A battery application management method

Abstract

The application discloses a battery application management and control method, which dynamically constructs a time sequence sequence prediction model cluster of multiple attribute characteristics based on multiple battery attribute characteristics, selects attribute characteristics most valuable for safety risk prediction online, adapts to safety risk prediction needs of different manufacturers, different models and different working conditions of batteries, and expands safety risk sample data by using a label propagation algorithm, deeply mines and fully utilizes unlabeled data, and can more accurately and stably perform safety risk prediction on lithium batteries.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of battery management system, and particularly relates to a battery application management and control method. BACKGROUND

[0002] At present, the traditional base station battery is only used for emergency standby power when the commercial power is off, and has not become a production tool for base station asset management, resulting in a large waste of silent assets of energy storage equipment of operators. At present, the power department divides 24 hours a day into several time periods such as peak, high peak, normal time and low valley according to the load change of the power grid, and formulates different electricity prices for each time period, so as to encourage users to reasonably arrange power consumption time, stagger power consumption, improve the utilization efficiency of equipment and save energy.

[0003] Although the intelligent peak-shaving technology can improve the utilization rate of the battery and save electricity bills, the primary role of the base station battery is to provide backup power. Therefore, the battery still needs to meet the backup power requirement after peak-shaving, and the peak-shaving power consumption strategy needs to be accurately calculated to ensure reliable backup power and maximize benefits. The backup power duration of the battery is affected by multiple factors such as load power, battery SOC, battery age, etc. The load power, battery capacity and health status of different sites are very different. Therefore, early judgment of battery risks and early prediction can guide the staff to handle the corresponding risks in time, which is crucial to ensure the safe operation of the base station in the network that uses the intelligent peak-shaving technology.

[0004] At present, the battery is facing the problem that the mechanism research lags behind the engineering implementation. The mechanism research aims to deeply understand the working principle and failure mechanism of the lithium battery inside, but it is often difficult to keep up with the rapid development of engineering technology in practice. This means that we cannot simply rely on theoretical models to predict and explain all types of lithium battery safety risks. Therefore, it is necessary to take a data-driven approach to gradually capture effective features and eliminate the interference of ineffective features.

[0005] Based on the data-driven approach, there are two key challenges in solving the problem of lithium battery safety risk.

[0006] First, it is difficult to determine the characteristics of battery safety risk.

[0007] Traditional battery safety risk prediction schemes rely solely on cross-sectional data at a single time point, lacking observation and analysis of early indicators of lithium battery safety risks. Before a battery exhibits obvious safety risks, early indicators of lithium battery safety risks can predict some specific indicators or patterns of potential problems. These features need to be observed over a period of time to extract the hidden patterns. For example, monitoring the timing characteristics of the battery's charge and discharge cycles, temperature changes, voltage fluctuations, etc. By collecting and analyzing these time-varying data, some patterns or trends related to battery safety risks can be identified. This method of digging deep into the hidden information behind the data is crucial for improving the safety and reliability of lithium batteries.

[0008] In addition, the early indicators of lithium battery safety risks are not completely the same for different manufacturers, models and batches. This is because there are differences in battery design, material selection, manufacturing process, etc. among various manufacturers, resulting in differences in performance and reliability of their products, which also makes it difficult to determine effective battery safety risk characteristics.

[0009] Second, there is a lack of labeled data, and a large amount of unlabeled data / samples cannot be effectively utilized

[0010] In current lithium battery safety risk analysis, there is relatively little labeled data, especially a limited number of samples with safety risks. This poses a challenge to building an accurate and reliable prediction model. In actual engineering deployment, it is difficult to accurately label the safety risk of each sample due to resource constraints or technical difficulties. This leads to the scarcity of samples with safety risks, further exacerbating the small sample problem.

[0011] In addition, there are a large number of samples in many engineering deployments that are not entirely clear, i.e. those that exhibit some risk characteristics but do not fully meet the typical safety risk patterns. These samples may be due to environmental factors, usage conditions or other unknown factors. However, due to the lack of clear labeling standards or expert judgment, these samples are often not effectively utilized. They cannot be considered normal samples and are excluded from analysis.

[0012] In summary, the effective characterization of battery safety risk characteristics and the low utilization rate of unlabeled samples are important challenges in lithium battery safety risk analysis technology, which need to be addressed. SUMMARY

[0013] The purpose of the present application is to solve the problems of the prior art, and to provide a battery application management and control method, which solves the problems of difficult determination of battery safety risk characteristics, small sample size and low utilization rate of a large amount of unlabeled samples in the battery risk prediction technology scheme in intelligent peak-shaving.

[0014] In order to solve the technical problem, the technical scheme of the present application is: a battery application control method, comprising the following steps:

[0015] Step 1: Obtain battery operation business data as a set of multi-attribute features of the battery;

[0016] Step 2: Battery operation business data preprocessing, preprocessing includes but is not limited to deleting error values, duplicate values, replacing outliers, and filling missing values;

[0017] Step 3: Construct a time series sequence prediction model cluster of multi-attribute features, and select an effective set of battery safety risk attribute features based on the prediction accuracy of each model in the model cluster;

[0018] Step 3-1: Construct a time series sequence prediction model cluster of multi-attribute features;

[0019] Establish a time series sequence prediction model of multi-attribute features, the model input includes the sampling sequence of multiple attribute features of the battery at a specific time, the model output is the prediction of battery safety risk, and the time series sequence prediction model of multi-attribute features is trained and evaluated using the safety risk sample or updated label data of the preprocessed multi-attribute feature set;

[0020] Iterate and combine the preprocessed multi-attribute feature set, train a time series sequence prediction model for each attribute feature combination, and combine all attribute feature combinations to form a time series sequence prediction model cluster of multi-attribute features;

[0021] Step 3-2: Online selection of an effective set of battery safety risk attribute features based on the prediction accuracy of each model in the time series sequence prediction model cluster of multi-attribute features;

[0022] Step 4: Label data expansion; based on the effective set of battery safety risk attribute features obtained by online selection, label the unlabeled samples in the preprocessed multi-attribute feature set by using a differential measurement algorithm and a label propagation algorithm, and obtain updated label data;

[0023] Step 5: Re-train the time series sequence prediction model cluster of multi-attribute features based on the updated label data, and iterate for multiple rounds until the proportion of unlabeled samples is lower than a threshold, and output a battery safety risk prediction model;

[0024] Step 6: Battery safety risk prediction; when a new battery working condition sample arrives, input it into the trained battery safety risk prediction model to obtain its corresponding safety risk prediction result, and the base station with high-risk battery no longer issues intelligent peak-shaving strategy to ensure the safety of the base station operation.

[0025] Preferably, the battery operation business data in step 1 is divided into three types, namely battery operation data, battery state data and battery alarm data, as a set of multi-attribute features of the battery, the battery operation data includes voltage, temperature and humidity data during the battery charging and discharging process, the battery state data includes capacity, remaining capacity, maximum available capacity, internal resistance and life, and the battery alarm data includes voltage anomaly and temperature anomaly.

[0026] Preferably, the pre-processing of the battery operation business data in step 2 is specifically as follows:

[0027] For error values and repeated values, delete them;

[0028] For outliers, based on the box plot, replace the high outliers in the data with the third quartile and replace the low outliers in the data with the first quartile;

[0029] For missing values, use linear interpolation to fill in the missing values.

[0030] Preferably, step 3 is specifically as follows:

[0031] Step 3-1: Construct a set of time series prediction models of multi-attribute features;

[0032] Select long short-term memory network (LSTM) as the time series prediction model of multi-attribute features. When establishing the time series prediction model of multi-attribute features, first define the input and output of the model, the input includes the sampling sequence of multiple attribute features of the battery within a certain time, and the model output is the prediction of the safety risk of the battery. Divide the safety risk sample or updated label data of the pre-processed multi-attribute feature set into two parts according to the ratio of 8:2, which are training set and validation set. Train the time series prediction model of multi-attribute features, the training set is used to train the model, so that it can learn the time series relationship between attribute features and the leading indicator features of safety risk, and the validation set is used to evaluate the performance of the model.

[0033] Iterate and combine the pre-processed multi-attribute feature set, train an LSTM time series prediction model for each attribute feature combination, combine all the LSTM time series prediction models constructed by attribute feature combinations to form an LSTM model cluster, and each model in the cluster is responsible for processing different attribute features or attribute feature combinations.

[0034] Step 3-2: Online selection of effective attribute feature set representing battery safety risk

[0035] According to the model prediction accuracy of the time sequence prediction model of the multi-attribute feature, the contribution value of each attribute feature is calculated, and the top-k attribute features strongly related to the safety risk prediction are selected according to the size, an attribute feature set effectively representing the battery safety risk is obtained, and the k attribute features are normalized.

[0036] Preferably, the contribution value of each attribute feature is calculated according to the following formula:

[0037]

[0038] wherein:

[0039] represents the prediction result;

[0040] represents the expected value of the prediction result;

[0041] v(S) represents the shap value of the attribute feature subset S;

[0042] represents the shap value of the new subset after adding the attribute feature i.

[0043] Preferably, the weight w[k] of the normalization processing is calculated according to the following formula:

[0044] w[k]=

[0045] wherein:

[0046] weight[k] represents the weight value of the kth attribute feature;

[0047] represents the sum of the weight values of all k attribute features.

[0048] Preferably, the step 4 is specifically:

[0049] Step 4-1: sample difference measurement;

[0050] In the battery safety risk prediction, DTW is used to compare whether the behavior patterns of two samples on a certain attribute feature are similar:

[0051] D(a,b)=Σ(w[k] DTW(a[k],b[k]))

[0052] wherein:

[0053] D(a,b) represents the difference measurement between sample a and sample b;

[0054] w[k] represents the normalized weight of the kth attribute feature;

[0055] DTW(a[k],b[k]) represents the dynamic time warping distance between sample a and sample b in the kth attribute feature representing the battery safety risk;

[0056] Step 4-2: label propagation

[0057] Initialization: assign an initial label to each unlabeled sample, usually by random assignment;

[0058] Iterative update: calculate the weight between each unlabeled sample and the surrounding labeled samples according to the difference measure D(a,b), the greater the weight, the more similar the two samples, and the greater the impact of label propagation;

[0059] Label propagation: update the label of the unlabeled sample according to the weight, which is realized by the weighted average aggregation method;

[0060] Convergence judgment: repeat the iterative update step until the convergence condition is met, the iteration number reaches the predetermined value or the label change is less than the threshold;

[0061] Label the unlabeled samples in the preprocessed multi-attribute feature set to obtain updated label data.

[0062] Preferably, the step 6 is specifically: when a new battery operating condition sample is collected and transmitted to the system, input the trained battery safety risk prediction model to obtain the safety risk prediction result, and perform real-time safety risk prediction and analysis on the new battery operating condition sample, and the intelligent peak-shaving strategy is no longer issued to the base station with high-risk batteries to ensure the safety of the base station operation.

[0063] Preferably, the battery application control method is executed by a battery control platform, and the battery control platform includes a basic function module, a battery health management module, a battery backup power analysis module, a battery life prediction module, an intelligent peak-shaving charging and discharging module, and a peak-shaving income calculation module.

[0064] The basic function module can obtain and view battery operation business data of the base station, including configuration, battery operation data, battery state data, and battery alarm data, and remotely control the battery;

[0065] The battery health management module can evaluate the health status of each group of batteries according to the battery operation business data to obtain a safety risk prediction result; support battery monitoring and management of the whole network / subnetwork / site, remind by notification and alarm, and provide related reports;

[0066] The battery backup analysis module can analyze the backup power capacity of the site or power supply system according to battery operation business data, load current and related alarms; support statistical analysis of the whole network / subnetwork / site; and provide backup power capacity analysis results.

[0067] The battery life prediction module can predict the remaining life of the battery according to the battery installation date, environmental temperature, battery type and battery operation working condition data.

[0068] The intelligent peak-shaving charging and discharging module can remotely set the peak-shaving power consumption parameters of each battery group, enable and disable the peak-shaving power consumption function, and query, set, modify and delete the peak-shaving power consumption mechanism to realize low-price self-charging and high-price self-discharging.

[0069] The peak-shaving income estimation module can view the detailed information of the peak-shaving power consumption of each battery group in real time, count and analyze the peak-shaving power consumption data and give the results, and evaluate the income of the peak-shaving power consumption.

[0070] Compared with the prior art, the application has the following advantages:

[0071] (1) The application discloses a battery application management and control method, which dynamically constructs a time sequence prediction model cluster of multiple attribute characteristics based on multiple battery attribute characteristics, selects the attribute characteristics most valuable for safety risk prediction online, adapts to the safety risk prediction needs of different manufacturers, different models of batteries and different working conditions, and expands safety risk sample data by using a label propagation algorithm, deeply mines and fully utilizes unlabeled data, and can more accurately and stably predict the safety risk of lithium batteries.

[0072] (2) The application constructs a time sequence prediction model cluster based on multiple battery attribute characteristics, selects multiple attribute leading indicator characteristics effectively representing the safety risk of the battery online, enhances the robustness of the model, and solves the problem of difficult determination of effective attribute characteristics.

[0073] (3) The application expands risk sample data based on the multiple attribute leading indicator characteristics effectively representing the safety risk of the battery obtained by online selection and the label propagation algorithm, iteratively trains the battery safety risk prediction model, improves the generalization ability and robustness of the model, and solves the problem of small sample size of the battery safety risk.

[0074] (4) The application extracts high-weight original attribute characteristics and eliminates invalid attribute characteristics from the battery multi-attribute characteristic sample data in a data-driven manner, which can enhance the stability of the model.

[0075] (5) In constructing the time sequence sequence prediction model cluster of multiple attribute characteristics, the preprocessed multiple attribute characteristic set is traversed and combined, an LSTM time sequence sequence prediction model is trained for each attribute combination, all attribute combinations are combined to form an LSTM model cluster, and each model in the cluster is responsible for processing different attribute characteristics or attribute characteristic combinations. This processing method can better utilize the complementary information between various attribute characteristics and improve the overall prediction ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0077] Figure 1 A flowchart of a battery application control method is provided. DETAILED DESCRIPTION

[0078] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0079] The present application provides a battery application control method, including battery business data acquisition, battery business data preprocessing, online selection of effective battery safety risk attribute characteristic set, label data expansion, output of battery safety risk prediction model, and further battery safety risk prediction. For base stations with high-risk batteries, intelligent peak-shaving strategy is no longer issued to ensure the safety of base station operation.

[0080] The present application dynamically constructs a multi-attribute characteristic time sequence sequence prediction model cluster based on multiple battery attribute characteristics, online selects the most valuable attribute characteristics for safety risk prediction, adapts to the safety risk prediction needs of different manufacturers, different models and different working conditions of batteries, and expands safety risk sample data using label propagation algorithm, deeply mines and fully utilizes unlabeled data, which can more accurately and stably predict the safety risk of lithium batteries.

[0081] The present application discloses a battery application control method, including the following steps:

[0082] Step 1: Obtain battery operation business data as a set of battery multi-attribute characteristics;

[0083] Step 2: Battery operation business data preprocessing, preprocessing includes but is not limited to deleting error values, duplicate values, replacing outliers, and filling missing values;

[0084] Step 3: Construct a time series sequence prediction model cluster of multi-attribute features, and select an effective set of battery safety risk attribute features based on the prediction accuracy of each model in the model cluster;

[0085] Step 3-1: Construct a time series sequence prediction model cluster of multi-attribute features;

[0086] Establish a time series sequence prediction model of multi-attribute features, the model input includes the sampling sequence of multiple attribute features of the battery at a specific time, the model output is the prediction of battery safety risk, and the time series sequence prediction model of multi-attribute features is trained and evaluated using the preprocessed safety risk sample or updated label data of the multi-attribute feature set;

[0087] Iterate the preprocessed multi-attribute feature set, train a time series sequence prediction model for each attribute feature combination, and combine all attribute feature combinations to form a time series sequence prediction model cluster of multi-attribute features;

[0088] Step 3-2: Online selection of an effective set of battery safety risk attribute features based on the prediction accuracy of each model in the time series sequence prediction model cluster of multi-attribute features;

[0089] Step 4: Label data expansion; based on the effective set of battery safety risk attribute features selected online, label the preprocessed multi-attribute feature set without label samples through a differential measurement algorithm and a label propagation algorithm, and obtain updated label data;

[0090] Step 5: Re-train the time series sequence prediction model cluster of multi-attribute features based on the updated label data, and iterate multiple rounds until the proportion of unlabeled samples is below a threshold, and output the battery safety risk prediction model;

[0091] Step 6: Battery safety risk prediction; when the battery operating condition data is collected by the battery control platform, it is input into the trained battery safety risk prediction model to obtain the corresponding safety risk prediction result, and the intelligent peak shaving strategy is not issued for the base station with high-risk batteries to ensure the safe operation of the base station.

[0092] Preferably, the battery operation business data in step 1 is divided into three types, namely battery operation data, battery state data and battery alarm data, as a set of multi-attribute features of the battery, the battery operation data includes voltage, temperature and humidity data during the battery charging and discharging process, the battery state data includes capacity, remaining capacity, maximum available capacity, internal resistance and life, and the battery alarm data includes voltage anomaly and temperature anomaly.

[0093] Preferably, the pre-processing of the battery operation business data in step 2 is specifically as follows:

[0094] For error values and repeated values, delete them;

[0095] For outliers, based on the box plot, replace the high outliers in the data with the third quartile and replace the low outliers in the data with the first quartile;

[0096] For missing values, use linear interpolation to fill in the missing values.

[0097] Preferably, step 3 is specifically as follows:

[0098] Step 3-1: Construct a set of time series prediction models of multi-attribute features;

[0099] Select long short-term memory network (LSTM) as the time series prediction model of multi-attribute features. When establishing the time series prediction model of multi-attribute features, first define the input and output of the model, the input includes the sampling sequence of multiple attribute features of the battery within a certain time, and the model output is the prediction of the safety risk of the battery. Divide the safety risk sample or updated label data of the pre-processed multi-attribute feature set into two parts according to the ratio of 8:2, which are training set and validation set. Train the time series prediction model of multi-attribute features, the training set is used to train the model, so that it can learn the time series relationship between attribute features and the leading indicator features of safety risk, and the validation set is used to evaluate the performance of the model.

[0100] Iterate and combine the pre-processed multi-attribute feature set, train an LSTM time series prediction model for each attribute feature combination, combine all the LSTM time series prediction models constructed by attribute feature combinations to form an LSTM model cluster, and each model in the cluster is responsible for processing different attribute features or attribute feature combinations.

[0101] Step 3-2: Online selection of effective attribute feature set representing battery safety risk

[0102] According to the model prediction accuracy of the time sequence prediction model of the multi-attribute feature, attribute features are selected, the contribution value of each attribute feature is calculated, and the top-k attribute features strongly related to the safety risk prediction are selected according to the size to obtain an attribute feature set effectively representing the battery safety risk, and the k attribute features are normalized.

[0103] Preferably, the contribution value of each attribute feature is calculated according to the following formula:

[0104]

[0105] wherein:

[0106] represents the prediction result;

[0107] represents the expected value of the prediction result;

[0108] v(S) represents the shap value of the attribute feature subset S;

[0109] represents the shap value of the new subset after adding the attribute feature i.

[0110] Preferably, the weight w[k] of the normalization processing is calculated according to the following formula:

[0111] w[k]=

[0112] wherein:

[0113] weight[k] represents the weight value of the kth attribute feature;

[0114] represents the sum of the weight values of all k attribute features.

[0115] Preferably, the step 4 is specifically:

[0116] Step 4-1: sample difference measurement;

[0117] In the battery safety risk prediction, DTW is used to compare whether the behavior patterns of two samples on a certain attribute feature are similar:

[0118] D(a,b)=Σ(w[k] DTW(a[k],b[k]))

[0119] wherein:

[0120] D(a,b) represents the difference measurement between sample a and sample b;

[0121] w[k] represents the normalized weight of the kth attribute feature;

[0122] DTW(a[k],b[k]) represents the dynamic time warping distance between sample a and sample b in the kth attribute feature representing the battery safety risk;

[0123] Step 4-2: label propagation

[0124] Initialization: assign an initial label to each unlabeled sample, usually by random assignment;

[0125] Iterative update: calculate the weight between each unlabeled sample and the surrounding labeled samples according to the difference measure D(a,b), the greater the weight, the more similar the two samples, and the greater the influence of label propagation;

[0126] Label propagation: update the label of the unlabeled sample according to the weight, which is realized by the weighted average aggregation method;

[0127] Convergence judgment: repeat the iterative update step until the convergence condition is met, the iteration number reaches the predetermined value or the label change is less than the threshold;

[0128] Label the unlabeled samples in the preprocessed multi-attribute feature set to obtain updated label data.

[0129] Preferably, step 6 is specifically: when a new battery operating condition sample is collected and transmitted to the system, input the trained battery safety risk prediction model to obtain the safety risk prediction result, and perform real-time safety risk prediction and analysis on the new battery operating condition sample. For base stations with high-risk batteries, the intelligent peak-shaving strategy is no longer issued to ensure the safety of base station operation.

[0130] Preferably, the battery application control method is executed by a battery control platform, which includes a basic function module, a battery health management module, a battery standby power analysis module, a battery life prediction module, an intelligent peak-shaving charging and discharging module, and a peak-shaving revenue calculation module.

[0131] The basic function module can obtain and view battery operation business data of the base station, including configuration, battery operation data, battery state data, and battery alarm data, and remotely control the battery;

[0132] The battery health management module can evaluate the health status of each group of batteries according to the battery operation business data to obtain the safety risk prediction result; support battery monitoring and management of the whole network / subnetwork / site, remind through notification and alarm, and provide related reports;

[0133] The battery backup analysis module can analyze the backup power capacity of a site or power supply system according to battery operation service data, load current and related alarms; support statistical analysis of the whole network / subnetwork / site; and provide backup power capacity analysis results.

[0134] The battery life prediction module can predict the remaining life of a battery according to the installation date of the battery, the environmental temperature, the battery type and the working condition data of the battery operation.

[0135] The intelligent peak-shaving charging and discharging module can remotely set the peak-shaving power consumption parameters of each battery group, enable and disable the peak-shaving power consumption function, query, set, modify and delete the peak-shaving power consumption mechanism, and realize low-price self-charging and high-price self-discharging.

[0136] The peak-shaving income calculation module can view the detailed information of the peak-shaving power consumption of each battery group in real time, count and analyze the peak-shaving power consumption data and give the results, and evaluate the income of the peak-shaving power consumption.

[0137] Brief introduction of the battery management and control platform of the application:

[0138] The battery management and control platform mainly includes the following functional modules:

[0139] 1) Basic functional module

[0140] The system can obtain and view the battery information of a base station, including configuration, battery operation data, battery state data and battery alarm data, and remotely control the battery.

[0141] 2) Battery health management module

[0142] The system can evaluate the health status of each battery group according to battery operation service data (voltage, current, SoH, SoC), obtain safety risk prediction results, support battery monitoring and management of the whole network / subnetwork / site, remind through notification and alarm, and provide related reports.

[0143] 3) Battery backup analysis module

[0144] The system can analyze the backup power capacity of a site or power supply system according to battery operation service data (voltage, current, SoH, SoC), load current and related alarms; support statistical analysis of the whole network / subnetwork / site; and provide backup power capacity analysis results.

[0145] 4) Battery life prediction module

[0146] The system can predict the remaining life of a battery according to the installation date of the battery, the environmental temperature, the battery type and the working condition data of the battery operation (for example, voltage, current, SoH, SoC, cycle number).

[0147] 5) Intelligent peak-shaving charging and discharging module

[0148] The system can remotely set the peak-shaving power consumption parameters of each group of batteries, enable and disable the peak-shaving power consumption function, query, set, modify, and delete the peak-shaving power consumption mechanism, and realize low-price self-charging and high-price self-discharging.

[0149] 6) Peak-shaving income estimation module

[0150] The system can view the detailed information of peak-shaving power consumption of each group of batteries in real time, count and analyze the peak-shaving power consumption data and give the results, and evaluate the income of peak-shaving power consumption.

[0151] Embodiment 1

[0152] The embodiment discloses a battery application management and control method, specifically comprising the following steps:

[0153] Step 1: Obtain battery operation business data

[0154] The battery operation business data is divided into three types:

[0155] (1) Battery operation data, such as voltage, temperature, humidity data during battery charging and discharging, etc.;

[0156] (2) Battery state data, such as capacity, remaining capacity, maximum available capacity, internal resistance, life, etc.;

[0157] (3) Battery alarm data: such as voltage anomaly, temperature anomaly, etc.;

[0158] Each type of data obtained as above is used as a battery multi-attribute feature set for modeling.

[0159] Step 2: Battery operation business data preprocessing

[0160] The quality of the data determines the effect of the prediction model, and is an important preparation work before model training. Through data preprocessing, some error values, duplicate values, missing values, and outliers can be avoided to have a negative impact on the construction of the model.

[0161] In this embodiment, for different types of data, the following processing is performed:

[0162] Error values and duplicate values: delete them;

[0163] Missing values: use linear interpolation to fill in missing values;

[0164] Outliers: based on the box plot, replace the high outliers in the data with the third quartile, and replace the low outliers in the data with the first quartile.

[0165] Step 3:

[0166] (3-1) Establish and train a set of multi-attribute feature time series prediction models

[0167] The multi-attribute feature time series sequence prediction model can adopt various architectures, and the embodiment selects a long short-term memory (LSTM) network. The LSTM is a variant of a recurrent neural network and is particularly suitable for processing sequence data with long-term dependencies. In battery safety risk prediction, the LSTM can capture the variation of battery multi-attribute features over time, thereby achieving early warning of safety risks.

[0168] In establishing the multi-attribute feature time series sequence prediction model, the input and output of the model are first defined. The input includes the sampling sequence of multiple key attribute features of the target lithium battery within a period of time, such as voltage, current, temperature, etc. The output is the prediction of the safety risk state of the battery, such as high, medium, and low safety risk. The preprocessed multi-attribute feature set of the safety risk sample is divided into two parts according to an 8:2 ratio, which are the training set and the validation set. The training set is used to train the model, so that it can learn the time series relationship between the attribute features and the leading indicator features of the safety risk. The validation set is used to evaluate the performance of the model.

[0169] The preprocessed multi-attribute feature set is iteratively combined. Each attribute combination trains an LSTM time series sequence prediction model. All attribute combinations are combined to form an LSTM model cluster. Each model in the cluster is responsible for processing different attribute features or attribute feature combinations. This processing method can better utilize the complementary information between various attribute features and improve the overall prediction ability of the model.

[0170] (3-2) Online selection of effective battery safety risk attribute feature set

[0171] In the LSTM neural network, each neuron has a weight value between connections, representing the strength of their association. These weight values are continuously updated and optimized through the backpropagation algorithm during the training process. Larger weight values usually mean that the corresponding attribute feature has a greater impact on the prediction ability of the model. From the battery multi-attribute feature sample data, a data-driven approach is used to extract high-weight original attribute features and eliminate invalid attribute features, which can enhance the stability of the model.

[0172] According to the model prediction accuracy of the multi-attribute feature time series sequence prediction model, attribute features are selected by calculating the contribution value of each original attribute feature and sorting them by size. The top-k attribute features that are strongly related to safety risk prediction can be selected. These k attribute features are considered the most critical part of the model and play an important role in the accuracy of safety risk prediction. The embodiment uses the shap algorithm to analyze the key attribute features that affect battery safety risk. For each attribute feature, the contribution value to the prediction result is:

[0173]

[0174] where f(x) represents the prediction result, E[f(x)] represents the expected value of the prediction result, and v(S) represents the shap value of the attribute feature subset S; represents the shap value of the new subset after adding the attribute feature i.

[0175] Normalization processing can make the contribution values of each attribute feature on a unified scale, facilitating subsequent analysis and comparison. At the same time, the normalized weight w[k] can also reflect the relative importance of each attribute feature to the safety risk prediction, which is helpful for further optimizing the model and extracting key features. Specifically, the weight value of each attribute feature can be divided by the sum of the weight values of all k attribute features to obtain the normalized weight w[k].

[0176] The calculation formula of the normalized weight w[k] is as follows:

[0177] w[k]=

[0178] where weight[k] represents the weight value of the kth attribute feature; represents the sum of the weight values of all k attribute features.

[0179] Step 4: Label data augmentation

[0180] (1) Sample difference measurement

[0181] Sample difference measurement refers to designing a method to quantify the difference between two lithium battery safety risk samples, which needs to consider multiple attribute features and weight them according to their importance. DTW is a method for measuring the similarity between two sequences, especially suitable for handling time series data with different lengths or varying rates. In lithium battery safety risk prediction, DTW can be used to compare whether the behavior patterns of two samples on a certain attribute feature are similar.

[0182] D(a,b)=Σ(w[k] DTW(a[k],b[k]))

[0183] where D(a,b) represents the difference measurement between sample a and sample b; w[k] represents the normalized weight of the kth attribute feature; and DTW(a[k],b[k]) represents the dynamic time warping (DTW) distance between sample a and sample b on the kth attribute feature that effectively represents battery safety risk.

[0184] (2) Label propagation

[0185] The unlabeled training samples are labeled using the previously proposed sample dissimilarity measure method D(a, b). Label propagation is a semi-supervised learning technique that enhances data and improves model performance by transferring known label information to unlabeled samples. Specifically, for an unlabeled training sample a, we can infer its possible label by calculating the dissimilarity measure D(a, b) between it and other labeled samples b, achieving data sample enhancement, where the dissimilarity measure uses the online selected effective representation battery safety risk attribute feature set obtained in step 3.

[0186] The specific process is as follows:

[0187] Initialization: Assign an initial label to each unlabeled sample, usually by random assignment or based on certain heuristic rules.

[0188] Iterative update: Calculate the weight between each unlabeled sample and the surrounding labeled samples according to the dissimilarity measure D(a, b). The greater the weight, the more similar the two samples, and the greater the impact of label transfer.

[0189] Label propagation: Update the label of the unlabeled sample according to the weight. This process is achieved through a weighted average aggregation method.

[0190] Convergence judgment: Repeat the iterative update step until the convergence condition is met (such as the number of iterations reaches a predetermined value or the label change is less than a certain threshold).

[0191] In this way, the unlabeled training samples are labeled, achieving label data expansion. These labeled samples can be added to the training set for further training and optimization of the model. Data augmentation can increase the number and diversity of training data, improving the generalization ability and robustness of the model.

[0192] Step 5: Retrain the time series prediction model cluster, iterate multiple times until the proportion of unlabeled samples is below the threshold, and output the battery safety risk prediction model.

[0193] Step 6: Battery safety risk prediction

[0194] When the battery control platform collects battery operating condition data, it needs to perform safety risk prediction and analysis. The data collection process usually involves real-time monitoring and recording of multiple key attribute features, such as battery voltage, current, temperature, etc.

[0195] The trained battery safety risk prediction model is inputted to obtain the corresponding safety risk prediction result, and real-time safety risk prediction and analysis are performed on the new lithium battery working condition sample. The base station with high-risk battery is no longer issued under the intelligent peak-shaving strategy to ensure the safety of the base station operation.

[0196] The battery safety risk prediction model outputs the prediction results as high-risk, medium-risk, and low-risk probabilities. The highest probability that is greater than a preset value is the battery risk prediction level. For example, if the high-risk probability is 60%, the medium-risk probability is 30%, and the low-risk probability is 10%, the final output result is high-risk.

[0197] Example 2

[0198] From September to November 2024, battery and data collection equipment were deployed in the existing machine room of Shaanxi Branch of China Tower Corporation, and management platform software was deployed, as follows:

[0199] Pilot Station 1: Xianyang Qindu Biyang Road West

[0200] On-site working condition: site load about 92.9A, working voltage: 54.4V;

[0201] Number of power supplies: one set of Vidy multi-tenant power supply, 6 50A rectifiers;

[0202] ZTE intelligent lithium battery: 3 groups of ZTE intelligent lithium battery ZXDC48FB100B3 are configured;

[0203] Other battery packs: 5 groups of Shengyang ordinary lithium batteries.

[0204] Pilot Station 2: Xianyang Weicheng 202 Institute

[0205] On-site working condition: site load about 80A, working voltage: 55.1V;

[0206] Number of power supplies: one set of Huawei multi-tenant power supply, 6 50A rectifiers;

[0207] ZTE intelligent lithium battery: 4 groups of ZTE intelligent lithium battery ZXDC48FB100B3 are configured;

[0208] Other battery packs: 1 group of Planar 200AH ordinary lithium batteries.

[0209] The utility conditions are shown in the following table:

[0210]

[0211] According to the test conditions of the pilot stations, the following data are obtained:

[0212] Lithium battery charging and discharging efficiency = discharge capacity / charging capacity ≈ 92%;

[0213] The Qin City Bijuang Road West Station has a high site load, with an average load of about 91A. The 3 sets of intelligent lithium batteries stop discharging when the discharge depth reaches 50%, and the actual single discharge is less than 2 hours;

[0214] The Weicheng 202 Institute has a low site load, with an average load of about 76A. The 4 sets of intelligent lithium batteries are discharged according to 2 hours, and the actual discharge depth does not reach 50%.

[0215] The economic benefit indicators are calculated as follows:

[0216] Scheme One: The station already has ZTE intelligent lithium batteries, and no additional batteries are added (calculated based on 2 sets of intelligent lithium batteries per station)

[0217] The station's intelligent lithium battery is used for peak shifting, with a discharge depth of 50%, a charge-discharge efficiency of 92%, and one charge and one discharge;

[0218] Single battery annual peak shifting income = 48V 100Ah / 1000 50% (1.1 yuan / kW·h-(0.38 yuan / kW·h / 92%)) 365=602 yuan;

[0219] Single station investment: BCUA purchase and installation about 1250 yuan, network management platform construction and operation and maintenance cost allocation 200 yuan / station / year, single station initial investment 1250 yuan, contract period operation and maintenance cost 1000 yuan;

[0220] According to the 5-year contract period, the static recovery period = 1250 yuan ÷ (602 2 5-(1250+1000)+1250) 10 years ≈ 2.49 years.

[0221] Scheme Two: The station has no ZTE intelligent lithium batteries, and additional batteries are added (calculated based on 2 sets of intelligent lithium batteries added per station)

[0222] The station adds new recycled intelligent lithium batteries for peak shifting, with a cycle number of not less than 6000 times, configured according to 2 hours of discharge, a discharge depth of 90%, a charge-discharge efficiency of 95%, and two charges and two discharges;

[0223] Single battery annual peak shifting income = 48V 100Ah / 1000 90% (1.1 yuan / kW·h-(0.38 yuan / kW·h / 92%)+1.1 yuan / kW·h-(0.7 yuan / kW·h / 92%) 365=1618 yuan;

[0224] Single station investment: new purchase of 5500 yuan per cycle of intelligent lithium battery, new power module 500 yuan per piece, battery installation cost about 720 yuan per set, power module installation about 400 yuan per piece, BCUA purchase and installation about 1250 yuan per station, network management platform construction and operation cost 200 yuan per station per year, station new battery and rectifier module per station operation and maintenance cost 400 yuan per station per year, according to the single station new 2 cycle of intelligent lithium battery, while increasing 2 power modules, the initial investment cost of single station is about 15490 yuan, and the operation and maintenance cost in the contract period is about 6000 yuan;

[0225] According to the 10-year project cycle, the static recovery period = 15490 yuan ÷ (1618 2 10-(15490+6000)+15490) 10 years ≈ 5.88 years.

[0226] The application discloses a battery application management and control method, which dynamically constructs a time sequence prediction model cluster of multiple attribute characteristics based on multiple battery attribute characteristics, selects attribute characteristics most valuable for safety risk prediction online, adapts to safety risk prediction needs of different manufacturers, different models and different working conditions of batteries, and expands safety risk sample data by using a label propagation algorithm, deeply mines and fully utilizes unlabeled data, so that the safety risk of the lithium battery can be more accurately and stably predicted.

[0227] The application constructs a time sequence prediction model cluster based on multiple battery attribute characteristics, selects multiple attribute leading indication characteristics effectively representing battery safety risks online, enhances model robustness, and solves the problem of difficult determination of effective attribute characteristics.

[0228] The application expands risk sample data based on the multiple attribute leading indication characteristics effectively representing battery safety risks obtained by online selection and the label propagation algorithm, iteratively trains the battery safety risk prediction model, improves the model generalization ability and robustness, and solves the problem of small sample size of battery safety risks.

[0229] The application extracts high-weight original attribute characteristics and eliminates invalid attribute characteristics from battery multi-attribute characteristic sample data in a data-driven manner, which can enhance the stability of the model.

[0230] In constructing the time sequence sequence prediction model cluster of the multi-attribute feature, the preprocessed multi-attribute feature set is traversed and combined, an LSTM time sequence sequence prediction model is trained for each attribute combination, all the LSTM models constructed by the attribute combinations are combined to form an LSTM model cluster, and each model in the cluster is responsible for processing different attribute features or attribute feature combinations. This processing method can better utilize the complementary information between various attribute features and improve the overall prediction ability of the model.

[0231] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A battery application management and control method, characterized in that, The method comprises the following steps: Step 1: obtaining battery operation business data as a battery multi-attribute feature set; Step 2: battery operation business data preprocessing, preprocessing including deleting error values, duplicate values, replacing outliers and filling missing values; Step 3: constructing a time series sequence prediction model cluster of multi-attribute features, selecting an effective battery safety risk attribute feature set based on the prediction accuracy of each model in the model cluster; Step 3-1: constructing a time series sequence prediction model cluster of multi-attribute features; A time series sequence prediction model of multi-attribute features is established, the model input includes the sampling sequence of multiple attribute features of the battery at a specific time, the model output is the battery safety risk prediction, and the time series sequence prediction model of multi-attribute features is trained and evaluated using the safety risk sample or updated label data of the preprocessed multi-attribute feature set; The preprocessed multi-attribute feature set is traversed and combined, a time series sequence prediction model is trained for each attribute feature combination, and all attribute feature combinations are combined to form a time series sequence prediction model cluster of multi-attribute features; Step 3-2: selecting an effective battery safety risk attribute feature set based on the prediction accuracy of each model in the time series sequence prediction model cluster of multi-attribute features; Step 4: label data expansion; based on the effective battery safety risk attribute feature set selected online, the label data of the preprocessed multi-attribute feature set without label sample is labeled by using a differential measurement algorithm and a label propagation algorithm, and updated label data is obtained; Step 5: retraining the time series sequence prediction model cluster of multi-attribute features based on the updated label data, and after multiple iterations, the proportion of label-free samples is lower than the threshold, and the battery safety risk prediction model is output; Step 6: battery safety risk prediction; When the battery control platform collects battery working condition data, it is input into the trained battery safety risk prediction model to obtain the corresponding safety risk prediction result, and the intelligent peak-shaving strategy is not issued to the base station with high-risk batteries to ensure the safe operation of the base station.

2. The method of claim 1, wherein, The battery operation business data in step 1 is divided into three types, namely battery operation data, battery state data and battery alarm data, which are used as the battery multi-attribute feature set. The battery operation data includes voltage, temperature and humidity data during battery charging and discharging. The battery state data includes capacity, remaining capacity, maximum available capacity, internal resistance and life. The battery alarm data includes voltage anomaly and temperature anomaly.

3. The method of claim 1, wherein, In step 2, the battery operation business data preprocessing is as follows: For error values and duplicate values, delete them; For outliers, based on the box plot, replace the high outliers in the data with the third and fourth quartiles, and replace the low outliers in the data with the first quartile; For missing values, use linear interpolation to fill in missing values.

4. The method of claim 1, wherein, Step 3 is as follows: Step 3-1: constructing a time series sequence prediction model cluster of multi-attribute features; Long Short Term Memory network (LSTM) is selected as a time series prediction model of multi-attribute features. In establishing the time series prediction model of multi-attribute features, the input and output of the model are defined first. The input includes the sampling sequence of multiple attribute features of the battery at a specific time, and the output of the model is the prediction of the safety risk of the battery. The safety risk samples of the preprocessed multi-attribute feature set or the updated label data are divided into two parts according to the ratio of 8:2, which are the training set and the validation set respectively. The time series prediction model of multi-attribute features is trained. The training set is used to train the model, so that it can learn the time series relationship between the attribute features and the leading indicator features of the safety risk. The validation set is used to evaluate the performance of the model. The preprocessed multi-attribute feature set is iteratively combined. An LSTM time series prediction model is trained for each attribute feature combination. The LSTM time series prediction models constructed by all attribute feature combinations are combined together to form an LSTM model cluster. Each model in the cluster is responsible for processing different attribute features or attribute feature combinations. Step 3-2: Online selection of attribute feature set effectively representing battery safety risk According to the model prediction accuracy of the time series prediction model of multi-attribute features, the contribution value of each attribute feature is calculated, and the top-k attribute features with strong correlation with safety risk prediction are selected according to the size, to obtain the attribute feature set effectively representing the battery safety risk, and the k attribute features are normalized.

5. The method of claim 4, wherein, The calculation formula of the contribution value of each attribute feature is as follows: wherein: represents the prediction result; denotes the expected value of the prediction result; v(S) represents the shap value of the attribute feature subset S. shap value representing the new subset after adding attribute feature i.

6. The method of claim 4, wherein, The calculation formula of the weight w[k] of the normalization processing is as follows: w[k]= wherein: weight[k] represents the weight value of the kth attribute feature. represents the sum of all k attribute feature weight values.

7. The method of claim 6, wherein, Step 4 is specifically: Step 4-1: Sample difference measurement In battery safety risk prediction, DTW is used to compare whether the behavior patterns of two samples on a certain attribute feature are similar: D(a,b) =∑(w[k]DTW(a[k],b[k])) DTW(a[k],b[k])) wherein: D(a,b) represents the difference measurement between sample a and sample b; w[k] represents the normalized weight of the kth attribute feature; DTW(a[k],b[k]) represents the dynamic time warping distance between sample a and sample b on the kth attribute feature effectively representing the battery safety risk. Step 4-2: Label propagation Initialization: an initial label is assigned to each unlabeled sample, usually by random assignment; Iterative update: according to the difference measurement D(a,b) between samples, the weight between each unlabeled sample and the surrounding labeled sample is calculated. The greater the weight, the more similar the two samples are, and the greater the influence of label transmission. Label propagation: update the label of the unlabeled sample according to the weight. This process is realized by weighted average aggregation method; Convergence judgment: repeat the iterative update step until the convergence condition is met, the iteration number reaches the predetermined value, or the label change is less than the threshold value; Label the unlabeled samples in the preprocessed multi-attribute feature set to obtain the updated label data.

8. The method of claim 1, wherein, The step 6 is specifically: when a new battery working condition sample is collected and transmitted to the system, the trained battery safety risk prediction model is inputted to obtain a safety risk prediction result, real-time safety risk prediction and analysis are performed on the new battery working condition sample, and the intelligent peak-shaving strategy is no longer issued to the base station with high-risk batteries to ensure the safe operation of the base station.

9. The method of claim 1, wherein, The battery application management and control method is executed by a battery management and control platform, and the battery management and control platform comprises a basic function module, a battery health management module, a battery backup power analysis module, a battery life prediction module, an intelligent peak-shaving charging and discharging module, and a peak-shaving income calculation module. The basic function module can acquire and view battery operation service data of the base station, including configuration, battery operation data, battery state data, and battery alarm data, and remotely control the battery. The battery health management module can evaluate the health status of each group of batteries according to the battery operation service data to obtain a safety risk prediction result, support battery monitoring and management of the whole network / sub-network / site, remind through notification and alarm, and provide related reports. The battery backup power analysis module can analyze the backup power capacity of the site or power supply system according to the battery operation service data, load current, and related alarms, support statistical analysis of the whole network / sub-network / site, and provide backup power capacity analysis results. The battery life prediction module can predict the remaining life of the battery according to the battery installation date, environmental temperature, battery type, and battery operation condition data. The intelligent peak-shaving charging and discharging module can remotely set the peak-shaving power consumption parameters of each group of batteries, enable and disable the peak-shaving power consumption function, query, set, modify, and delete the peak-shaving power consumption mechanism, and realize low-price self-charging and high-price self-discharging. The peak-shaving income calculation module can view the detailed information of the peak-shaving power consumption of each group of batteries in real time, statistically analyze the peak-shaving power consumption data, and give the results to evaluate the income of the peak-shaving power consumption.

Images