural areas), the clustering radius is dynamically adjusted to ensure more consistent spatial characteristics within the same subset; weights are assigned to each feature (such as voltage, current, power factor), and the determination of the weights is based on the correlation between the feature and the target variable (such as power metering error). During the training process, as the importance of the features changes, the weights are dynamically adjusted to avoid the adverse effects of irrelevant features on data partitioning; environmental data (such as lightning activity frequency) is used to assist in detecting outliers. If lightning activity is frequent, it will affect the accuracy of power metering data, and thus mark the relevant data as potential outliers; these outliers are repaired by interpolation or replacement with neighboring values to ensure the integrity and consistency of the data within the subset; the electricity data in different environments or geographical locations may follow different distributions. Directly mixing and training will lead to model bias or a decrease in generalization ability, and will also result in unreasonable subset partitioning (such as over-fine stratification or too large clustering radius); the similarity of feature distributions between different subsets is tested by using KL divergence or JS divergence; if the feature distributions between some subsets are too similar, the stratification or clustering parameters are adjusted to increase the differences between different subsets and enhance the generalization ability of the model; time series features in environmental data (such as daily changes in temperature, seasonal fluctuations in humidity) are extracted and used as auxiliary inputs for cross-validation. Methods such as autoencoders or LSTM (Long Short-Term Memory) networks are used to embed these time series features into the subset partitioning process, making the data partitioning more compliant with time dependence; sensitivity analysis is used to calculate the degree of influence of changes in environmental physical quantities on subset partitioning (such as the influence of temperature volatility on the partitioning result). When the changes in environmental physical quantities exceed a certain threshold, re-optimizing subset partitioning is triggered to ensure that the partitioning result is more sensitive to environmental changes; during the model training process, the changes in environmental physical quantities are monitored in real time, and the parameters of cross-validation (such as the number of subsets, stratification threshold) are dynamically adjusted. According to the change pattern of environmental physical quantities, a cross-validation strategy is selected to improve the accuracy of the model; the consistency of the distribution of environmental physical quantities between subsets is verified.Use statistical methods such as chi-square test or t-test to quantify the consistency of environmental characteristics among subsets and ensure that the integrity of environmental characteristics is not damaged during the data partitioning process.

[0042] The effects of the above technical solutions are as follows: By dynamically adjusting the data stratification threshold in combination with changes in environmental data such as temperature and humidity, the consistency of each subset in terms of environmental characteristics is ensured, reducing the heterogeneity of data under different environmental conditions and improving the stability of model training; Through the sliding window method and geographical location clustering, the correlation of time series and geographical location is considered, making each subset not only coherent in time but also having consistent characteristics in space, thus enhancing the spatio-temporal consistency of data modeling; By combining environmental factors (such as lightning activity frequency) for outlier detection, potential abnormal data can be identified more accurately and repaired through interpolation or neighboring value replacement, avoiding the negative impact of abnormal data on model training; By introducing measurement methods such as KL divergence or JS divergence, the distribution differences between subsets can be monitored, and the stratification or clustering parameters can be dynamically adjusted according to environmental changes, ensuring the independence and difference between subsets and preventing problems such as overfitting or uneven data distribution; By extracting the time series characteristics of environmental data (such as temperature change patterns and humidity fluctuations) and using them as auxiliary inputs for cross-validation, the accuracy and generalization ability of the model are further improved, especially in terms of long time scales and seasonal changes; By real-time monitoring the changes in environmental physical quantities and dynamically adjusting the cross-validation strategy, the most appropriate validation method can be selected under different environmental conditions, further improving the reliability and generalization ability of the model; By calculating the influence degree of environmental physical quantities on subset partitioning, the sensitivity of the model can be evaluated, and the data partitioning strategy can be flexibly adjusted according to the sensitivity analysis results to avoid unnecessary errors. Through the above technical solutions, not only the randomness problem of subset partitioning is solved, but also the robustness and physical meaning of cross-validation are enhanced through technical means such as the introduction of environmental physical quantities, multi-dimensional feature weighting, and time series embedding; At the same time, by combining the sensitivity and consistency verification of physical quantity changes, the reliability and interpretability of cross-validation results are ensured.

[0043] In one embodiment of the present invention, the S43 includes: S431. Further optimize and adjust the error correction model according to the evaluation of the correction result; If the performance of the model does not meet the requirements, the structure, parameters or training algorithm of the model can be adjusted to improve the correction effect of the model.

[0044] S432. Optimize the feature extraction method in the data fusion process, convert different feature combinations and feature selection methods, and improve the quality and representativeness of features; S433. Feed back the evaluation of the correction result to the model training and data acquisition processes to provide a basis for further optimizing the model and adjusting the data acquisition strategy; form a closed-loop feedback mechanism to continuously optimize and improve the error correction model.

[0045] The working principle of the above technical solution is as follows: During the data processing and analysis, the error correction model evaluates the difference between the actual result and the expected result and feeds back the correction effect of the model. If the corrected performance does not meet the expectation, the system will further optimize the model. The optimization content may include adjusting the model structure, parameters, or training algorithm. The purpose of this process is to improve the accuracy and adaptability of the correction model, thereby enhancing the overall system accuracy. In the data fusion process, feature extraction is a crucial step, and the quality and representativeness of the features directly affect the performance of the model. By optimizing different feature combinations and feature selection methods, the expression ability and discrimination of the features can be effectively improved. The optimized features can better reflect the core information of the data, thus providing more effective input for the error correction model. While continuously optimizing the correction model and feature extraction method, the correction result will also be fed back to the model training and data acquisition processes. By evaluating the correction result, it can guide the subsequent data acquisition strategy and model training process. For example, if the correction effect is poor, it is necessary to adjust the data acquisition method or further adjust the model structure. Through this feedback mechanism, a closed loop is formed to ensure that the system can self-adjust and continuously optimize.

[0046] The effect of the above technical solution is as follows: By optimizing and adjusting the error correction model, the performance of the correction model can be improved for different data environments and requirements, ensuring that the model can correct errors more accurately, thereby enhancing the overall correction effect of the system. By optimizing the feature extraction method and converting different feature combinations and selection methods, more representative and discriminative features can be extracted from the data, thereby improving the quality of the model input and enhancing the performance of the model in data fusion. Feeding back the correction result to the model training and data acquisition processes can not only provide a basis for further optimizing the model but also adjust the data acquisition strategy in real time. This closed-loop feedback mechanism ensures that the system can dynamically adapt to different environmental changes, continuously optimize the error correction effect, and thereby enhance the intelligence and adaptability of the entire system. Through continuous optimization and adjustment, the system can achieve self-improvement and adapt to the changing data environment and requirements. This adaptive ability improves the long-term effectiveness of the system, reduces the need for manual intervention, and lowers the system maintenance cost; by continuously optimizing and adjusting the parameters and features of the model, the correction model can better adapt to different types of data and complex application scenarios, enhancing the generalization ability of the model and improving its stability and reliability in practical applications.

[0047] In one embodiment of the present invention, the S432 includes: Introduce environmental physical quantities as a bridge to map different data sources to a unified physical feature space; distribution differences in the space; Use the DTW algorithm to align time series data to solve the problem of inconsistent sampling frequencies of different data sources; dynamically adjust the interpolation strategy according to the changes in environmental physical quantities; Conduct Granger causality tests on different data sources to identify the causal relationships between data sources and avoid interference from irrelevant data sources in feature extraction; based on the causal relationships, construct Bayesian networks or structural equation models to quantify the dependence strength between data sources and guide feature selection; Generate new features using environmental physical quantities, and use polynomial regression or factorization machines to model the interaction effects between different features and mine the non-linear relationships between data sources; Design a multi-modal autoencoder, input different data sources into the shared encoding layer to learn cross-modal joint feature representations; introduce an attention mechanism in the autoencoder to dynamically allocate weights to different data sources, highlight key features, and suppress noise features; According to the correlation between environmental physical quantities and the target variable, screen out features that are highly correlated with the changes in physical quantities; Use environmental physical quantities as fusion weights to weight-fuse the features of different data sources to generate a comprehensive feature vector; use mutual information or conditional entropy to evaluate the amount of information between the fused features and the target variable; Verify the distribution consistency of the fused features under different environmental physical quantity conditions to ensure that the feature extraction does not damage the physical associations between data sources; use KL divergence or JS divergence to quantify the distribution differences of the fused features under different physical quantity conditions; use environmental physical quantities as an auxiliary evaluation index for model performance.

[0048] The working principle of the above technical solution is: by introducing environmental physical quantities (such as temperature, humidity, etc.) as a bridge, different data sources (such as voltage, current, meteorological data) are uniformly mapped to a shared physical feature space. This allows various types of data to be compared and analyzed in the same physical background, thereby reducing the differences between different data sources; using technologies such as non-negative matrix factorization (NMF) or generative adversarial networks (GAN) to learn the optimal mapping matrix between data sources and minimize the distribution differences of different data sources in the physical feature space. In this way, different data sources can be better unified into a common representation space; using the dynamic time warping (DTW) algorithm to align time series data (such as electricity metering data and meteorological data) to solve the problem of inconsistent sampling frequencies between different data sources. Through this alignment method, the consistency and comparability of data in the time dimension can be ensured; according to the sudden changes in environmental physical quantities (such as temperature changes, etc.), the data interpolation strategy is dynamically adjusted (for example, using linear interpolation or spline interpolation) to ensure that the smoothness of the data is not affected after time alignment; through the Granger causality test, the causal relationship between different data sources is identified to avoid the interference of irrelevant data sources on feature extraction. Using causal relationships, further construct Bayesian networks or structural equation models to quantify the dependencies between different data sources and guide the subsequent feature selection process; use environmental physical quantities (such as light intensity, etc.) to generate new features, such as combining light intensity and voltage data to generate coupling features. Use polynomial regression or factorization machine (FM) and other technologies to model the interaction effects between features and explore the nonlinear relationship between data sources; design a multimodal autoencoder to input different data sources into the shared encoding layer and learn cross-modal joint feature representation. Introduce an attention mechanism in the autoencoder to dynamically adjust the weights of different data sources, highlight key information, and suppress noise features; based on the correlation between environmental physical quantities (such as wind speed) and target variables (such as electric energy metering errors), screen out highly correlated features. Use Shapley value or LIME method to dynamically evaluate the importance of features, avoiding the limitations of static feature selection; use environmental physical quantities as fusion weights to weightedly fuse features from different data sources to generate a comprehensive feature vector. The information between the fused features and the target variables is evaluated through mutual information or conditional entropy to ensure that the fused features have high validity; the distribution consistency of the fused features under different environmental physical quantity conditions is verified through methods such as KL divergence or JS divergence to ensure that feature extraction does not destroy the physical association between data sources. At the same time, environmental physical quantities can be used as auxiliary indicators for model performance evaluation to ensure that the model is adaptable under different environmental conditions.

[0049] The effects of the above technical solutions are as follows: By introducing environmental physical quantities as a bridge, mapping data from different sources to a unified physical feature space can effectively reduce the differences between data sources, improve the consistency and comparability between different data sources; By methods such as non-negative matrix factorization or generative adversarial networks, learn the optimal mapping matrix between data sources to minimize the distribution differences of data sources in the feature space, making the feature extraction process more accurate and meaningful; By using the dynamic time warping (DTW) algorithm to process the alignment of time series data, solve the problem of inconsistent sampling frequencies of different data sources; And dynamically adjust the interpolation strategy according to the changes in environmental physical quantities to ensure that the data after time alignment has good smoothness and stability; Through Granger causality test, the causal relationship between different data sources can be identified to avoid interference from irrelevant data sources on feature extraction. At the same time, use Bayesian networks or structural equation models to quantify the dependence relationship between data sources, thereby guiding the feature selection process and enhancing the interpretability of the model; By combining environmental physical quantities with other data sources to generate new features and using methods such as polynomial regression or factorization machines, the non-linear relationship between data sources can be captured, thus enriching the feature space and improving the expressiveness of the model; Design a multi-modal autoencoder to input different data sources into the shared encoding layer to effectively learn the cross-modal joint feature representation. In addition, introducing an attention mechanism can dynamically allocate the weights of different data sources, highlight key features, reduce the influence of noise, and further improve the quality of feature selection; Through the Shapley value or LIME method, the importance of features under different environmental conditions can be dynamically evaluated, avoiding the limitations of traditional static feature selection methods. By screening out features highly correlated with the target variable according to the changes in environmental physical quantities, the model can better adapt to environmental changes; Use environmental physical quantities as fusion weights to weight and fuse the features of different data sources to generate a more comprehensive feature vector. In addition, use mutual information or conditional entropy to evaluate the information volume between the fused features and the target variable to ensure that the generated fused features have strong correlation and effectiveness; By verifying the distribution consistency of the fused features under different environmental physical quantity conditions, ensure that feature extraction does not destroy the physical association between data sources. Quantify the distribution differences of the fused features through KL divergence or JS divergence to further improve the robustness of the model under various environmental conditions; Use environmental physical quantities as an auxiliary evaluation index for model performance. When the environmental conditions change drastically, the model can have higher fault tolerance and adaptability, so as to better cope with the challenges in different actual environments.

[0050] An embodiment of the present invention, an electric energy metering error correction system based on data fusion, includes a memory, a processor, and a computer program stored on the memory and executable on the memory. The processor executes the program to implement the electric energy metering error correction method based on data fusion as described in any one of the above.

[0051] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A method for correcting power metering errors based on data fusion, characterized in that The method includes: S1. Collect multi-dimensional data and preprocess the collected multi-dimensional data; S2. Use a convolutional neural network to extract features from the operation data and mine the spatial features in the data; use a recurrent neural network to extract the temporal features of the secondary circuit and environmental data and establish the temporal correlation between the data; S3. Integrate the extracted spatial features and temporal features; conduct in-depth mining and analysis of the multi-dimensional data; S4. Train the built-in error correction model; evaluate and verify the trained error correction model; feedback the evaluation results to form a closed-loop feedback mechanism.

2. The method for correcting the power metering error based on data fusion according to claim 1, wherein The said S1 includes: S11. Collect multi-dimensional data at a preset time interval; S12. During the data collection process, check the integrity of the collected data; S13. Preprocess the collected data.

3. The method for correcting the power metering error based on data fusion according to claim 1, wherein The said S2 includes: S21. Slice the operation data of the device according to a certain time window, and perform convolution operations on the sliced data through multiple convolutional layers using convolutional kernels of different sizes to extract the spatial features in the data; S22. Add a pooling layer after the convolutional layer to reduce the dimension of the convolutional features while retaining the feature information; S23. Construct the secondary circuit and environmental data into sequence data in chronological order and process the sequence data through a recurrent unit; S24. Encode the features output by the recurrent unit and convert the temporal features into a vector representation of a fixed length.

4. The method for correcting the power metering error based on data fusion according to claim 3, wherein, The said S21 includes: Align the operation data of the device with the environmental data in a unified timestamp format; and use the timestamp interpolation method to linearly interpolate the missing timestamps; Calculate the autocorrelation coefficient of the operation data of each device to determine the time dependence range of the data; at the same time, divide multiple time windows to capture features of different time granularities; Calculate the information entropy of the data within each time window, introduce the environmental data as an auxiliary variable, and dynamically reduce the time window when the temperature change rate exceeds the threshold; Use convolutional kernels of different sizes to extract local, mesoscopic, and global spatial features respectively; introduce depthwise separable convolution in the convolutional layer to decompose the standard convolution into depth convolution and pointwise convolution; Embed a channel attention mechanism in the convolutional layer to assign different weights to the features of different channels and strengthen the extraction of key features; map the environmental data to the initial parameters of the convolutional kernel; Add a cross-modal feature interaction layer after the convolutional layer to integrate the convolutional features of the device operation data and the environmental data; Construct a physical quantity-window size joint optimization model, take the environmental data and the window size as joint variables, and determine the optimal window size through an optimization algorithm.

5. The method for correcting the power metering error based on data fusion according to claim 3, wherein The said S23 includes: Align the secondary circuit data with the environmental data according to the timestamp; Normalize the data with different dimensions and map them to a unified interval; use the Z-score standardization method to eliminate the skewness of the data distribution; Divide the time series into windows of a fixed length; on the basis of the fixed window, introduce a sliding window mechanism to dynamically adjust the window step size; capture the dependence relationships of different time granularities; Calculate the decay rate of the autocorrelation coefficient of the time series. When the autocorrelation decays to the threshold, determine the effective length of the current sequence and dynamically adjust the window size; Adopt a hybrid structure of the long short-term memory network LSTM and GRU to process short-term and long-term dependencies respectively; Introduce a self-attention mechanism into the recurrent unit to assign higher weights to key time points in the time series; Use the environmental data as an external input and jointly input it into the recurrent unit with the secondary circuit data; construct a physical quantity-time series coupling model to map temperature-related physical quantities and current and voltage data into a unified feature space; During the training process, dynamically monitor the gradient change of the recurrent unit. When the gradient vanishes or explodes, automatically truncate the overly long sequence; based on the reinforcement learning method, dynamically adjust the sequence length according to the model performance.

6. The method for correcting the power metering error based on data fusion according to claim 1, wherein The S3 includes: S31. Concatenate the spatial features and temporal features to form a comprehensive feature vector; S32. On the basis of feature concatenation, introduce an attention mechanism to perform weighted processing on the importance of different features and highlight the key features; S33. Design the number of network layers and neurons according to the features of the data and the task requirements; S34. Use the fused data to train the deep learning model; S35. During the training process, regularly evaluate the model and optimize the structure and parameters of the model according to the evaluation results.

7. The method for correcting the power metering error based on data fusion according to claim 6, characterized in that The S31 includes: S311. Preprocess the obtained spatial data and temporal data to obtain a spatial feature set and a temporal feature set; S312. Expand the dimension of the spatial feature set to obtain an enhanced vector of spatial features, and further extract the temporal feature set to obtain an enhanced vector of temporal features; S313. Based on different concatenation strategies, concatenate the enhanced vector of spatial features and the enhanced vector of temporal features respectively to form a preliminary comprehensive feature vector; S314. Normalize the formed preliminary comprehensive feature vector to obtain the final comprehensive feature vector; perform further feature analysis and processing according to the final comprehensive feature vector to obtain the final multi-modal feature set; S315. Based on the final multi-modal feature set, model and infer the spatio-temporal dependence relationship to obtain a spatio-temporal dependence relationship graph; based on the spatio-temporal dependence relationship graph, perform relationship inference and optimization to obtain the final relationship inference score.

8. The method for correcting the power metering error based on data fusion according to claim 6, wherein, The S313 includes: Represent the time information clockwise through time dependence and represent the timestamp counterclockwise; perform the first concatenation of the clockwise represented coordinate information and the counterclockwise represented timestamp based on the relationship between position and time to obtain the first layer of concatenation; The spatial distance is divided into X equal small segments through a reference point, where X > 2. The odd segments are represented in the positive direction, and the even segments are represented in the negative direction. The spatially-distanced segments with different representations are spliced to obtain a first spatial distance. The time interval is divided into N equal parts of equal length through a gradual transition function, where N > 2. The odd equal parts and the even equal parts are respectively spliced, and then the splicing results are spliced again to obtain a first time interval. Based on the relative relationship between positions and time changes, the first spatial distance and the first time interval are second-spliced to obtain a second-layer splicing. Based on the direction and change speed of the object's movement, the direction feature and the change rate are third-spliced to obtain a third-layer splicing. Extract the first 30% of the dimensions of the first-layer splicing and splice them with the second-layer splicing; extract the last 70% of the dimensions of the first-layer splicing and splice them with the third-layer splicing; splice the two spliced parts again to form a preliminary comprehensive feature vector.

9. The method for correcting the power metering error based on data fusion according to claim 1, wherein The S4 includes: S41. Train the built-in error correction model according to the depth data fusion result; S42. Evaluate and verify the trained error correction model; S43. Feedback the evaluation result to form a closed-loop feedback mechanism.

10. The electric energy metering error correction system based on data fusion is characterized in that It includes a memory, a processor, and a computer program stored on the memory and executable on the memory. The processor executes the program to implement the power metering error correction method based on data fusion according to any one of claims 1-9.

Images