Box multi-parameter prediction method and system integrated with big data processing

Abstract

The application relates to a meter box multi-parameter prediction method and system integrated with big data processing, which comprises the following steps: collecting multi-dimensional original data of a meter box in real time through an RTU acquisition module built in the meter box, uploading the original data to a cloud server after encryption, and performing cleaning, standardization and feature extraction to obtain structured data; storing the structured data to a distributed storage system and performing index and compression processing on the data; calling the structured data after compression processing to construct a multi-task prediction model, and performing supervised training on the multi-task prediction model by using historical statistical data to obtain a trained prediction model; inputting the original data collected in real time into the trained prediction model to output a single-parameter prediction result, and performing multi-parameter collaborative analysis based on the single-parameter prediction result and the original data to generate a collaborative prediction result.

Description

Technical Field

[0001] This application relates to the field of smart watch box technology, and in particular to a multi-parameter prediction method and system for watch boxes that integrates big data processing. Background Technology

[0002] Meter boxes are key equipment in the final distribution stage of the power system, undertaking the core functions of electricity metering, user electricity distribution, and security protection. Their operational status directly affects the safety and stability of the power grid, the reliability of user electricity consumption, and the efficiency of anti-theft management. With the in-depth advancement of smart grid construction and the continuous growth of electricity load, the number of meter boxes has surged and their distribution range has expanded. The traditional management model relying on manual inspection and single-parameter monitoring is no longer sufficient to meet the needs of modern power operation and maintenance.

[0003] Currently, meter boxes face problems such as overheating of live / neutral wires, abnormal current and voltage, unauthorized opening of the box door, unbalanced three-phase loads, and aging of the box. These issues can easily lead to short circuits, power outages, electricity theft, and equipment damage. Furthermore, manual inspections are time-consuming and cannot capture short-term faults in real time. Single-parameter monitoring within the box focuses only on local indicators, ignoring the combined effects of multiple parameters such as current, voltage, phase, and box door status, which can easily lead to misdiagnosis or missed faults.

[0004] Based on this, this application provides a multi-parameter prediction method and system for bins that integrates big data processing. Summary of the Invention

[0005] To address the issues of long manual inspection cycles, inability to capture short-term faults in real time, and the tendency for single-parameter monitoring within meter boxes to focus only on local indicators while ignoring the combined effects of multiple parameters such as current, voltage, phase, and door status, which can easily lead to misjudgment or missed fault detection, this application provides a multi-parameter prediction method and system for meter boxes that integrates big data processing.

[0006] Firstly, this application provides a multi-parameter prediction method for tables and bins that integrates big data processing, employing the following technical solution: including:

[0007] The RTU acquisition module built into the meter box collects multi-dimensional raw data of the meter box operation in real time. The raw data includes: live wire and neutral wire temperature, current and voltage, phase information, door opening and closing status, meter box location coordinates and meter measurement data.

[0008] The raw data is transmitted to an IoT device via a dedicated communication chip, encrypted, and then uploaded to a cloud server. The raw data received from the cloud is then cleaned, standardized, and its features are extracted to obtain structured data.

[0009] The structured data is stored in a distributed storage system, and the data is indexed and compressed. The distributed storage system uses a time-series database to store real-time monitoring data and a relational database to store device information and historical statistical data.

[0010] The structured data after compression is used to construct a multi-task prediction model. The multi-task prediction model includes an LSTM neural network for predicting short-term trends of temperature and current and voltage, an XGBoost model for identifying fault types, and a degradation model for assessing the aging degree of the meter box. The multi-task prediction model is trained under supervision using historical statistical data to obtain the trained prediction model.

[0011] The original data is collected in real time and input into the trained prediction model to output a single-parameter prediction result. Based on the single-parameter prediction result and the original data, a multi-parameter collaborative analysis is performed to generate a collaborative prediction result.

[0012] Preferably, the raw data is transmitted to an IoT device via a dedicated communication chip, encrypted, and then uploaded to a cloud server. The raw data received from the cloud is then cleaned, standardized, and its features are extracted to obtain structured data, including:

[0013] The raw data is transmitted to the IoT device via a dedicated communication chip built into the RTU, and the raw data transmitted to the IoT device is output. The transmission rate of the dedicated communication chip is matched with the sampling frequency of the raw data.

[0014] The original data is encrypted using a preset encryption algorithm, and the encrypted original data is uploaded to the cloud server via an Internet of Things device;

[0015] The original data is processed in the cloud to remove outliers and fill in missing values, resulting in cleaned preprocessed data.

[0016] The parameters of different dimensions in the preprocessed data are converted into a unified dimension, and standardized data is output. Then, time features and state features are extracted from the standardized data, and structured data is output.

[0017] Preferably, the step of storing the structured data in a distributed storage system and indexing and compressing the data includes storing real-time monitoring data in a time-series database, storing device information and historical statistical data in a relational database, and the following:

[0018] The structured data is allocated to different databases in the distributed storage system according to its data type: real-time monitoring data is allocated to a time-series database, and device information and historical statistical data are allocated to a relational database.

[0019] Based on the real-time monitoring data, downsampling and compression processing is performed on the high-frequency data to retain the peak, valley and timestamp key nodes of the data sequence, and lossless compression is performed on the low-frequency data to output the compressed real-time monitoring data.

[0020] Based on the device information and historical statistics data in the relational database, a data index is constructed. A primary key index for device ID is established for the device information table, and a composite index for timestamp and bin ID is established for the historical statistics data. Indexed information is then output.

[0021] Based on the compressed real-time monitoring data and the index information, the two types of data are associated and stored through the collaborative storage mechanism of the distributed storage system, and the structured data stored in the distributed storage system is output.

[0022] Preferably, the multi-task prediction model is constructed by calling the compressed structured data. The multi-task prediction model includes an LSTM neural network for predicting short-term trends in temperature and current / voltage, an XGBoost model for identifying fault types, and a degradation model for assessing the aging degree of the meter box. The multi-task prediction model is then trained under supervised training using historical statistical data to obtain a trained prediction model, including:

[0023] The historical statistical data is extracted from the distributed storage system and divided into a training dataset and a validation dataset according to a preset ratio. The training dataset must contain normal operation samples and failure samples.

[0024] Using the training dataset as input, an initial multi-task prediction model containing three collaborative sub-models is constructed, and supervised training is performed using a joint loss function;

[0025] Based on the pre-trained multi-task prediction model, the model performance is evaluated using the validation dataset: the trend prediction accuracy of the LSTM sub-model is verified, and the parameters of the sub-models that do not meet the standard are fine-tuned, and the validated trained prediction model is output.

[0026] Preferably, the step of constructing an initial multi-task prediction model containing three collaborative sub-models using the training dataset as input, and performing supervised training using a joint loss function, includes:

[0027] An LSTM neural network sub-model is constructed. The input layer receives real-time parameters and time features of current, voltage, and temperature. There are two hidden layers with 64 neurons in each hidden layer and three output neurons. The three neurons in the output layer correspond to the predicted trend values ​​of temperature, current, and voltage for the next hour, respectively.

[0028] An XGBoost sub-model is constructed. The input layer receives features of temperature, current, and door status, as well as fault labels. The mapping relationship between features and fault types is learned through 10 decision trees to achieve fault type identification.

[0029] A degradation model is constructed, and the input layer receives aging-related features such as temperature fluctuation frequency, humidity, and years of operation. The degradation parameters are used to fit the performance degradation curve of the meter box to assess the degree of aging.

[0030] An initial multi-task prediction model containing the above three collaborative sub-models is obtained, wherein each sub-model achieves data collaboration by sharing an input feature layer and is trained under supervision using a joint loss function.

[0031] Preferably, the initial multi-task prediction model obtained includes the above three collaborative sub-models, wherein each sub-model achieves data collaboration by sharing an input feature layer and is trained under supervision using a joint loss function, including:

[0032] For the LSTM sub-model, the mean square error between the predicted trend value and the actual value is used as the loss term. For the XGBoost sub-model, the cross-entropy loss function is used with the fault identification accuracy as the optimization objective. For the degradation model, the absolute error between the predicted aging degree value and the actual evaluation value is used as the loss term.

[0033] The Adam optimizer is used to jointly minimize the total loss function, iteratively optimize the parameters of each sub-model, and output the pre-trained multi-task prediction model. The parameters of each sub-model are shared and their performance is linked through collaborative training.

[0034] Preferably, the step of real-time acquisition of the raw data and inputting it into the trained prediction model to output a single-parameter prediction result, and performing multi-parameter collaborative analysis based on the single-parameter prediction result and the raw data to generate a collaborative prediction result, includes:

[0035] Based on the original data, the data is preprocessed into feature vectors compatible with the trained prediction model through format conversion and feature matching, and the preprocessed real-time input data is output.

[0036] Based on the real-time input data and the trained prediction model, the predicted values ​​of temperature, current and voltage trends, fault type probability distribution and aging degree are output by the LSTM neural network, XGBoost model and degradation model respectively, to obtain the single-parameter prediction results.

[0037] Based on the single-parameter prediction results and the original data, a collaborative analysis rule for trend, state and aging is constructed, and the collaborative analysis rule is output.

[0038] Based on the aforementioned collaborative analysis rules and the single-parameter prediction results, a collaborative prediction result is generated that includes short-term trend prediction, fault warning level, and aging maintenance recommendations.

[0039] Secondly, this application discloses a multi-parameter prediction device for a data box that integrates big data processing, employing the following technical solution:

[0040] The data acquisition module is used to collect multi-dimensional raw data of the meter box operation in real time through the RTU acquisition module built into the meter box. The raw data includes: live wire and neutral wire temperature, current and voltage, phase information, door opening and closing status, meter box location coordinates and meter measurement data.

[0041] The data processing module is used to transmit the raw data to the Internet of Things device through a dedicated communication chip, encrypt it and upload it to the cloud server, and clean, standardize and extract features from the raw data received from the cloud to obtain structured data.

[0042] The data storage module is used to store the structured data in a distributed storage system and to index and compress the data. The distributed storage system uses a time-series database to store real-time monitoring data and a relational database to store device information and historical statistical data.

[0043] The model training module is used to call the compressed structured data to construct a multi-task prediction model. The multi-task prediction model includes an LSTM neural network for predicting short-term trends of temperature and current voltage, an XGBoost model for identifying fault types, and a degradation model for assessing the aging degree of the meter box. The multi-task prediction model is trained under supervision using historical statistical data to obtain the trained prediction model.

[0044] The result prediction module is used to input the real-time collected raw data into the trained prediction model, output a single-parameter prediction result, and perform multi-parameter collaborative analysis based on the single-parameter prediction result and the raw data to generate a collaborative prediction result.

[0045] Thirdly, this application also provides a control device, the device comprising:

[0046] It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed, such as the bin multi-parameter prediction method for integrated big data processing described above.

[0047] Fourthly, this application also provides a computer-readable storage medium storing a computer program capable of being loaded by a processor and executing the multi-parameter prediction method for integrated big data processing as described above.

[0048] In summary, this application relies on RTUs to acquire multi-dimensional data such as live / neutral wire temperature, current, voltage, phase, and enclosure door status in real time. After encrypted transmission and preprocessing, efficient management is achieved through a distributed storage system. Simultaneously, a multi-task prediction model is constructed based on the stored data: an LSTM neural network predicts short-term trends in temperature, current, and voltage; XGBoost identifies fault types; and a degradation model assesses aging levels. Parameters are optimized through joint training. After real-time data is input into the model, multi-parameter collaborative analysis generates a collaborative result including trend prediction, fault warning, and maintenance suggestions. This overcomes the limitations of traditional single-parameter monitoring and delayed early warning, achieving real-time collaborative monitoring of multiple parameters, early trend prediction, and proactive early warning. Data-driven optimization of load scheduling, fault handling, and lifespan management improves the operational safety, maintenance efficiency, and power grid reliability of the meter enclosure. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a multi-parameter prediction method for a data bin that integrates big data processing.

[0050] Figure 2 This is a structural block diagram of a multi-parameter prediction device for a data processing bin. Detailed Implementation

[0051] The following combination Figure 1 - Figure 2 This application will be described in further detail.

[0052] This invention remotely monitors the operating status of existing meter boxes by collecting relevant data, downloading an APP software, and displaying the data on a network TV screen through a management center: (1) The device has a meter box location positioning and navigation function and a navigation module. (2) The device has a function to detect the temperature of the live wire and neutral wire of the meter inside the meter box. (3) The device has a function to detect the current and voltage of the meter inside the meter. (4) The device has a function to detect the opening and closing of the meter box door. (5) The device has an automatic phase recognition function. (6) The device has the function to read the meter data by connecting to the meter's infrared and R485 interfaces. By analyzing the metering data collected by the RTU and comparing it with the meter data, the metering accuracy of the meter can be analyzed and compared.

[0053] This device transmits real-time data changes on the temperature, current, voltage, position, and door status of the live and neutral wires within the meter box to an IoT communication device via a dedicated communication chip (ALLDGO). The IoT then sends this data to a cloud server, allowing for 24 / 7 monitoring and alerts via mobile phone or internet access to the management center's large screen. This application enables real-time monitoring of the meter box's open / closed status; real-time monitoring of conductor temperature changes; real-time monitoring of meter voltage fluctuations; and real-time monitoring of user current changes. It also sends SMS messages and makes phone calls to management personnel when the meter box is illegally opened / closed, or when the voltage is below or above normal, promptly detecting electricity theft and improving the timeliness and accuracy of meter box anti-theft measures. Furthermore, it allows for timely checks of user loads to prevent overheating, conductor overheating, insulation damage, phase-to-phase short circuits, fires, or power outages. After the meters are connected and energized in the meter box, the device identifies the phase of each meter and simultaneously adjusts the three-phase load balance to prevent neutral wire meltdowns, large-scale power outages, and equipment damage due to three-phase load imbalance. By using location positioning and navigation, we can reach the meter box installation location in a timely and accurate manner, enabling efficient maintenance, inspection, and examination.

[0054] Reference Figure 1 The embodiments of this application include at least steps S10 to S50.

[0055] The S10 uses the built-in RTU acquisition module to collect multi-dimensional raw data of the meter box operation in real time.

[0056] S20 transmits the raw data to the IoT device via a dedicated communication chip, encrypts it, and uploads it to the cloud server. The raw data received from the cloud is then cleaned, standardized, and its features are extracted to obtain structured data.

[0057] S30 stores structured data in a distributed storage system and performs indexing and compression on the data. The distributed storage system uses a time-series database to store real-time monitoring data and a relational database to store device information and historical statistical data.

[0058] S40: The compressed structured data is called to build a multi-task prediction model. The multi-task prediction model includes an LSTM neural network for predicting short-term trends of temperature, current and voltage, an XGBoost model for identifying fault types, and a degradation model for assessing the aging degree of the meter box. The multi-task prediction model is trained in a supervised manner using historical statistical data to obtain the trained prediction model.

[0059] The S50 will collect raw data in real time and input it into the trained prediction model, output single-parameter prediction results, and perform multi-parameter collaborative analysis based on the single-parameter prediction results and raw data to generate collaborative prediction results.

[0060] The raw data includes: live and neutral wire temperatures, current and voltage, phase information, door switch status, meter box location coordinates, and meter readings.

[0061] Specifically, the system collects multi-dimensional data such as live / neutral wire temperature, current, voltage, and door status through the RTU module to capture the real-time operating status of the meter box. The data is encrypted and uploaded to the cloud and preprocessed into structured data to ensure data security and usability, laying the foundation for subsequent analysis. Distributed storage using a time-series database and a relational database efficiently manages real-time monitoring data and equipment / historical data, supporting rapid queries (such as historical fault records for a specific meter box). A multi-task model is constructed: LSTM predicts temperature / current / pressure trends, XGBoost identifies fault types, and a degradation model assesses aging. After training, it possesses "prediction-diagnosis-evaluation" capabilities. Combining real-time data and model output, multi-parameter collaborative analysis (such as temperature trend + current value + door status) generates early warnings (such as overheating risk) and maintenance suggestions.

[0062] In addition, by collecting and analyzing data from individual meters, analyzing data from meter boxes, and analyzing the power consumption relationship of a single power supply point and three layers (meters, meter boxes, and transformers), we can accurately analyze power load and identify fault points. We can also locate specific faulty equipment in a timely manner through equipment information, so that staff can promptly carry out on-site repairs.

[0063] This enables a shift from passive inspection to proactive early warning, allowing for the early detection of issues such as electricity theft, overheating, and three-phase imbalance, thereby reducing power outages. Furthermore, data-driven optimization of maintenance plans reduces operational costs and ensures the safe and stable operation of end-point equipment in the power grid.

[0064] In some embodiments, step S10 specifically includes the following steps: transmitting the raw data to the IoT device via a dedicated communication chip built into the RTU, and outputting the raw data transmitted to the IoT device, wherein the transmission rate of the dedicated communication chip matches the sampling frequency of the raw data; encrypting the raw data using a preset encryption algorithm, and uploading the encrypted raw data to a cloud server via the IoT device; performing outlier removal and missing value filling on the raw data in the cloud to obtain cleaned preprocessed data; converting parameters of different dimensions in the preprocessed data into a unified dimension, outputting standardized data, and extracting time features and state features from the standardized data to output structured data.

[0065] Specifically, in the context of meter box maintenance, the above steps are the core hub connecting front-end data collection and back-end intelligent analysis. Their overall function is to achieve secure transmission, quality optimization, and structured empowerment of raw data, providing reliable input for subsequent predictive model training and maintenance decisions.

[0066] The technical features are reflected in three aspects: First, real-time adaptation. By dynamically matching the transmission rate of the dedicated communication chip with the sampling frequency of the original data (e.g., current 5 seconds / time, temperature 1 minute / time), high-frequency power parameters (e.g., current fluctuations) are uploaded without delay, meeting the timeliness requirements of fault early warning. Second, security protection. Pre-set encryption algorithms are used to encrypt the entire data transmission link, preventing the leakage or tampering of meter box operation data (e.g., power load, phase information), ensuring the security of power data. Third, data value-added processing. Cloud preprocessing repairs data noise by removing outliers (e.g., sensor fault jump data) and filling missing values ​​(time series interpolation), standardizing and unifying units (e.g., temperature ℃, current A) to eliminate parameter differences, and then extracting time features and state features (e.g., door opening frequency), transforming fragmented raw data into structured feature vectors, directly supporting the accurate training of subsequent multi-task prediction models.

[0067] In some embodiments, step S20 specifically includes the following steps: allocating structured data to different databases in the distributed storage system according to data type: allocating real-time monitoring data to a time-series database, and allocating device information and historical statistical data to a relational database; based on the real-time monitoring data, performing downsampling compression on the high-frequency acquisition data, retaining the peak, valley, and timestamp key nodes of the data sequence, performing lossless compression on the low-frequency data, and outputting the compressed real-time monitoring data; based on the device information and historical statistical data in the relational database, constructing a data index, establishing a primary key index for device ID in the device information table, establishing a joint index for timestamp and bin ID in the historical statistical data, and outputting the index information; based on the compressed real-time monitoring data and index information, storing the two types of data together through the collaborative storage mechanism of the distributed storage system, and outputting the structured data stored in the distributed storage system.

[0068] Specifically, data is allocated by type. Real-time monitoring data is stored in a time-series database to meet high-frequency writing and trend query requirements. Equipment information (box ID, model) and historical statistics (monthly fault records) are stored in a relational database to support structured queries. High-frequency data is downsampled and compressed to retain peak, valley, and timestamp key nodes, reducing storage usage. A "device ID primary key index" and a "timestamp + box ID joint index" are built for structured data to achieve traceability. The two types of data are associated through box ID. Historical data is extracted from the relational database to divide the training / validation sets. An LSTM (for predicting 1-hour temperature / flow / pressure trends), an XGBoost (for fault identification), and a degradation model (for aging assessment) are built and trained together to output a high-precision prediction model.

[0069] In some embodiments, step S30 specifically includes the following steps: extracting historical statistical data from the distributed storage system and dividing it into a training dataset and a validation dataset according to a preset ratio, wherein the training dataset must contain normal operation samples and failure samples; using the training dataset as input, constructing an initial multi-task prediction model containing three collaborative sub-models, and performing supervised training using a joint loss function; based on the initially trained multi-task prediction model, evaluating the model performance through the validation dataset: verifying the trend prediction accuracy of the LSTM sub-models, fine-tuning the parameters of sub-models that do not meet the standards, and outputting the validated trained prediction model.

[0070] Furthermore, step S30 also includes the following steps: constructing an LSTM neural network sub-model, where the input layer receives real-time parameters and time features of current, voltage, and temperature; two hidden layers with 64 neurons each; and three output layers with three neurons, each corresponding to the predicted trend values ​​of temperature, current, and voltage for the next hour; constructing an XGBoost sub-model, where the input layer receives features of temperature, current, and door status, as well as fault labels, and learns the mapping relationship between features and fault types through 10 decision trees to achieve fault type identification; constructing a degradation model, where the input layer receives aging-related features such as temperature fluctuation frequency, humidity, and years of operation, and assesses the degree of aging by fitting the meter box performance degradation curve through degradation parameters; and obtaining an initial multi-task prediction model containing the above three collaborative sub-models, where each sub-model achieves data collaboration by sharing an input feature layer and is trained under supervision using a joint loss function.

[0071] Specifically, historical statistical data is extracted from distributed storage and divided into training and validation sets containing normal and faulty samples according to a set ratio. This ensures that the model learns fault characteristics and normal patterns, avoiding overfitting. A collaborative sub-model is constructed, consisting of LSTM (predicting 1-hour trends for temperature / flow / pressure), XGBoost (identifying fault types), and a degradation model (assessing aging levels). Through shared feature layers (e.g., LSTM temperature trends provide time-series data for XGBoost overheat identification), this addresses the challenge of a single model simultaneously meeting the "prediction-diagnosis-assessment" operational needs. Parameters are fine-tuned based on validation set evaluation to ensure model accuracy. In application, multi-task prediction supports proactive operational decision-making: early trend warnings, accurate fault location, and maintenance plan development based on aging levels, reducing power outages.

[0072] In some embodiments, step S40 specifically includes the following steps: for the LSTM sub-model, the mean square error between the predicted trend value and the actual value is used as the loss term; for the XGBoost sub-model, the cross-entropy loss function is used as the optimization objective for fault identification accuracy; for the degradation model, the absolute error between the predicted aging degree value and the actual evaluation value is used as the loss term; by jointly minimizing the total loss function through the Adam optimizer, the parameters of each sub-model are iteratively optimized to output the pre-trained multi-task prediction model, wherein the parameters of each sub-model achieve feature sharing and performance linkage through collaborative training.

[0073] Specifically, the loss function design adapts to the needs of different scenarios: when the LSTM sub-model predicts temperature / current / voltage trends, it uses mean squared error (MSE) as the loss term to focus on numerical deviations and avoid overheating warning delays caused by trend shifts; when the XGBoost sub-model identifies fault types, it uses cross-entropy loss to optimize accuracy and reduce "false alarms of normal fluctuations leading to faults" or "missed reports of latent faults" in operation and maintenance; when the degradation model assesses the degree of aging, it uses absolute error as the loss term to ensure quantification of deviations and support differentiated maintenance plans.

[0074] The optimization process relies on the Adam optimizer to jointly minimize the total loss. By iteratively adjusting the parameters of each sub-model (such as the LSTM weight matrix and XGBoost tree node thresholds), the linkage of "trend-fault-aging" features is achieved: when the LSTM predicts that the temperature continues to rise, XGBoost automatically increases the weight of "overheating fault" identification; when the degradation model outputs a high aging score, the LSTM becomes more sensitive to temperature fluctuations, avoiding misjudgment of trends due to the performance degradation of aging equipment.

[0075] In some embodiments, step S50 specifically includes the following steps: based on the original data, preprocessing it into feature vectors compatible with the trained prediction model through format conversion and feature matching, and outputting the preprocessed real-time input data; based on the real-time input data and the trained prediction model, outputting temperature, current, and voltage trend prediction values, fault type probability distribution, and aging degree prediction values ​​respectively through LSTM neural network, XGBoost model, and degradation model to obtain single-parameter prediction results; based on the single-parameter prediction results and the original data, constructing trend, state, and aging collaborative analysis rules, and outputting collaborative analysis rules; based on the collaborative analysis rules and the single-parameter prediction results, integrating and generating collaborative prediction results that include short-term trend prediction, fault warning level, and aging maintenance suggestions.

[0076] Specifically, data preprocessing solves the problem of data format differences between different meter boxes by converting and matching formats (converting real-time parameters such as current and temperature into 16-dimensional feature vectors), ensuring that real-time input is compatible with the trained model and laying the foundation for subsequent prediction accuracy. In operation and maintenance, this step avoids model output deviation due to inconsistent data formats (misjudgment caused by confusion of temperature units).

[0077] Single-parameter prediction relies on a multi-task model after training. LSTM outputs an hourly trend, XGBoost provides a fault probability distribution, and the degradation model assesses the degree of aging, but a single parameter is difficult to support decision-making.

[0078] Collaborative rule construction overcomes the limitations of isolated data through multi-dimensional correlation: if the LSTM predicts a temperature trend >70℃, the XGBoost overheat probability >80%, and the real-time current >120% of the rated value, a high-risk overheat rule is triggered; if the XGBoost electricity theft probability >75% and the real-time unauthorized door opening >30 seconds, suspected electricity theft is identified. These rules link single-parameter predictions with real-time status, avoiding false alarms that only warn of high temperatures.

[0079] The integrated output includes short-term trend curves, fault warning levels, and aging maintenance recommendations, directly supporting proactive prevention and control in operations and maintenance.

[0080] The implementation principle of the multi-parameter prediction method for meter boxes integrating big data processing in this application embodiment is as follows: This application relies on RTU to acquire multi-dimensional data such as live / neutral wire temperature, current, voltage, phase, and box door status in real time. After encrypted transmission and preprocessing, it achieves efficient management through a distributed storage system. At the same time, a multi-task prediction model is constructed based on the stored data: LSTM neural network predicts short-term trends of temperature / current / voltage, XGBoost identifies fault types, and a degradation model assesses the degree of aging. Parameters are optimized through joint training. After real-time data is input into the model, multi-parameter collaborative analysis generates a collaborative result including trend prediction, fault warning, and maintenance suggestions. This breaks through the limitations of traditional single-parameter monitoring and delayed early warning, realizing real-time collaborative monitoring of multiple parameters, early trend prediction, and proactive early warning. Through data-driven optimization of load scheduling, fault handling, and lifespan management, it improves the operational safety, maintenance efficiency, and power grid reliability of the meter boxes.

[0081] Figure 1 This is a flowchart illustrating a multi-parameter prediction method for a data bin that integrates big data processing in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows; unless explicitly stated otherwise, there is no strict order requirement for the execution of these steps, and they can be executed in other orders; and Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0082] Based on the same technical concept, referring to Figure 2 This application also provides a multi-parameter prediction device for a data box that integrates big data processing, employing the following technical solution: The device includes:

[0083] The data acquisition module is used to collect multi-dimensional raw data of the meter box in real time through the built-in RTU acquisition module. The raw data includes: live wire and neutral wire temperature, current and voltage, phase information, door opening and closing status, meter box location coordinates and meter measurement data.

[0084] The data processing module is used to transmit raw data to IoT devices via a dedicated communication chip, encrypt the data, upload it to the cloud server, and clean, standardize, and extract features from the raw data received from the cloud to obtain structured data.

[0085] The data storage module is used to store structured data in a distributed storage system and to index and compress the data. The distributed storage system uses a time-series database to store real-time monitoring data and a relational database to store device information and historical statistical data.

[0086] The model training module is used to call the compressed structured data to build a multi-task prediction model. The multi-task prediction model includes an LSTM neural network for predicting short-term trends of temperature, current and voltage, an XGBoost model for identifying fault types, and a degradation model for assessing the aging degree of the meter box. The multi-task prediction model is trained in a supervised manner using historical statistical data to obtain the trained prediction model.

[0087] The result prediction module is used to input the real-time collected raw data into the trained prediction model, output single-parameter prediction results, and perform multi-parameter collaborative analysis based on the single-parameter prediction results and the raw data to generate collaborative prediction results.

[0088] In some embodiments, the data acquisition module is specifically used to transmit raw data to an IoT device via a dedicated communication chip built into the RTU, and output the raw data transmitted to the IoT device, wherein the transmission rate of the dedicated communication chip is matched with the sampling frequency of the raw data.

[0089] The original data is encrypted using a preset encryption algorithm, and the encrypted original data is uploaded to the cloud server via an IoT device.

[0090] Outlier removal and missing value filling are performed on the raw data in the cloud to obtain cleaned preprocessed data.

[0091] The parameters of different dimensions in the preprocessed data are converted into a unified dimension, and standardized data is output. Then, time features and state features are extracted from the standardized data to output structured data.

[0092] In some embodiments, the data processing module is specifically used to allocate structured data to different databases in a distributed storage system according to the data type: real-time monitoring data is allocated to a time-series database, and device information and historical statistical data are allocated to a relational database;

[0093] Based on real-time monitoring data, downsampling and compression processing is performed on the high-frequency data to retain the peak, valley and timestamp key nodes of the data sequence, while lossless compression is performed on the low-frequency data to output the compressed real-time monitoring data.

[0094] Based on the equipment information and historical statistics of a relational database, a data index is constructed. A primary key index for the equipment ID is created in the equipment information table, and a composite index for the timestamp and bin ID is created in the historical statistics table. The index information is then output.

[0095] Based on compressed real-time monitoring data and index information, the two types of data are associated and stored through the collaborative storage mechanism of the distributed storage system, and the structured data stored in the distributed storage system is output.

[0096] In some embodiments, the data storage module is specifically used to extract historical statistical data from the distributed storage system and divide it into a training dataset and a validation dataset according to a preset ratio, wherein the training dataset must contain normal operation samples and fault samples.

[0097] Using the training dataset as input, an initial multi-task prediction model containing three collaborative sub-models is constructed, and supervised training is performed using a joint loss function;

[0098] Based on the initially trained multi-task prediction model, the model performance is evaluated using a validation dataset: the trend prediction accuracy of the LSTM sub-model is verified, and the parameters of the sub-models that do not meet the standard are fine-tuned, outputting the validated trained prediction model.

[0099] In some embodiments, the model training module is specifically used to construct an LSTM neural network sub-model. The input layer receives real-time parameters and time features of current, voltage, and temperature. There are two hidden layers with 64 neurons in each hidden layer and three output neurons. The three neurons in the output layer correspond to the predicted trend values ​​of temperature, current, and voltage for the next hour, respectively.

[0100] An XGBoost sub-model is constructed. The input layer receives features of temperature, current, and door status, as well as fault labels. The mapping relationship between features and fault types is learned through 10 decision trees to achieve fault type identification.

[0101] A degradation model is constructed, and the input layer receives aging-related features such as temperature fluctuation frequency, humidity, and years of operation. The degradation parameters are used to fit the performance degradation curve of the meter box to assess the degree of aging.

[0102] An initial multi-task prediction model containing the above three collaborative sub-models is obtained, wherein each sub-model achieves data collaboration by sharing an input feature layer and is trained under supervision using a joint loss function.

[0103] In some embodiments, the model training module is specifically used for the LSTM sub-model, with the mean square error between the predicted trend value and the actual value as the loss term; for the XGBoost sub-model, with the cross-entropy loss function as the optimization objective of fault identification accuracy; and for the degradation model, with the absolute error between the predicted aging degree value and the actual evaluation value as the loss term.

[0104] By jointly minimizing the total loss function using the Adam optimizer, the parameters of each sub-model are iteratively optimized, and a pre-trained multi-task prediction model is output. The parameters of each sub-model are collaboratively trained to achieve feature sharing and performance linkage.

[0105] In some embodiments, the result prediction module is specifically used to preprocess the original data into a feature vector compatible with the trained prediction model through format conversion and feature matching, and output the preprocessed real-time input data.

[0106] Based on real-time input data and trained prediction models, the predicted values ​​of temperature, current and voltage trends, fault type probability distribution and aging degree are output by LSTM neural network, XGBoost model and degradation model respectively, to obtain single-parameter prediction results.

[0107] Based on the single-parameter prediction results and raw data, we construct a collaborative analysis rule for trend, state and aging, and output the collaborative analysis rule.

[0108] Based on collaborative analysis rules and single-parameter prediction results, a collaborative prediction result is generated that includes short-term trend prediction, fault warning level, and aging maintenance recommendations.

[0109] This application also discloses a control device.

[0110] Specifically, the control device includes a memory and a processor, the memory storing a computer program that can be loaded by the processor and executed to perform the aforementioned integrated big data processing multi-parameter prediction method for tables.

[0111] This application also discloses a computer-readable storage medium.

[0112] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed as described above for the multi-parameter prediction method of integrated big data processing. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A multi-parameter prediction method for a data bin integrating big data processing, characterized in that, include: The RTU acquisition module built into the meter box collects multi-dimensional raw data of the meter box operation in real time. The raw data includes: live wire and neutral wire temperature, current and voltage, phase information, door opening and closing status, meter box location coordinates and meter measurement data. The raw data is transmitted to an IoT device via a dedicated communication chip, encrypted, and then uploaded to a cloud server. The raw data received from the cloud is then cleaned, standardized, and its features are extracted to obtain structured data. The structured data is stored in a distributed storage system, and the data is indexed and compressed. The distributed storage system uses a time-series database to store real-time monitoring data and a relational database to store device information and historical statistical data. The structured data after compression is used to construct a multi-task prediction model. The multi-task prediction model includes an LSTM neural network for predicting short-term trends of temperature and current and voltage, an XGBoost model for identifying fault types, and a degradation model for assessing the aging degree of the meter box. The multi-task prediction model is trained under supervision using historical statistical data to obtain the trained prediction model. The raw data collected in real time is input into the trained prediction model, which outputs a single-parameter prediction result. Based on the single-parameter prediction result and the raw data, a multi-parameter collaborative analysis is performed to generate a collaborative prediction result. The process involves using the compressed structured data to construct a multi-task prediction model. This model includes an LSTM neural network for predicting short-term trends in temperature and current / voltage, an XGBoost model for identifying fault types, and a degradation model for assessing the aging degree of the meter enclosure. The multi-task prediction model is then trained under supervised supervision using historical statistical data to obtain the trained prediction model, which includes: The historical statistical data is extracted from the distributed storage system and divided into a training dataset and a validation dataset according to a preset ratio. The training dataset must contain normal operation samples and failure samples. Using the training dataset as input, an initial multi-task prediction model containing three collaborative sub-models is constructed, and supervised training is performed using a joint loss function; Based on the pre-trained multi-task prediction model, the model performance is evaluated using the validation dataset: the trend prediction accuracy of the LSTM sub-model is verified, and the parameters of the sub-models that do not meet the standard are fine-tuned, and the trained prediction model that passes the validation is output. The step of constructing an initial multi-task prediction model containing three collaborative sub-models using the training dataset as input, and conducting supervised training using a joint loss function, includes: An LSTM neural network sub-model is constructed. The input layer receives real-time parameters and time features of current, voltage, and temperature. There are two hidden layers with 64 neurons in each hidden layer and three output neurons. The three neurons in the output layer correspond to the predicted trend values ​​of temperature, current, and voltage for the next hour, respectively. An XGBoost sub-model is constructed. The input layer receives features of temperature, current, and door status, as well as fault labels. The mapping relationship between features and fault types is learned through 10 decision trees to achieve fault type identification. A degradation model is constructed, and the input layer receives aging-related features such as temperature fluctuation frequency, humidity, and years of operation. The degradation parameters are used to fit the performance degradation curve of the meter box to assess the degree of aging. An initial multi-task prediction model containing the above three collaborative sub-models is obtained, wherein each sub-model achieves data collaboration by sharing an input feature layer and is trained under supervision using a joint loss function; The process of collecting raw data in real time and inputting it into the trained prediction model, outputting single-parameter prediction results, and performing multi-parameter collaborative analysis based on the single-parameter prediction results and the raw data to generate collaborative prediction results includes: Based on the original data, the data is preprocessed into feature vectors compatible with the trained prediction model through format conversion and feature matching, and the preprocessed real-time input data is output. Based on the real-time input data and the trained prediction model, the predicted values ​​of temperature, current and voltage trends, fault type probability distribution and aging degree are output by the LSTM neural network, XGBoost model and degradation model respectively, to obtain the single-parameter prediction results. Based on the single-parameter prediction results and the original data, a collaborative analysis rule for trend, state and aging is constructed, and the collaborative analysis rule is output. Based on the aforementioned collaborative analysis rules and the single-parameter prediction results, a collaborative prediction result is generated that includes short-term trend prediction, fault warning level, and aging maintenance recommendations.

2. The multi-parameter prediction method for integrated big data processing according to claim 1, characterized in that, The raw data is transmitted to an IoT device via a dedicated communication chip, encrypted, and then uploaded to a cloud server. The raw data received from the cloud is then cleaned, standardized, and its features are extracted to obtain structured data, including: The raw data is transmitted to the IoT device via a dedicated communication chip built into the RTU, and the raw data transmitted to the IoT device is output. The transmission rate of the dedicated communication chip is matched with the sampling frequency of the raw data. The original data is encrypted using a preset encryption algorithm, and the encrypted original data is uploaded to the cloud server via an Internet of Things device; The original data is processed in the cloud to remove outliers and fill in missing values, resulting in cleaned preprocessed data. The parameters of different dimensions in the preprocessed data are converted into a unified dimension, and standardized data is output. Then, time features and state features are extracted from the standardized data, and structured data is output.

3. The multi-parameter prediction method for integrated big data processing according to claim 2, characterized in that, The structured data is stored in a distributed storage system, and the data is indexed and compressed. The distributed storage system uses a time-series database to store real-time monitoring data and a relational database to store device information and historical statistical data, including: The structured data is allocated to different databases in the distributed storage system according to its data type: real-time monitoring data is allocated to a time-series database, and device information and historical statistical data are allocated to a relational database. Based on the real-time monitoring data, downsampling and compression processing is performed on the high-frequency data to retain the peak, valley and timestamp key nodes of the data sequence, and lossless compression is performed on the low-frequency data to output the compressed real-time monitoring data. Based on the device information and historical statistics in the relational database, a data index is constructed. A primary key index of device ID is established for the device information, and a composite index of timestamp and bin ID is established for the historical statistics. Indexed information is then output. Based on the compressed real-time monitoring data and the index information, the two types of data are associated and stored through the collaborative storage mechanism of the distributed storage system, and the structured data stored in the distributed storage system is output.

4. The multi-parameter prediction method for integrated big data processing according to claim 3, characterized in that, The resulting initial multi-task prediction model comprises the three collaborative sub-models described above. Each sub-model achieves data collaboration by sharing an input feature layer and undergoes supervised training using a joint loss function, including: For the LSTM sub-model, the mean square error between the predicted trend value and the actual value is used as the loss term. For the XGBoost sub-model, the cross-entropy loss function is used with the fault identification accuracy as the optimization objective. For the degradation model, the absolute error between the predicted aging degree value and the actual evaluation value is used as the loss term. The Adam optimizer is used to jointly minimize the total loss function, iteratively optimize the parameters of each sub-model, and output the pre-trained multi-task prediction model. The parameters of each sub-model are shared and their performance is linked through collaborative training.

5. A multi-parameter prediction device for a data box integrating big data processing, characterized in that, The device includes: The data acquisition module is used to collect multi-dimensional raw data of the meter box operation in real time through the RTU acquisition module built into the meter box. The raw data includes: live wire and neutral wire temperature, current and voltage, phase information, door opening and closing status, meter box location coordinates and meter measurement data. The data processing module is used to transmit the raw data to the Internet of Things device through a dedicated communication chip, encrypt it and upload it to the cloud server, and clean, standardize and extract features from the raw data received from the cloud to obtain structured data. The data storage module is used to store the structured data in a distributed storage system and to index and compress the data. The distributed storage system uses a time-series database to store real-time monitoring data and a relational database to store device information and historical statistical data. The model training module is used to call the compressed structured data to construct a multi-task prediction model. The multi-task prediction model includes an LSTM neural network for predicting short-term trends of temperature and current voltage, an XGBoost model for identifying fault types, and a degradation model for assessing the aging degree of the meter box. The multi-task prediction model is trained under supervision using historical statistical data to obtain the trained prediction model. The result prediction module is used to input the real-time collected raw data into the trained prediction model, output a single-parameter prediction result, and perform multi-parameter collaborative analysis based on the single-parameter prediction result and the raw data to generate a collaborative prediction result. The model training module is specifically used to extract the historical statistical data from the distributed storage system and divide it into a training dataset and a validation dataset according to a preset ratio. The training dataset must include normal operation samples and failure samples. Using the training dataset as input, an initial multi-task prediction model containing three collaborative sub-models is constructed, and supervised training is performed using a joint loss function; Based on the pre-trained multi-task prediction model, the model performance is evaluated using the validation dataset: the trend prediction accuracy of the LSTM sub-model is verified, and the parameters of the sub-models that do not meet the standard are fine-tuned, and the trained prediction model that passes the validation is output. The model training module is also used to construct an LSTM neural network sub-model. The input layer receives real-time parameters and time features of current, voltage, and temperature. There are two hidden layers with 64 neurons in each hidden layer and three output neurons. The three neurons in the output layer correspond to the predicted trend values ​​of temperature, current, and voltage for the next hour, respectively. An XGBoost sub-model is constructed. The input layer receives features of temperature, current, and door status, as well as fault labels. The mapping relationship between features and fault types is learned through 10 decision trees to achieve fault type identification. A degradation model is constructed, and the input layer receives aging-related features such as temperature fluctuation frequency, humidity, and years of operation. The degradation parameters are used to fit the performance degradation curve of the meter box to assess the degree of aging. An initial multi-task prediction model containing the above three collaborative sub-models is obtained, wherein each sub-model achieves data collaboration by sharing an input feature layer and is trained under supervision using a joint loss function; Specifically, the result prediction module is used to preprocess the original data into a feature vector compatible with the trained prediction model through format conversion and feature matching, and output the preprocessed real-time input data. Based on the real-time input data and the trained prediction model, the predicted values ​​of temperature, current and voltage trends, fault type probability distribution and aging degree are output by the LSTM neural network, XGBoost model and degradation model respectively, to obtain the single-parameter prediction results. Based on the single-parameter prediction results and the original data, a collaborative analysis rule for trend, state and aging is constructed, and the collaborative analysis rule is output. Based on the aforementioned collaborative analysis rules and the single-parameter prediction results, a collaborative prediction result is generated that includes short-term trend prediction, fault warning level, and aging maintenance recommendations.

6. A control device, characterized in that, The device includes: It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 4.

Images