Method and System for Activating Distribution Terminal Batteries

Abstract

This invention relates to a method and system for activating batteries in distribution terminals, belonging to the field of distribution automation technology. The system is deployed at a distribution automation master station and includes modules for activation strategy management, task execution and monitoring, battery health status assessment, historical record query and archiving, and multiple anti-error verification. The method involves remotely configuring and issuing activation tasks through the master station, monitoring the voltage sequence during battery activation in real time, and recording the voltage U1 before activation, U2 after discharge, and U3 after charging. It constructs a multi-source feature system covering voltage timing, inherent battery characteristics, operating environment, and historical performance, and uses a deep learning model for dynamic assessment and anomaly detection of battery health status. After the task is completed, activation records are automatically generated and archived. This invention achieves remote, accurate, and safe activation testing and intelligent operation and maintenance management of massive numbers of distribution terminal batteries, effectively improving the reliability of power distribution networks.

Description

Technical Field

[0001] This invention belongs to the field of power distribution automation technology, specifically relating to a method and system for testing the activation of batteries in power distribution terminals. Background Technology

[0002] The power distribution network deploys a large number of distribution terminals, and the health status of their backup power supplies directly affects the emergency power supply capability and grid reliability in the event of a fault. Currently, the maintenance of terminal batteries mainly relies on periodic on-site measurements or simple remote voltage monitoring. The measured data lacks visualization, data management, and comparability, making it difficult to trace and analyze maintenance data, and lacking effective remote in-depth diagnostics and activation methods. On-site battery testing is complicated by the complexity of manual operations, inconsistent methods leading to different results, disconnection and reconnection issues, and time constraints. Furthermore, high-power discharge testing equipment cannot be carried to the field, resulting in limited on-site testing capabilities and an inability to quickly determine the battery's health status.

[0003] Many remote activation solutions rely solely on single criteria such as voltage drop rate, failing to comprehensively consider multi-dimensional information such as battery internal resistance, temperature, historical performance, SOC, and SOH. This results in insufficient accuracy in assessment and a risk of false activation. Furthermore, the diverse models and operating environments of field terminals necessitate a comprehensive, intelligent, and safe activation testing and management system for distribution terminal batteries. Improving battery operation and maintenance efficiency and the accuracy of status assessment has become an urgent need for intelligent operation and maintenance of distribution networks. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for activating batteries in power distribution terminals, so as to realize accurate, safe, intelligent remote activation testing and health status assessment and management of a large number of power distribution terminal batteries.

[0005] To achieve the above objectives, the technical solution of the present invention is: a method for testing the activation of a power distribution terminal battery, comprising:

[0006] The terminal battery voltage data is sent to the master station, which monitors the battery voltage data and controls the battery activation soft pressure plate.

[0007] Configure battery activation tasks, including creating target terminal activation tasks, setting activation strategies, and batch sending tasks to power distribution terminals;

[0008] Perform battery activation tasks, monitor the real-time voltage of the battery in the power distribution terminal, record the voltage U1 before activation, the voltage U2 after discharge, and the voltage U3 after charging, and monitor the status of the activation soft plate and the activation time at the same time.

[0009] Based on real-time voltage sequences and multi-source feature data, a deep learning model is used to assess battery health status, which includes the determination of abnormal battery activation.

[0010] After the task is completed, an activation history is generated and archived;

[0011] Perform error prevention checks during task execution;

[0012] The multi-source feature data includes voltage time-series features, inherent battery features, operating environment features, historical performance features, charge and discharge parameter features, real-time status features, charge and discharge cycle features, battery material characteristics, and grid fluctuation correlation features.

[0013] Furthermore, during the data acquisition process of multi-source feature data, environmental data needs to be corrected using a comprehensive environmental correction coefficient. , ,in:

[0014] This is a temperature correction function, based on the ambient temperature. The interval to be selected is a piecewise function, including:

[0015] Low temperature range ≤0℃): Correction function ;

[0016] Normal temperature range, i.e., 0℃ < ≤ 35℃): Correction function ;

[0017] High temperature range > 35℃: Correction function ;

[0018] This is a humidity correction function, based on the ambient humidity. Choose a piecewise function, including:

[0019] Low humidity range Correction function ;

[0020] medium humidity range Correction function ;

[0021] High humidity range Correction function ;

[0022] This is the barometric pressure correction function. , The actual ambient pressure is given; 101325 Pa is the standard atmospheric pressure.

[0023] Furthermore, the data acquisition process for multi-source feature data also includes correction for battery type and terminal model, using a battery type correction function f(type) and a terminal model correction coefficient. ,in:

[0024] f(type) takes a value based on the specific battery type, including: 1.0 for valve-regulated lead-acid batteries, 1.1 for gel lead-acid batteries, 1.2 for ternary lithium-ion batteries, and 1.3 for lithium iron phosphate batteries;

[0025] ,in For standard terminal power management efficiency, This refers to the power management efficiency of the current terminal model.

[0026] Furthermore, after the multi-source feature data acquisition is completed, data preprocessing is required. Data preprocessing operations include:

[0027] For missing data, interpolation is performed using the K-nearest neighbor interpolation method. The interpolation formula is as follows:

[0028]

[0029] in , The feature distance between the sample to be interpolated and its j-th neighbor sample is calculated using Euclidean distance: , For feature dimension, For the sample to be interpolated 3D eigenvalues For the first The neighbor sample number 3D eigenvalues;

[0030] And outlier detection is performed using the 3σ criterion and box plot method, where:

[0031] The outlier determination criteria of the 3σ criterion are: ,in The characteristic mean, The characteristic standard deviation;

[0032] The outlier detection criteria for box plot method are: or ,in The lower quartile, It is the upper quartile. .

[0033] Furthermore, after the multi-source feature data acquisition and preprocessing are completed, feature engineering optimization is required, including derived feature calculation and feature selection. Derived feature calculation includes:

[0034] The formula for calculating the sliding window mean of a voltage series is as follows:

[0035]

[0036] in, for Mean of sliding window over time (unit: V) The current time (in seconds). Index of time within the window (value) -4 to ), for Real-time voltage (unit: V) at any given moment, with a fixed window size of 5 consecutive sampling points, used to smooth voltage time series data.

[0037] The formula for calculating the rate of change of the sliding window slope is:

[0038]

[0039] in express The rate of change of the voltage slope of the sliding window at any given time reflects the trend of voltage change (rising, falling, or remaining stable) within that window. express Real-time voltage value at any given moment (unit: V); express Voltage value (unit: V) four time units prior to the time point. It represents the time span of the sliding window (fixed to 4 units of time, such as 4 seconds), used to normalize the voltage change, so that the slope change rate has a uniform dimension (unit: V / unit time).

[0040] The time-domain voltage sequence is converted into a frequency-domain signal using the Fast Fourier Transform (FFT), as shown in the following formula:

[0041]

[0042] in, It is a time-domain voltage sequence. For sequence length, It is a frequency domain signal;

[0043] Calculate the skewness, kurtosis, and autocorrelation coefficient of the voltage series, where:

[0044] The skewness formula is:

[0045]

[0046] The kurtosis formula is:

[0047]

[0048] The formula for the autocorrelation coefficient with a lag of k steps is:

[0049]

[0050] in, For voltage sequence data, The mean, Standard deviation, The sequence length;

[0051] Feature selection is performed using mutual information (MI). The formula for mutual information is:

[0052]

[0053] in, As characteristic variables, For the target variable, For joint probability distribution, , These are the marginal probability distributions.

[0054] Furthermore, the training of the deep learning model includes:

[0055] 1) Construct a training dataset by collecting historical data on battery activation of different models, operating years, and operating environments of power distribution terminals for more than 3 years, and collecting valid samples. Each sample contains complete characteristic data of the battery activation process of the power distribution terminal, dynamic threshold labels, and abnormal type labels. The valid samples are randomly divided into training dataset and test dataset in an 8:2 ratio.

[0056] Dynamic threshold label The calculation formula is:

[0057]

[0058] in, As a baseline threshold based on battery type, The correction threshold is based on ambient temperature. The correction threshold is based on charge and discharge parameters. , , Let be the weighting coefficient, satisfying + + =1, the weight value is determined through grid search optimization;

[0059] Anomaly type labels: These are manually labeled based on the activation process log and on-site verification results.

[0060] 2) Configure training parameters, including:

[0061] Learning rate decay formula:

[0062]

[0063] in =0.001 is the initial learning rate. =0.9 is the attenuation coefficient. For training rounds;

[0064] L2 regularization loss term: ,in =0.001 is the weight decay coefficient. These are the model weight parameters;

[0065] 3) Model Training: Input the training dataset; initialize the dual model parameters. Use the He initialization method to initialize the weights of the fully connected layers, and randomly initialize the weights of the LSTM layers. The He initialization formula is: ,in The number of neurons is input to the fully connected layer; iterative training is performed, and dynamic threshold predictions and anomaly type probability distributions are obtained based on the training dataset. The total loss is calculated according to the joint loss function, and the model parameters are updated using gradient descent to obtain the trained model; the performance metrics of the trained model are calculated using the test dataset to validate the model; the model with the best performance on the validation set is saved as the final training model.

[0066] Furthermore, the determination of abnormal battery activation includes:

[0067] The formula for determining abnormalities during the discharge process is as follows:

[0068]

[0069] Where Ut is the real-time voltage at time t during the discharge process. The dynamic deviation threshold for the discharge phase output by the model; The discharge correction factor is calculated using the following formula:

[0070]

[0071] in , , Let be the weighting coefficient, satisfying + + =1, This is a battery type correction function. For ambient temperature correction function, =1+0.01×(25− ), For ambient temperature, This is the discharge current correction function. =1+0.005×( - ), Standard discharge current, This is the real-time discharge current;

[0072] The formula for determining abnormalities during the charging process is as follows:

[0073]

[0074] in This is the dynamic deviation threshold for the discharge phase output by the model. The charging correction factor is calculated using the following formula:

[0075]

[0076] in , , Let be the weighting coefficient, satisfying + + =1, This is a function to correct the number of iterations. , This indicates the number of charge-discharge cycles (unit: times) of the battery in the power distribution terminal, reflecting the degree of battery wear and tear, and is obtained from the battery's operating history. For charging time correction function, , The actual duration of this charging process (in minutes) is recorded in real time by the task execution and monitoring module. Standard charging time, This is the correction function for the rate of change of internal resistance. , , For real-time internal resistance, This is the reference value for internal resistance.

[0077] Furthermore, the generation of activation history records includes calculating the activation duration:

[0078] For activation tasks that are completed normally, the activation duration formula is:

[0079]

[0080] in = - , For discharge start-up time, This is the discharge end time. = - . For charging start-up time, This is the charging end time;

[0081] For activation tasks that terminate abnormally, the activation duration formula is:

[0082] =

[0083] in This represents the runtime before the exception occurred.

[0084] Furthermore, the error prevention verification includes:

[0085] Activation is disabled when the battery voltage is 0.

[0086] The battery activation soft pressure plate is in a state where it is not included in this activation process when it is put into operation.

[0087] Remote control is prohibited when the battery activation soft pressure plate switching point number is not 0.

[0088] The present invention also provides a battery activation testing system for a power distribution terminal, comprising:

[0089] Activation Strategy Management Module: Used to configure battery activation strategies, including setting charge and discharge duration and charge and discharge voltage deviation ratio, and supports batch filtering of target terminals, creation of activation tasks, one-click batch distribution of tasks to power distribution terminals, and export of task list.

[0090] Task execution and monitoring module: It communicates with the activation strategy management module to receive activation tasks and monitor the task execution process, including monitoring the real-time voltage sequence of the distribution terminal battery, recording the voltage U1 before activation, the voltage U2 after discharge, the voltage U3 after charging, the soft pressure plate status, the activation time, and the switch status.

[0091] Battery health status assessment module: Communicates with the task execution and monitoring module, and is used to assess the battery health status and determine anomalies based on real-time voltage sequences and multi-source feature data using a deep learning model. The multi-source feature data includes voltage time series features, inherent battery features, operating environment features, historical performance features, charge and discharge parameter features, real-time status features, charge and discharge cycle features, battery material characteristics, and grid fluctuation correlation features.

[0092] Historical record query and archiving module: It is connected to the task execution and monitoring module and the battery health status assessment module. It is used to generate and store the activation history record after the activation task is completed. The record includes activation duration, key voltage data U1, U2, U3, voltage deviation ratio and abnormal marking information.

[0093] Multiple anti-misoperation verification module: Interacts with the activation strategy management module and the task execution and monitoring module to perform anti-misoperation verification before and during the activation task issuance, including: prohibiting the issuance of activation tasks when the battery voltage is 0, not including the current activation task when the battery activation soft pressure plate switching status is in the "on" state, prohibiting the issuance of remote control commands when the battery activation soft pressure plate switching point number is not 0, and locking the operation of terminals that are already in the activation process.

[0094] Compared with the prior art, the present invention has the following beneficial effects:

[0095] (1) Precision of state monitoring: By tracking the voltage and the voltage of its key process nodes (U1, U2, U3) in time sequence and integrating multi-source data such as battery SOC and SOH, a multi-dimensional battery health status assessment model was constructed, which greatly improved the accuracy of battery status diagnosis.

[0096] (2) Intelligent battery management: Battery activation strategies can be flexibly configured through the system main station, and differentiated strategy deployment and one-click batch execution can be supported by region, manufacturer, model and other dimensions. This realizes large-scale, lean intelligent operation and maintenance of massive terminal batteries, significantly improving operation and maintenance efficiency and terminal online rate.

[0097] (3) Operational safety: Multiple anti-misoperation verification logic and emergency interruption mechanism for voltage abnormality are designed to cover the pre-start, during-process and post-start, fundamentally eliminating the safety risks caused by misoperation and activation with defects. Attached Figure Description

[0098] Figure 1 A flowchart of the battery activation test method for power distribution terminals;

[0099] Figure 2 This is a schematic diagram of battery activation in a power distribution terminal. Detailed Implementation

[0100] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0101] This invention provides a method for testing the activation of batteries in a power distribution terminal, comprising:

[0102] The terminal battery voltage data is sent to the master station, which monitors the battery voltage data and controls the battery activation soft pressure plate.

[0103] Configure battery activation tasks, including creating target terminal activation tasks, setting activation strategies, and batch sending tasks to power distribution terminals;

[0104] Perform battery activation tasks, monitor the real-time voltage of the battery in the power distribution terminal, record the voltage U1 before activation, the voltage U2 after discharge, and the voltage U3 after charging, and monitor the status of the activation soft plate and the activation time at the same time.

[0105] Based on real-time voltage sequences and multi-source feature data, a deep learning model is used to assess battery health status, which includes the determination of abnormal battery activation.

[0106] After the task is completed, an activation history is generated and archived;

[0107] Perform error prevention checks during task execution;

[0108] The multi-source feature data includes voltage time-series features, inherent battery features, operating environment features, historical performance features, charge and discharge parameter features, real-time status features, charge and discharge cycle features, battery material characteristics, and grid fluctuation correlation features.

[0109] The present invention also provides a battery activation testing system for a power distribution terminal, comprising:

[0110] Activation Strategy Management Module: Used to configure battery activation strategies, including setting charge and discharge duration and charge and discharge voltage deviation ratio, and supports batch filtering of target terminals, creation of activation tasks, one-click batch distribution of tasks to power distribution terminals, and export of task list.

[0111] Task execution and monitoring module: It communicates with the activation strategy management module to receive activation tasks and monitor the task execution process, including monitoring the real-time voltage sequence of the distribution terminal battery, recording the voltage U1 before activation, the voltage U2 after discharge, the voltage U3 after charging, the soft pressure plate status, the activation time, and the switch status.

[0112] Battery health status assessment module: Communicates with the task execution and monitoring module, and is used to assess the battery health status and determine anomalies based on real-time voltage sequences and multi-source feature data using a deep learning model. The multi-source feature data includes voltage time series features, inherent battery features, operating environment features, historical performance features, charge and discharge parameter features, real-time status features, charge and discharge cycle features, battery material characteristics, and grid fluctuation correlation features.

[0113] Historical record query and archiving module: It is connected to the task execution and monitoring module and the battery health status assessment module. It is used to generate and store the activation history record after the activation task is completed. The record includes activation duration, key voltage data U1, U2, U3, voltage deviation ratio and abnormal marking information.

[0114] Multiple anti-misoperation verification module: Interacts with the activation strategy management module and the task execution and monitoring module to perform anti-misoperation verification before and during the activation task issuance, including: prohibiting the issuance of activation tasks when the battery voltage is 0, not including the current activation task when the battery activation soft pressure plate switching status is in the "on" state, prohibiting the issuance of remote control commands when the battery activation soft pressure plate switching point number is not 0, and locking the operation of terminals that are already in the activation process.

[0115] The following is a detailed implementation process of the present invention.

[0116] like Figure 1 , 2 As shown, the present invention provides a method for testing the activation of a power distribution terminal battery, comprising the following steps:

[0117] Step S1: The terminal battery voltage data is sent to the main station. The main station has the functions of monitoring battery voltage data and remote control of the battery activation soft pressure plate.

[0118] Furthermore, the battery activation soft pressure plate refers to a functional control switch in the power distribution terminal or master station system, which triggers the mode switching of the battery activation process through two state signals: "on" and "off". The on state corresponds to the battery discharge mode, and the off state corresponds to the battery charging mode.

[0119] Step S2: Configure battery activation task. Add filter conditions through the battery activation function module, create target terminal activation task, complete the addition, deletion and modification of activation strategy and selection status verification, execute one-click batch activation task and export task list.

[0120] Furthermore, the main station activation task involves selecting terminals and adding strategies to the remote control queue. The strategies include configuration of charging / discharging duration and charging / discharging voltage deviation ratio, allowing batch addition or removal of terminals for one-click parallel control, and providing execution step analysis and display as well as real-time monitoring of remote control tasks.

[0121] Step S3: Battery activation task execution monitoring process. After the one-click control is completed in step S2, the terminal performing the activation task enters the activation task monitoring module. Terminals in normal charging and discharging states can be monitored normally, while terminals with abnormal voltage or remote control malfunctions are directly archived to the history record. If a terminal malfunctions during the activation task execution, the activation task can be terminated with a one-click button, i.e., the system closes the activation pressure plate to stop battery discharge.

[0122] Furthermore, the activation task monitoring system simultaneously monitors the soft platen status, activation time, pre-activation voltage (U1), real-time voltage during the activation process, and key node voltages (U2 after discharge and U3 after charging), switch status, and voltage deviation. To improve the accuracy of the U1 data, the voltage three minutes after the activation task begins is taken as the initial monitoring voltage, which is used as the pre-activation voltage (U1).

[0123] Step S4: Handling voltage anomalies during the activation task execution in Step S3. The handling includes determining battery activation anomalies, abnormal voltage increases during battery discharge, and abnormal voltage decreases during battery charging.

[0124] Step S4 includes the construction of a multi-source feature system, the construction and training of a deep learning model, the dynamic anomaly determination process, and anomaly tracing and self-learning optimization.

[0125] (I) Construction of multi-source feature system (Table 1 shows multi-source features)

[0126] Table 1

[0127] Feature categories Specific features Feature Description Voltage timing characteristics Average voltage (U1) in the first 3 minutes before activation, real-time voltage sequence (U_t, t=1,2,...,T), voltage change rate (ΔU / Δt), and voltage fluctuation variance (σ_U). Reflects the dynamic trend of voltage changes and avoids the randomness of data at a single point in time. Battery inherent characteristics Battery type (lead-acid / lithium-ion), rated capacity (C), service life (Y), charge / discharge cycle count (N), factory parameters (rated voltage U0, internal resistance reference value R0) This reflects the individual differences in batteries and provides a basis for determining their differences. Operating environment characteristics Activation parameters include ambient temperature (T_env), humidity (H_env), terminal load rate (L), and master station power supply stability (S, 0-1 score). Environmental factors directly affect the battery's charge and discharge characteristics and need to be included in the calibration. Historical performance characteristics Historical activation anomaly count (A), historical maximum voltage deviation ratio (R_max), previous activation health score (S_prev), historical voltage recovery speed (V_rec) The probability of this anomaly is predicted based on historical performance. Charge and discharge parameter characteristics Set the charge / discharge current (I), charge / discharge duration threshold (T_charge / T_discharge), and activation strategy type (P, such as deep activation / shallow activation). Associate activation operation parameters to distinguish between "abnormal under normal parameters" and "normal under abnormal parameters". Real-time status characteristics Soft switch status (S_switch, engaged=1 / disengaged=0), remote control command response status (R_resp, normal=1 / abnormal=0), real-time battery internal resistance value (R_real) It reflects the operating status of the equipment during the activation process and eliminates voltage abnormalities not caused by the battery itself. Charge-discharge cycle characteristics Cycle depth distribution (DOD_distribution), cycle consistency coefficient (C_consistency) Cycle depth distribution reflects the proportion of cycles at different depths during battery charging and discharging, while cycle consistency coefficient quantifies the stability of charging and discharging parameters across multiple cycles. Both are strongly correlated with battery aging rate. Battery material characteristics Electrode material degradation rate (E_degradation), electrolyte conductivity (E_conductivity) The electrode material degradation rate is estimated from the battery's service life and charge / discharge intensity, while the electrolyte conductivity is derived from historical voltage-current data, directly affecting the battery's charge / discharge efficiency and capacity retention. Power grid fluctuation correlation characteristics Grid voltage fluctuation amplitude (G_fluctuation), grid frequency stability (G_frequency) (G_frequency) The power supply quality on the grid side affects battery charging efficiency and voltage stability, and should be included in the correction of interference factors for anomaly detection.

[0128] (II) Deep Learning Model Construction and Training

[0129] A dual-model architecture of "time series prediction model + anomaly classification model" is adopted to achieve dynamic threshold calculation and accurate anomaly type identification.

[0130] The specific construction and training process for (I) and (II) is as follows:

[0131] 1. Standardized construction of multi-source feature sets

[0132] Multi-source features need to undergo standardized processing throughout the entire process of "data acquisition - preprocessing - feature engineering - feature selection" to form a feature set with unified structure, complete information, and suitable for model input, providing a high-quality data foundation for model training.

[0133] (1) Data collection:

[0134] ① Real-time dynamic characteristics: Voltage time sequence characteristics (U1, Ut, voltage change rate, etc.) and real-time status characteristics (soft plate status, real-time internal resistance value, etc.) are collected through the real-time communication interface between the main station and the power distribution terminal. The collection frequency is 30 seconds / time to ensure the continuity and timeliness of the time sequence data.

[0135] ②Static inherent characteristics: Battery type, rated capacity, service life, etc. are extracted from the equipment file database, collected once, and updated synchronously with the equipment parameters;

[0136] ③ Environmental and operational characteristics: Ambient temperature, humidity, terminal load rate, etc., are collected through the terminal's built-in sensors or the distribution network environment monitoring system. The collection frequency is 5 minutes / time. Missing data is supplemented using linear interpolation. The difference formula is:

[0137] in For missing moments eigenvalues, , These are the two most recent valid feature values ​​before and after the missing time point. , These are the acquisition times for the corresponding valid feature values;

[0138] ④ Historical performance characteristics: The number of historical activation anomalies and the last health score are extracted from the activation history database and aggregated according to the device's unique identifier to ensure data traceability.

[0139] (2) Data acquisition and correction

[0140] ① Temperature correction optimization: The original linear correction, which only applied to ambient temperature, has been expanded into a piecewise nonlinear correction model to adapt to the differences in battery characteristics across different temperature ranges.

[0141] Low temperature range ( ≤0℃): Correction function

[0142] normal temperature range (0℃ < ≤ 35℃): Correction function (Preserve the original logic and adapt to normal environments)

[0143] High temperature range ( > 35℃): Correction function

[0144] ② Humidity Correction Model: A new piecewise humidity correction function has been added to address the impact of decreased battery insulation in high humidity environments and water loss in low humidity environments.

[0145] low humidity range Correction function

[0146] medium humidity range Correction function (No calibration required, optimal operating humidity for the battery)

[0147] High humidity range Correction function

[0148] ③ Atmospheric pressure correction model: A new environmental atmospheric pressure correction function has been added to adapt to the atmospheric pressure differences at different altitudes (atmospheric pressure affects the gas diffusion rate inside the battery): Correction function ,in The actual ambient pressure (Pa) is given, and 101325Pa is the standard atmospheric pressure.

[0149] ④ Comprehensive Environmental Correction Factor: Integrates temperature, humidity, and air pressure corrections to form a comprehensive environmental correction factor. Used for the acquisition and calibration of core data such as voltage and current:

[0150] in This is a temperature correction function (select the corresponding piecewise function according to the temperature range). For humidity correction function, This is the air pressure correction function.

[0151] ⑤ Battery Type Sub-categorization Correction: Based on the existing lead-acid / lithium-ion battery classification, further subdivide the models and optimize the correction function: | Battery Sub-categorization | Type Correction Function f(type) | Supplementary Notes |

[0152] |Valve-regulated lead-acid battery|1.0|Conventional lead-acid battery, retaining the original calibration logic|

[0153] |Gel lead-acid batteries|1.1|Different electrolyte characteristics result in slightly higher charge / discharge efficiency|

[0154] |Ternary lithium-ion battery|1.2|High energy density, good voltage stability|

[0155] |Lithium iron phosphate battery|1.3|Long cycle life, stable discharge voltage platform, and correction coefficient adapted to its characteristics|

[0156] Terminal model compatibility calibration: Add terminal model compatibility calibration coefficient. Corrections are made based on the efficiency differences of power management modules in different terminals: ,in The standard terminal power management efficiency is taken as the industry average of 92%. The power management efficiency of the current terminal model (extracted from the device profile database).

[0157] ⑥ Data acquisition process correction procedure:

[0158] Raw data acquisition: Collect data such as voltage, current, temperature, humidity, and air pressure at the original frequency, and record equipment information such as battery model and terminal model at the same time;

[0159] Environmental correction: Multiply the original voltage and current data by a comprehensive environmental correction factor. We obtain environmentally corrected data;

[0160] Device characteristic correction: Multiply the environmentally corrected data by the battery type correction function f(type) and the terminal model correction coefficient. This yields the final correction data;

[0161] Anomaly detection: For the corrected data, the original data is used. Outlier identification is performed using criteria and box plots to ensure data validity.

[0162] (3) Data preprocessing:

[0163] ① Continuous features (voltage, temperature, internal resistance, etc.): K-nearest neighbor (K=5) interpolation is used to supplement these features. Neighbor samples are selected based on the similarity of equipment model and operating years to avoid distribution distortion caused by mean interpolation. The interpolation formula is as follows:

[0164]

[0165] in , The feature distance between the sample to be interpolated and its j-th neighbor sample is calculated using Euclidean distance: , For feature dimension, For the sample to be interpolated 3D eigenvalues For the first The neighbor sample number 3D eigenvalues;

[0166] ②Key features (charging and discharging current, real-time internal resistance): If the missing rate exceeds 10%, it is marked as an invalid sample and removed. At the same time, the device identification is recorded for subsequent manual verification.

[0167] ③ Numerical characteristics: Extreme outliers are identified based on the 3σ criterion (such as voltage exceeding the rated voltage by ±30% or internal resistance exceeding the factory reference value by ±50%), and verified by the interquartile range (IQR) method of box plot to ensure the accuracy of outlier identification; The criteria for outlier identification in the guidelines are as follows: ,in The characteristic mean, The characteristic standard deviation; in the box plot method, the outlier detection criteria are: or ,in The lower quartile, It is the upper quartile. ;

[0168] ④ Logical outliers: Filtered by business rules (e.g., negative discharge duration, charge / discharge cycle count exceeding the equipment design limit, negative voltage deviation ratio);

[0169] ⑤ Timing characteristics: Time alignment processing is performed to uniformly truncate the voltage sequences of different activation tasks to a fixed length (e.g., if the activation process lasts for T minutes, sampling is performed at 30 seconds / time, resulting in 2T data points). Any insufficient data points are filled with the last data point, and any excess data points are truncated from the core data of the latter part to ensure the consistency of timing characteristic length.

[0170] (4) Feature engineering optimization:

[0171] ① Trend-based derived features: Based on the voltage sequence, calculate the mean, maximum, minimum, and slope change rate of a sliding window (window size = 5) to capture local voltage fluctuation trends; the formula for the sliding window mean is:

[0172]

[0173] The formula for the rate of change of the slope of the sliding window is:

[0174]

[0175] in, for The voltage value at that moment;

[0176] ② Frequency Domain Feature Extraction: The time-domain voltage sequence is converted into a frequency-domain signal using Fast Fourier Transform (FFT). The amplitude and phase of the first five dominant frequency components are taken as features to reflect the periodicity of voltage fluctuations. The FFT formula is:

[0177] in, It is a time-domain voltage sequence. For sequence length, It is a frequency domain signal;

[0178] ③ Statistical derived characteristics: Calculate the skewness, kurtosis, and autocorrelation coefficient (lag 1-3 steps) of the voltage series to quantify the voltage distribution characteristics and time correlation. The skewness formula is:

[0179]

[0180] The kurtosis formula is:

[0181]

[0182] Autocorrelation coefficient (lag) The formula for the first step is:

[0183]

[0184] in, For voltage sequence data, The mean, Standard deviation, The sequence length is given.

[0185] (5) Feature selection: Redundant and weakly correlated features are removed to reduce the computational complexity of the model and improve generalization ability. The formula for calculating mutual information is:

[0186]

[0187] in, As characteristic variables, For the target variable, For joint probability distribution, , These are the marginal probability distributions.

[0188] 2. The entire model training process

[0189] (1) Dataset construction

[0190] ①Data source: Collected historical activation data of power distribution terminal batteries of different models (lead-acid batteries, lithium-ion batteries), different operating years (0-8 years), and different operating environments (high temperature, humidity, high load, etc.) for more than 3 years, with a total of 15,000 valid samples (each sample contains complete activation process characteristic data, dynamic threshold labels, and anomaly type labels).

[0191] ② Data splitting: The training set (12,000 records) and the test set (3,000 records) were randomly split in an 8:2 ratio to ensure that the sample distribution of the training set and the test set is consistent (verified by chi-square test, p value > 0.05).

[0192] ③ Labeling:

[0193] Dynamic threshold label ( Based on expert experience and actual activation results, combined with battery type, ambient temperature, and charge / discharge parameters, the reasonable voltage deviation threshold range for different scenarios is calculated, and the median value of the range is taken as the label. The calculation formula is as follows:

[0194]

[0195] in, As a baseline threshold based on battery type, The correction threshold is based on ambient temperature. The correction threshold is based on charge and discharge parameters. , , Let be the weighting coefficient, satisfying + + =1, the weight value is determined through grid search optimization;

[0196] Anomaly type labels: These labels are manually created based on the activation process logs and on-site verification results to ensure label accuracy (labeling accuracy rate ≥ 98%).

[0197] (2) Training parameter configuration

[0198] ① Optimizer: The Adam optimizer is used, with an initial learning rate of 0.001, and a learning rate decay strategy is employed (every 5 seconds of learning rate decay). (Attenuation of 10%), the attenuation formula is:

[0199]

[0200] in, =0.001 is the initial learning rate. =0.9 is the attenuation coefficient. For training rounds;

[0201] ② Training batch: Batch Size=64, training rounds ( =50, adopting an early stopping strategy, when the test set loses 3 consecutive times. Stop training if there is no decline.

[0202] ③ Regularization: In addition to dropout, L2 regularization (weight decay coefficient = 0.001) is added to the fully connected layer to prevent overfitting. The L2 regularization loss term is: ,in =0.001 is the weight decay coefficient. These are the model weight parameters;

[0203] ④ Equipment environment: Implemented based on the PyTorch framework, using GPU (NVIDIA Tesla V100) to accelerate training.

[0204] (3) Training steps

[0205] Step 1: Data preprocessing and feature set construction (follow the above "standardized construction of multi-source feature sets" process);

[0206] Step 2: Initialize the dual model parameters (use the He initialization method to initialize the weights of the fully connected layers, and randomly initialize the weights of the LSTM layers). The He initialization formula is: ,in The number of input neurons for the fully connected layer;

[0207] Step 3: Iterative Training

[0208] Forward propagation: Input the features of the training set into the model to obtain the dynamic threshold prediction value and the probability distribution of the anomaly type, respectively;

[0209] Loss calculation: Calculate the total loss based on the joint loss function;

[0210] Backpropagation: The model parameters are updated using gradient descent (the parameters of the shared layer are updated jointly by the gradients of the two sub-models).

[0211] Step 4: Model Validation: After training one... Calculate model performance metrics using the test set (regression task: R²; Classification task: accuracy, recall, F1 score). The calculation formula is

[0212]

[0213] The calculation formula is:

[0214]

[0215] in, The number of samples in the test set. For the true value, For predicted values, The mean of the true values;

[0216]

[0217] Recall rate ( The calculation formula is:

[0218]

[0219] The formula for calculating the F1 score is:

[0220]

[0221] in, = , It is a true positive. It is a true negative. It was a false positive. It is a false negative;

[0222] Step 5: Model selection: Save the model with the best performance on the validation set as the final training model, and ensure that the accuracy on the test set is ≥95% and the recall is ≥92% (meeting the design requirements).

[0223] (4) Model incremental update mechanism

[0224] ① Data accumulation: Collect newly activated data monthly (estimated 1000-1500 records), and after preprocessing and feature engineering, form an incremental dataset;

[0225] ②Incremental training:

[0226] A "warm start" strategy is adopted, loading the historically optimal model parameters as initial weights;

[0227] The incremental dataset is mixed with 20% (2400 records) of the original training set to form a new training set (to avoid catastrophic forgetting).

[0228] The training epochs were set to 10, the learning rate was adjusted to 0.0005, and other parameters remained unchanged;

[0229] ③ Model evaluation: After incremental training, the performance is verified using the latest test set (20% of the newly added data in the current month). If the accuracy or recall drops by more than 2%, full training is performed again.

[0230] ④ Parameter Deployment: Synchronize the updated model parameters to the main station system, replace the original model, and ensure that the model adapts to changes in battery performance and environmental differences.

[0231] 3. Model Deployment and Inference Process

[0232] (1) Deployment method

[0233] TensorRT is used to accelerate the model. After the trained PyTorch model is converted to ONNX format, it is deployed to the GPU inference engine of the main server to ensure that the real-time inference latency is ≤100ms (meeting the real-time requirement of receiving feature data once every 30 seconds).

[0234] (2) Reasoning process

[0235] Data input: The latest feature data is received every 30 seconds. After preprocessing and feature fusion, a 19-dimensional fused feature vector is generated.

[0236] Parallel inference: Two sub-models are input simultaneously. The time-series prediction sub-model outputs the dynamic bias threshold (\(n_{dynamic}\)), and the anomaly classification sub-model outputs the anomaly type probability.

[0237] Output results: Input the dynamic deviation threshold into the anomaly determination formula (formulas (1) and (2) in step S4), and combine it with the anomaly type probability to generate the final anomaly determination result and type identifier.

[0238] (3) Model monitoring

[0239] The main station system monitors the model inference performance (latency and accuracy) in real time. When the inference latency exceeds 200ms or the prediction accuracy of 100 consecutive samples is lower than 90%, an alarm is triggered and the system automatically switches to the backup model, while simultaneously initiating the model troubleshooting process.

[0240] (III) Dynamic Anomaly Detection Logic

[0241] Real-time anomaly detection formula:

[0242] The real-time anomaly detection during the discharge process is as follows:

[0243] (1)

[0244] in, This represents the real-time voltage at time t during the discharge process. This is the discharge correction factor, calculated by the model based on battery type, ambient temperature, and discharge current; the specific formula is:

[0245]

[0246] in, This is a battery type correction function (1.0 for lead-acid batteries and 1.2 for lithium-ion batteries). For ambient temperature correction function, =1+0.01×(25− (Base temperature is 25℃). This is the discharge current correction function. =1+0.005×( - )( (where a is the standard discharge current), a1, a2, and a3 are weighting coefficients, determined through model training and optimization, satisfying... + + =1; This is the dynamic deviation threshold for the discharge phase output by the model.

[0247] The real-time anomaly detection during the charging process is as follows:

[0248] (2)

[0249] in, The voltage at time t during the charging process is the real-time voltage. This is the charging correction factor, calculated by the model based on the battery cycle count, charging time, and internal resistance change rate. The specific calculation formula is as follows:

[0250]

[0251] in, This is a function to correct the number of iterations. , For charging time correction function, ( (Standard charging time) This is the correction function for the rate of change of internal resistance. ( , For real-time internal resistance, (The internal resistance reference value) , , The weighting coefficients are determined through model training and optimization, satisfying the following conditions: ; This is the dynamic deviation threshold for the discharge phase output by the model.

[0252] (iv) Anomaly tracing and self-learning optimization

[0253] After each anomaly is identified, the system automatically initiates source tracing analysis, combines the anomaly type probability output by the model, matches similar historical anomaly cases, and generates a source tracing report; the current anomaly data, handling results, and source tracing conclusions are used as new samples to supplement the training set, optimize the model's recognition accuracy for similar anomalies, and form a closed-loop mechanism of "identification-handling-source tracing-optimization".

[0254] Step S5: After the battery activation task is completed, the task is written to the history record. Different colors are used to display the situation such as battery voltage of 0, voltage deviation value of negative, voltage deviation value exceeding the threshold abnormality, and soft pressure plate control abnormality, and the record is recorded.

[0255] Furthermore, activation historical data is used to generate activation conclusions. Abnormal activation terminals are marked with the activation anomaly time point and charging or discharging status, and U1 (voltage before discharge), U2 (voltage after discharge), U3 (voltage after charging), discharge voltage deviation ratio, and charging voltage deviation ratio are displayed. The activation duration is calculated based on the activation start and end times. Abnormal terminals can initiate defect procedures within the battery activation module, achieving seamless handling of the entire battery module process for abnormal terminals.

[0256] Activation task completed normally:

[0257] Activation time = Discharge time + Charging time

[0258] (3)

[0259] in, = - (in For discharge start-up time, (Discharge end time) = - (in For charging start-up time, (This refers to the charging end time).

[0260] Activation task that terminated abnormally:

[0261] Activation duration = runtime before the anomaly occurred

[0262] = (4)

[0263] Step S6: The activation task supports error prevention verification. Activation is prohibited when the battery voltage is 0. Activation is not included in this activation if the battery activation soft pressure plate is in the "in" state. Remote control is prohibited when the battery activation soft pressure plate switching point number is not 0.

[0264] Furthermore, the terminals that are already in the activation process are locked to prevent the activation conclusion from being abnormal due to mis-control of the pressure plate during the activation process. The soft pressure plate cutting point number is set to 0 to form a point table specification.

[0265] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for testing the activation of batteries in a power distribution terminal, characterized in that, include: The terminal battery voltage data is sent to the master station, which monitors the battery voltage data and controls the battery activation soft pressure plate. Configure battery activation tasks, including creating target terminal activation tasks, setting activation strategies, and batch sending tasks to power distribution terminals; Perform battery activation tasks, monitor the real-time voltage of the battery in the power distribution terminal, record the voltage U1 before activation, the voltage U2 after discharge, and the voltage U3 after charging, and monitor the status of the activation soft plate and the activation time at the same time. Based on real-time voltage sequences and multi-source feature data, a deep learning model is used to assess battery health status, which includes the determination of abnormal battery activation. After the task is completed, an activation history is generated and archived; Perform error prevention checks during task execution; The multi-source feature data includes voltage time-series features, battery inherent features, operating environment features, historical performance features, charge and discharge parameter features, real-time status features, charge and discharge cycle features, battery material characteristics, and grid fluctuation correlation features.

2. The method for testing the activation of a distribution terminal battery according to claim 1, characterized in that, During the data acquisition process of multi-source feature data, environmental data needs to be corrected using a comprehensive environmental correction coefficient. , ,in: This is a temperature correction function, based on the ambient temperature. The interval to be selected is a piecewise function, including: Low temperature range ≤0℃): Correction function ; Normal temperature range, i.e., 0℃ < ≤ 35℃): Correction function ; High temperature range > 35℃: Correction function ; This is a humidity correction function, based on the ambient humidity. Choose a piecewise function, including: Low humidity range Correction function ; medium humidity range Correction function ; High humidity range Correction function ; This is the barometric pressure correction function. , The actual ambient pressure is given; 101325 Pa is the standard atmospheric pressure.

3. The method for testing the activation of a distribution terminal battery according to claim 2, characterized in that, The data acquisition process for multi-source feature data also includes correction for battery type and terminal model, using a battery type correction function f(type) and a terminal model correction coefficient. ,in: f(type) takes a value based on the specific battery type, including: 1.0 for valve-regulated lead-acid batteries, 1.1 for gel lead-acid batteries, 1.2 for ternary lithium-ion batteries, and 1.3 for lithium iron phosphate batteries; ,in For standard terminal power management efficiency, This refers to the power management efficiency of the current terminal model.

4. The method for testing the activation of a distribution terminal battery according to claim 1, characterized in that, After the multi-source feature data is collected, data preprocessing is required. The data preprocessing operations include: For missing data, interpolation is performed using the K-nearest neighbor interpolation method. The interpolation formula is as follows: in , The feature distance between the sample to be interpolated and its j-th neighbor sample is calculated using Euclidean distance: , For feature dimension, For the sample to be interpolated 3D eigenvalues For the first The neighbor sample number 3D eigenvalues; And outlier detection is performed using the 3σ criterion and box plot method, where: The outlier determination criteria of the 3σ criterion are: ,in The characteristic mean, The characteristic standard deviation; The outlier detection criteria for box plot method are: or ,in The lower quartile, It is the upper quartile. .

5. The method for testing the activation of a distribution terminal battery according to claim 1, characterized in that, After the multi-source feature data is collected and preprocessed, feature engineering optimization is required, including derived feature calculation and feature selection. Derived feature calculation includes: The formula for calculating the sliding window mean of a voltage series is: in, for Sliding window mean at all times For the current moment, The index of the time within the window, the value. -4 to , for Real-time voltage is used, with a fixed window size of 5 consecutive sampling points to smooth voltage time-series data; The formula for calculating the rate of change of the sliding window slope is: in express The rate of change of the voltage slope of the sliding window at any time reflects the trend of voltage change within that window; express Real-time voltage value at any given moment; express Voltage values ​​four time units prior to the current moment; It represents the time span of the sliding window and is used to normalize the voltage change, so that the rate of change of the slope has a uniform dimension. The time-domain voltage sequence is converted into a frequency-domain signal using the Fast Fourier Transform (FFT), as shown in the following formula: in, It is a time-domain voltage sequence. For sequence length, It is a frequency domain signal; Calculate the skewness, kurtosis, and autocorrelation coefficient of the voltage series, where: The skewness formula is: The kurtosis formula is: The formula for the autocorrelation coefficient with a lag of k steps is: in, For voltage sequence data, The mean, Standard deviation, The sequence length; Feature selection is performed using mutual information (MI). The formula for mutual information is: in, As characteristic variables, For the target variable, For joint probability distribution, , These are the marginal probability distributions.

6. The method for testing the activation of a distribution terminal battery according to claim 1, characterized in that, The training of the deep learning model includes: 1) Construct a training dataset by collecting historical data on battery activation of different models, operating years, and operating environments of power distribution terminals for more than 3 years, and collecting valid samples. Each sample contains complete characteristic data of the battery activation process of the power distribution terminal, dynamic threshold labels, and abnormal type labels. The valid samples are randomly divided into training dataset and test dataset in an 8:2 ratio. Dynamic threshold label The calculation formula is: in, As a baseline threshold based on battery type, The correction threshold is based on ambient temperature. The correction threshold is based on charge and discharge parameters. , , Let be the weighting coefficient, satisfying + + =1, the weight value is determined through grid search optimization; Anomaly type labels: These are manually labeled based on the activation process log and on-site verification results. 2) Configure training parameters, including: Learning rate decay formula: in =0.001 is the initial learning rate. =0.9 is the attenuation coefficient. For training rounds; L2 regularization loss term: ,in =0.001 is the weight decay coefficient. These are the model weight parameters; 3) Model Training: Input the training dataset; initialize the dual model parameters. Use the He initialization method to initialize the weights of the fully connected layers, and randomly initialize the weights of the LSTM layers. The He initialization formula is: ,in The number of neurons is input to the fully connected layer; iterative training is performed, and dynamic threshold predictions and anomaly type probability distributions are obtained based on the training dataset. The total loss is calculated according to the joint loss function, and the model parameters are updated using gradient descent to obtain the trained model; the performance metrics of the trained model are calculated using the test dataset to validate the model; the model with the best performance on the validation set is saved as the final training model.

7. The method for testing the activation of a distribution terminal battery according to claim 1, characterized in that, The battery activation anomaly determination includes: The formula for determining abnormalities during the discharge process is as follows: Where Ut is the real-time voltage at time t during the discharge process. The dynamic deviation threshold for the discharge phase output by the model; The discharge correction factor is calculated using the following formula: in , , Let be the weighting coefficient, satisfying + + =1, This is a battery type correction function. This is the ambient temperature correction function. =1+0.01×(25− ), For ambient temperature, This is the discharge current correction function. =1+0.005×( - ), Standard discharge current, This is the real-time discharge current; The formula for determining abnormalities during the charging process is as follows: in This is the dynamic deviation threshold for the discharge phase output by the model. The charging correction factor is calculated using the following formula: in , , Let be the weighting coefficient, satisfying + + =1, This is a function to correct the number of iterations. , This indicates the number of charge-discharge cycles of the battery in the power distribution terminal, reflecting the degree of battery wear and tear, and is obtained from the battery's operating history. For charging time correction function, , The actual duration of this charging process is recorded in real time by the task execution and monitoring module. Standard charging time, This is the correction function for the rate of change of internal resistance. , , For real-time internal resistance, This is the reference value for internal resistance.

8. The method for testing the activation of a distribution terminal battery according to claim 1, characterized in that, The generation of activation history records includes calculating the activation duration: For activation tasks that are completed normally, the activation duration formula is: in = - , For discharge start-up time, This is the discharge end time. = - . For charging start-up time, This is the charging end time; For activation tasks that terminate abnormally, the activation duration formula is: = in This represents the runtime before the exception occurred.

9. The method for testing the activation of a distribution terminal battery according to claim 1, characterized in that, The error prevention verification includes: Activation is disabled when the battery voltage is 0. The battery activation soft pressure plate is in a state where it is not included in this activation process when it is put into operation. Remote control is prohibited when the battery activation soft pressure plate switching point number is not 0.

10. A battery activation testing system for a power distribution terminal, characterized in that, include: Activation Strategy Management Module: Used to configure battery activation strategies, including setting charge and discharge duration and charge and discharge voltage deviation ratio, and supports batch filtering of target terminals, creation of activation tasks, one-click batch distribution of tasks to power distribution terminals, and export of task list. Task execution and monitoring module: It communicates with the activation strategy management module to receive activation tasks and monitor the task execution process, including monitoring the real-time voltage sequence of the distribution terminal battery, recording the voltage U1 before activation, the voltage U2 after discharge, the voltage U3 after charging, the soft pressure plate status, the activation time, and the switch status. Battery health status assessment module: Communicates with the task execution and monitoring module, and is used to assess the battery health status and determine anomalies based on real-time voltage sequences and multi-source feature data using a deep learning model. The multi-source feature data includes voltage time series features, inherent battery features, operating environment features, historical performance features, charge and discharge parameter features, real-time status features, charge and discharge cycle features, battery material characteristics, and grid fluctuation correlation features. Historical record query and archiving module: It is connected to the task execution and monitoring module and the battery health status assessment module. It is used to generate and store the activation history record after the activation task is completed. The record includes activation duration, key voltage data U1, U2, U3, voltage deviation ratio and abnormal marking information. Multiple anti-misoperation verification module: Interacts with the activation strategy management module and the task execution and monitoring module to perform anti-misoperation verification before and during the activation task issuance, including: prohibiting the issuance of activation tasks when the battery voltage is 0, not including the current activation task when the battery activation soft pressure plate switching status is in the "on" state, prohibiting the issuance of remote control commands when the battery activation soft pressure plate switching point number is not 0, and locking the operation of terminals that are already in the activation process.

Images