Energy storage in-situ monitoring method and system for energy storage power station

Abstract

The present application relates to the technical field of energy storage system, and particularly relates to an energy storage power station energy storage on-site monitoring method and system, which is based on pre-constructed equipment semantic meta-models of various types of equipment of the energy storage power station, and a unified protocol analysis table and semantic mapping table of various types of equipment of the energy storage power station; including realizing multi-source heterogeneous monitoring data collection and unified formatting: collecting data streams generated by the equipment in real time through the equipment communication interface for each equipment of the energy storage power station; analyzing original fields in each data stream in real time based on the protocol analysis table; mapping each original field analyzed based on the semantic mapping table to a unified standard field according to the equipment semantic meta-models; performing time-space alignment resampling on the mapped standard field, obtaining standard structured data of each equipment of the energy storage power station and outputting. The present application is used to solve the problems of weak multi-source data interaction ability, information island and difficulty in coping with complex operation trends based on a general judgment mechanism based on static threshold.

Description

Technical Field

[0001] This invention relates to the field of energy storage system technology, and specifically to a method and system for on-site monitoring of energy storage in an energy storage power station. Background Technology

[0002] With the large-scale integration of renewable energy, energy storage power stations have been widely used in scenarios such as grid peak shaving, valley filling, and microgrid regulation. To ensure the operational efficiency and safety of energy storage power stations, multiple monitoring devices are typically deployed inside the station, such as battery management systems (BMS), data acquisition modules, environmental sensors, and fire protection systems.

[0003] Although existing energy storage power stations are equipped with local station-level monitoring systems for data acquisition and monitoring, they only perform functions such as data aggregation, and their functionality is still relatively weak. Currently, the following problems exist:

[0004] 1) The data interaction capabilities of various monitoring devices within the system are weak, resulting in information silos;

[0005] 2) Judgment mechanisms are generally based on static thresholds, which are difficult to cope with complex operational trends;

[0006] 3) AI diagnostics are deployed in the cloud, which is highly dependent on communication, lacks real-time performance and stability, and cannot respond locally. Summary of the Invention

[0007] To address the above shortcomings, this invention provides a method and system for on-site monitoring of energy storage in an energy storage power station, thereby solving at least one of the aforementioned problems.

[0008] In a first aspect, the present invention provides a method for on-site monitoring of energy storage in an energy storage power station. The method is based on a pre-constructed device semantic meta-model of various types of equipment in the energy storage power station, as well as a unified protocol parsing table and semantic mapping table for various types of equipment in the energy storage power station; the device semantic meta-model includes a standard field dictionary.

[0009] The methods include:

[0010] S1. Acquisition and unified formatting of multi-source heterogeneous monitoring data:

[0011] Real-time data streams generated by each device in the energy storage power station are collected through the device communication interface;

[0012] Real-time, based on the protocol parsing table, the raw fields in each collected data stream are parsed in a structured manner;

[0013] In real time, based on the semantic mapping table, the parsed original fields are mapped to unified standard fields according to the standard field dictionary of the device semantic meta-model, so as to obtain the standard fields of each device in the energy storage power station.

[0014] By employing a master clock anchor point and sliding time window strategy, the standard fields of each device in the energy storage power station are spatiotemporally aligned and resampled to obtain and output the standard structured data of each device in the energy storage power station.

[0015] S2. Perform data fusion and feature extraction on the standard structured data of each device output in S1 to obtain the structured feature vector of each device in the energy storage power station.

[0016] S3. Edge-side intelligent diagnosis: The structured feature vectors of each device in the energy storage power station are input into the pre-trained multi-task diagnostic model deployed on the local diagnostic node. After model calculation, the current health status, current fault trend and current fault risk level of the energy storage power station are predicted.

[0017] Furthermore, S2 includes:

[0018] For each device in the energy storage power station, a unified sliding time window is adopted to statistically analyze the continuous characteristics and discrete state characteristics of the standard structured data of each device output in S1. Then, the key derived quantities are derived by combining physical formulas and device logic rules, and the derived key derived quantities are collected as multi-dimensional aggregated features.

[0019] For the standard structured data of each device in the energy storage power station within the sliding time window, based on the collaborative relationship between the devices in the energy storage power station, the standard structured data is multidimensionally fused through the field association strategy. Then, based on the multidimensional fusion, the cross-device coupling behavior between the devices in the energy storage power station is analyzed, and key joint features are extracted from it.

[0020] By constructing a feature configuration template, the multidimensional aggregated features and key joint features obtained under the sliding time window are uniformly encoded and fixed-length structured to obtain the structured feature vectors of each device in the energy storage power station.

[0021] S3, Edge-side Intelligent Diagnosis:

[0022] The structured feature vectors of each device in the energy storage power station are input into a pre-trained multi-task diagnostic model deployed on the local diagnostic node. After model calculation, the current health status, current fault trend, and current fault risk level of the energy storage power station are predicted.

[0023] Furthermore, the method also includes:

[0024] S4. Based on the predicted current health status, current failure trend, and current failure risk level, trigger the corresponding local response strategy.

[0025] Furthermore, the method for obtaining the pre-trained multi-task diagnostic model in S3 includes:

[0026] S31. Construct the training set;

[0027] S32. Construct the network architecture for a multi-task diagnostic model;

[0028] S33. Train the network architecture using the training set to obtain a trained multi-task diagnostic model;

[0029] In S31, the methods for constructing the training set include:

[0030] S311. Collect historical operating data of all equipment in the energy storage power station throughout its entire life cycle;

[0031] S312. Using the same method as S1 and S2, process the collected historical operation data of each device in the energy storage power station throughout its entire life cycle to obtain several historical structured feature vectors of the energy storage power station.

[0032] S313. Each historical structured feature vector is labeled. The labels include the energy storage power station health status label, fault trend label, and fault risk score label. The health status label includes "101", "102", and "103", which represent normal, slightly abnormal, and severely abnormal, respectively. The fault trend label includes "201", "202", and "203", which represent fault increase, fault decrease, and stable trend, respectively. The risk score label includes "301", "302", "303", "304", "305", and "306", which correspond to the score ranges [0,20), [20,40), [40,60), [60,80), [80,95), and [95,100], respectively, representing extremely low risk, low risk, medium risk, medium-high risk, high risk, and extremely high risk.

[0033] The labeled historical structured feature vectors are collected and sorted in ascending order according to the data collection time to form a training set.

[0034] Furthermore, the network architecture of the multi-task diagnostic model includes:

[0035] The input layer is used to input structured feature vectors, providing input for subsequent feature extraction;

[0036] A shared hidden layer, connected to the input layer, is used to extract a global representation of the structured feature vector of the input.

[0037] The task branch layer, connected to the shared hidden layer, specifically includes a state classification branch, a trend prediction time series branch, and a risk scoring branch. The state classification branch is used to predict the probability distribution of the health state categories of the energy storage power station; the trend prediction time series branch is used to predict the state change of the energy storage power station; and the risk scoring branch is used to predict the continuous distribution of the risk score of the energy storage power station.

[0038] The output layer includes a first softmax function layer, a second softmax function layer, and a Linear layer. The first softmax function layer is connected to the output of the state classification branch and is used to obtain and output the health state classification corresponding to the energy storage power station based on the output of the state classification branch. The second softmax function layer is connected to the output of the trend prediction time series branch and is used to obtain and output the fault trend classification corresponding to the energy storage power station based on the output of the trend prediction time series branch. The Linear layer is connected to the output of the risk scoring branch and is used to map the output of the risk scoring branch to the corresponding risk scoring classification and output it.

[0039] S33. Train the network architecture using the training set to obtain a trained multi-task diagnostic model.

[0040] Furthermore, S33 includes:

[0041] S331. Initialize the network architecture parameters to obtain the initial model;

[0042] S332. Input the training set into the initial model for forward propagation iterative training. In each iteration, calculate the loss of each task branch, and then calculate the total loss by fusing the losses of each branch through dynamic weights. Among them, the state classification branch adopts the cross-entropy loss function, the trend prediction time series branch adopts the MSE loss function, and the risk scoring branch adopts the MSE loss function.

[0043] S333. Backpropagate to update the model parameters until the total loss converges or the preset number of training iterations are reached, and a well-trained multi-task diagnostic model is obtained.

[0044] Furthermore, the device semantic meta-model also includes a device semantic processing module;

[0045] The device semantic processing module is used to perform standard field derivation calculation, field addition, device registration, and semantic conflict handling in the standard field dictionary.

[0046] Furthermore, when there are multiple local diagnostic nodes as described in S3, the trained multi-task diagnostic model is deployed on each local diagnostic node. S3 includes:

[0047] The structured feature vectors of each device in the energy storage power station are input into the trained multi-task diagnostic model deployed on each local diagnostic node.

[0048] The trained multi-task diagnostic model deployed on each local diagnostic node performs calculations on the input structured feature vector to obtain the predicted current health status, current fault trend, and current fault risk level of the energy storage power station, and sends them to the collaborative judgment module in a structured form.

[0049] The collaborative judgment module forms a candidate result set from the prediction results sent by each local diagnostic node. For the same type of abnormal event, the final prediction result is determined by majority voting or weighted voting mechanism.

[0050] Furthermore, the method also includes:

[0051] After the local response strategy is triggered and executed, the changes of preset target fields of each device in the energy storage power station are continuously monitored for a preset time period, the degree of improvement after the strategy response is calculated, and a feedback score is generated based on this.

[0052] The control strength of the local response strategy is adjusted based on feedback scores, forming a self-closing response optimization mechanism.

[0053] Secondly, this invention provides an on-site monitoring system for energy storage power stations. The system includes a pre-constructed device semantic meta-model for various types of equipment in the energy storage power station, as well as a unified protocol parsing table and semantic mapping table for various types of equipment in the energy storage power station. The device semantic meta-model includes a standard field dictionary.

[0054] The system also includes:

[0055] The monitoring data acquisition module is used for the acquisition and unified formatting of multi-source heterogeneous monitoring data in energy storage power stations.

[0056] The monitoring data acquisition module is configured to perform the following steps:

[0057] Real-time data streams generated by each device in the energy storage power station are collected through the device communication interface;

[0058] Real-time, based on the protocol parsing table, the raw fields in each collected data stream are parsed in a structured manner;

[0059] In real time, based on the semantic mapping table, the parsed original fields are mapped to unified standard fields according to the standard field dictionary of the device semantic meta-model, so as to obtain the standard fields of each device in the energy storage power station.

[0060] By employing a master clock anchor point and sliding time window strategy, the standard fields of each device in the energy storage power station are spatiotemporally aligned and resampled to obtain and output the standard structured data of each device in the energy storage power station.

[0061] The system also includes:

[0062] The feature extraction module is used to perform data fusion and feature extraction on the standard structured data of each device output by the monitoring data acquisition module to obtain the structured feature vectors of each device in the energy storage power station.

[0063] The edge-side intelligent diagnostic module is used to input the structured feature vectors of each device in the energy storage power station into a pre-trained multi-task diagnostic model deployed on the local diagnostic node. After model calculation, the current health status, current fault trend, and current fault risk level of the energy storage power station are predicted.

[0064] As can be seen from the above technical solutions, the present invention has the following advantages:

[0065] This invention constructs a device semantic meta-model, a protocol parsing table, and a semantic mapping table, which can adapt to device data from different manufacturers and using different protocols, eliminating the problem of weak interaction between multi-source data and the existence of information silos. Furthermore, compared to traditional methods that rely on static mapping relationships, this mechanism has stronger scalability and adaptability, making it suitable for the data integration needs of complex heterogeneous systems.

[0066] This invention deploys diagnostic models at edge nodes. Compared with traditional solutions that rely on cloud inference, this mechanism reduces the dependence on cloud communication, enhances real-time performance and stability, and facilitates local response. Detailed Implementation

[0067] The present invention will now be described in detail. Specific details, such as particular system structures and techniques, are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will recognize that the invention may be implemented in other embodiments without these specific details.

[0068] The key terms used in this invention will be explained below.

[0069] BMS: Battery Management System.

[0070] PCS: Energy Storage Converter.

[0071] SOC: System On Chip, a system-on-a-chip or system-on-a-chip.

[0072] PCCK: Pack Control Circuit Kit.

[0073] It should be understood that, when used in this specification, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0074] The terms "one embodiment" or "some embodiments" used in this invention mean that one or more embodiments of the invention include the specific features, structures, or characteristics described in that embodiment. Therefore, phrases such as "in some embodiments" or "in other embodiments" appearing in different parts of the invention do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0075] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] This embodiment provides a method for on-site monitoring of energy storage in an energy storage power station, the execution entity of which can be an on-site monitoring system for the energy storage power station. The on-site monitoring method for the energy storage power station provided in this embodiment is executed by computer equipment; correspondingly, the on-site monitoring system for the energy storage power station runs on the computer equipment.

[0077] In one embodiment of the present invention, the method is based on a pre-constructed device semantic meta-model of various types of equipment in an energy storage power station, as well as a unified protocol parsing table and semantic mapping table for various types of equipment in an energy storage power station.

[0078] Construction and registration of device semantic meta-models: Based on industry standards and relevant device protocol documents, abstract and construct device semantic meta-models for various types of equipment in energy storage power stations (such as BMS, PCS, liquid cooling, etc.).

[0079] The device semantic metamodel includes a standard field dictionary.

[0080] The standard field dictionary defines the core parameters of various devices in an energy storage power station. Specifically, the standard field dictionary defines the core parameters of the devices, including standard field names (including field name, physical meaning, standard unit, data type, etc.), field hierarchical structure (cell, PCCK, cluster), field type (state quantity, control quantity, environmental quantity), source device identifier, etc., as the basis for subsequent device semantic mapping and data normalization.

[0081] Meanwhile, the field dictionary supports a scope constraint mechanism to specify the validity of a field under specific states or conditions, and supports a dynamic semantic annotation mechanism that can dynamically adjust the semantic interpretation of a field based on the device's operating status, scene labels, or environmental variables, so that the same field has differentiated semantic expressions in different contexts.

[0082] For various devices in energy storage power stations (such as BMS, PCS, liquid chillers, fire protection systems, dehumidifiers, and meters), a unified "protocol parsing-semantic mapping table" is constructed to realize the decoding and semantic binding of device data.

[0083] The "protocol parsing-semantic mapping table" refers to both the protocol parsing table and the semantic mapping table.

[0084] The protocol parsing table is used to structure and parse the raw fields of the data streams generated by various devices in the energy storage power station. Specifically, it describes the physical location, data type, offset, unit conversion, and other rules of each field in the raw message. This allows for structured parsing of the raw data streams collected through the communication interfaces of various devices in the energy storage power station, such as Modbus TCP / RTU and CAN, to extract raw fields such as field name, actual value, timestamp, and source device ID.

[0085] The semantic mapping table is used to map the parsed raw fields to unified standard fields based on the standard field dictionary of the device semantic metamodel, achieving field semantic alignment and standardized reconstruction. This includes unified processing of field names, units, dimensions, precision, data ranges, etc. This process supports preserving the original field names and contextual information for subsequent data fusion, tracing, and consistency analysis.

[0086] The method in this embodiment includes step S1: multi-source heterogeneous monitoring data acquisition and unified formatting.

[0087] This step is used to achieve data semantic unification, providing a high-quality, dimensionally unified data foundation for downstream fusion analysis.

[0088] Specifically, step S1 includes:

[0089] Step S11: Collect the data streams generated by each device in the energy storage power station in real time through the device communication interface (such as Modbus TCP / RTU, CAN, etc.);

[0090] Step S12: Real-time structured parsing of the original fields in each collected data stream based on the protocol parsing table;

[0091] Step S13: In real time, based on the semantic mapping table, the parsed original fields are mapped to unified standard fields according to the standard field dictionary of the device semantic meta-model, so as to obtain the standard fields of each device in the energy storage power station.

[0092] Step S14: Using the master clock anchor point and sliding time window strategy, the standard fields of each device in the energy storage power station are spatiotemporally aligned and resampled to obtain the standard structured data of each device in the energy storage power station and output it.

[0093] In step S1, the reconstructed data is semantically consistent, formatted, and spatiotemporally aligned. The output format is JSON, and local synchronous caching is also supported.

[0094] In the specific implementation of step S14, a sliding time window is constructed using the BMS or another high-frequency device in the energy storage power station as the time anchor point. The parsed data from each device is projected into the target window according to the timestamp, generating a snapshot of the entire station's status at a unified sampling time. For fields not within the time window, interpolation, preserving previous values, or filling with default values ​​are used to fill in the missing fields based on their characteristics.

[0095] In addition, in the actual implementation, depending on the actual situation, you can choose to add a time integrity label to each frame of data and generate a data quality score (such as whether it contains interpolation fields, the proportion of missing fields, etc.) to ensure the consistency of the "unified time state" in downstream semantic calculation, state evaluation and other modules.

[0096] In this embodiment, the method further includes:

[0097] Step S2: Data fusion and feature extraction;

[0098] Step S3: Edge-side intelligent diagnosis.

[0099] In step S1, the standard structured data of each device in the energy storage power station is output to step S2. Step S2 performs data fusion and feature extraction to obtain the structured feature vector of each device in the energy storage power station.

[0100] In the method of this embodiment, step S3 includes:

[0101] The structured feature vectors of each device in the energy storage power station obtained in step S2 are input into the pre-trained multi-task diagnostic model deployed on the local diagnostic node. After model calculation, the predicted current health status, current fault trend, and current fault risk level of the energy storage power station are obtained.

[0102] Step S2 collects the multi-source data output in step S1, and performs information extraction and fusion processing on the collected multi-source data to mine potential risk signals and generate high-quality feature vectors that can be used for model calculation.

[0103] Specifically, step S2 includes:

[0104] Step S21: For each device in the energy storage power station, a unified sliding time window is used to statistically analyze the continuous quantity statistical characteristics and discrete state quantity behavioral characteristics in the standard structured data of each device output in step S1. Then, the key derived quantities are derived by combining physical formulas and device logic rules, and the derived key derived quantities are collected as multi-dimensional aggregated features.

[0105] Step S22: For the standard structured data of each device in the energy storage power station within the sliding time window, based on the collaborative relationship between the devices in the energy storage power station, multi-dimensional fusion of the standard structured data is carried out through the field association strategy. Then, based on the multi-dimensional fusion, the cross-device coupling behavior between the devices in the energy storage power station is analyzed, and then key joint features are extracted from them.

[0106] Step S23: Using the constructed feature configuration template, the multidimensional aggregated features and key joint features obtained under the sliding time window are uniformly encoded and fixed-length structured to obtain the structured feature vectors of each device in the energy storage power station.

[0107] Understandably, step S21 employs field aggregation and derived feature calculation to derive more valuable statistical quantities and behavioral features from a large number of basic fields, thereby obtaining a high-dimensional but sparse statistical feature vector. Specifically: based on setting a sliding time window, the statistical characteristics of continuous quantities (including cell voltage, current, temperature, SOC, PCS power, and ambient temperature and humidity) in the standard structured data of each device output in step S1 are statistically analyzed to characterize the stability and changing trends of device operation; at the same time, the behavioral characteristics (including switching frequency, duration, and proportion) of discrete state quantities (including dehumidification, cooling, discharge, and fire protection) in the standard structured data of each device output in step S1 are identified to reflect their working status and response mode; then, based on physical formulas (such as P=U×I, energy integral) and device logic rules, key derived quantities are derived, laying a semantic foundation for multi-source data fusion.

[0108] In this embodiment, the statistical characteristics of the continuous quantities include mean, maximum value, slope, standard deviation, skewness, and range. Key derived quantities include power, energy, pressure difference, temperature difference, consistency index, and cooling response delay.

[0109] Understandably, step S22 involves cross-device semantic fusion and anomalous signal identification to uncover collaborative relationships between devices within the same time window. Specifically, this includes:

[0110] The "full-site status snapshot" constructed using the unified time reference described above is used to perform multi-dimensional fusion of key fields (including temperature, current, power, SOC, operating status, and alarm flags) through field association strategies. This allows for the mining of semantic associations between devices in terms of operating behavior, response relationships, and abnormal behavior. Cross-device coupling behaviors are analyzed (e.g., whether high temperature triggers the cooling system, whether rapid changes in SOC are accompanied by PCS start-up and shutdown, whether alarms occur synchronously, etc.). Based on rule-driven, frequency statistics, and graph structure modeling methods, these cross-device semantic behaviors are modeled and matched to identify abnormal linkage patterns (e.g., temperature rise + cooling failure + power drop) and extract key joint features.

[0111] Optionally, based on rule-driven, frequency statistics, and graph structure modeling methods, these cross-device semantic behaviors are modeled and matched to identify abnormal linkage patterns and extract key joint features, including:

[0112] Step 1: Rule engine matching.

[0113] First, key fields from the panoramic status (including cell voltage, temperature, SOC, PCS power, liquid cooling status, dehumidification status, and fire alarm) are input into the rule engine according to time windows. The rule engine has a built-in predefined trigger-response rule library. For example, when the cell temperature exceeds a threshold, the liquid cooling responds at full speed; or when the SOC drops rapidly, the PCS reduces its load. Whenever the data within the time window meets a rule condition in the trigger-response rule library, the corresponding trigger event (such as the cell temperature exceeding the threshold or the SOC dropping rapidly) is recorded, along with its fields, device ID, timestamp, and trigger type, providing basic data for subsequent cross-device correlation analysis.

[0114] Step 2: Frequency modeling.

[0115] Based on the recorded triggering events, frequency statistics and quantitative modeling are performed for each type of triggering event. Frequency statistics indicators include event trigger frequency, response delay, or amplitude variation range, used to determine the stability and repeatability of linked events. Through this quantitative analysis, high-probability, high-impact abnormal linkage patterns can be selected based on threshold conditions as candidate patterns for subsequent coding.

[0116] Step 3: Graph structure construction and reasoning.

[0117] Triggering events and their response relationships are constructed as a directed graph, where nodes represent device fields, edges represent triggering or response relationships, and the weights of the edges are the frequency statistics mentioned above. Through directed graph traversal and path matching algorithms, chain-like anomalies across devices and multiple fields are identified, such as "cell temperature rise → liquid cooling unresponsive → PCS power decrease". This structured representation can reveal complex systemic anomaly paths.

[0118] Step 4: High-dimensional joint feature vector encoding.

[0119] The rule engine trigger records, frequency statistics, and directed graph structure path features are integrated and encoded into a high-dimensional joint feature vector (i.e., key joint features) according to a preset field template. In the encoding process, continuous quantities are normalized, discrete quantities are encoded and mapped, and missing fields are filled with default values ​​and marked with missing bitmaps.

[0120] Understandably, step S23 compresses the multidimensional data into a fixed-structure, uniform-format model input vector: a fixed-length, structured feature vector array that the model can directly use.

[0121] In a specific implementation, step S23 may include:

[0122] By constructing a feature configuration template, feature values ​​are extracted according to the preset field order. Continuous values ​​are normalized, discrete fields are encoded and mapped, and missing fields are filled with default values ​​and marked with missing bitmaps to ensure that the vector length is fixed and consistent. Finally, a structured fixed-length feature array is output.

[0123] The output of step S23 can support serialization in formats such as JSON and Protobuf.

[0124] Optionally, the method for obtaining the above-mentioned trained multi-task diagnostic model includes:

[0125] Step S31: Construct the training set;

[0126] Step S32: Construct the network architecture of the multi-task diagnostic model;

[0127] Step S33: Train the network architecture using the training set to obtain a trained multi-task diagnostic model.

[0128] In S31, the methods for constructing the training set include:

[0129] Step S311: Collect historical operating data of each device in the energy storage power station throughout its entire life cycle;

[0130] Step S312: Using the same method as steps S1 and S2, process the collected historical operation data of each device in the energy storage power station throughout its entire life cycle to obtain several historical structured feature vectors of the energy storage power station.

[0131] Step S313: Label each historical structured feature vector. The labels include the energy storage power station health status label, fault trend label, and fault risk score label. The health status label includes "101", "102", and "103", which represent normal, slightly abnormal, and severely abnormal, respectively. The fault trend label includes "201", "202", and "203", which represent fault increase, fault decrease, and stable trend, respectively. The risk score label includes "301", "302", "303", "304", "305", and "306", which correspond to the score ranges [0,20), [20,40), [40,60), [60,80), [80,95), and [95,100], respectively, representing extremely low risk, low risk, medium risk, medium-high risk, high risk, and extremely high risk.

[0132] The labeled historical structured feature vectors are collected and sorted in ascending order according to the data collection time to form a training set.

[0133] In practice, a label set can be constructed based on expert experience or on-site records. This label set contains labels for the health status of the energy storage power station, fault trend labels, and fault risk score labels. During labeling, the various labels recorded in the label set are used to label the sample data.

[0134] For example, the network architecture of the multi-task diagnostic model constructed above includes:

[0135] The input layer is used to input structured feature vectors, providing input for subsequent feature extraction;

[0136] A shared hidden layer, connected to the input layer, is used to extract a global representation of the structured feature vector of the input.

[0137] The task branch layer, connected to the shared hidden layer, specifically includes a state classification branch, a trend prediction time series branch, and a risk scoring branch. The state classification branch is used to predict the probability distribution of the health state categories of the energy storage power station; the trend prediction time series branch is used to predict the state change of the energy storage power station; and the risk scoring branch is used to predict the continuous distribution of the risk score of the energy storage power station.

[0138] The output layer includes a first softmax function layer, a second softmax function layer, and a Linear layer. The first softmax function layer is connected to the output of the state classification branch and is used to obtain and output the health status classification of the energy storage power station based on the output of the state classification branch. The second softmax function layer is connected to the output of the trend prediction time series branch and is used to obtain and output the fault trend classification of the energy storage power station based on the output of the trend prediction time series branch. The Linear layer is connected to the output of the risk scoring branch and is used to map the output of the risk scoring branch to the corresponding risk scoring classification and output it.

[0139] Optionally, step S33 involves training the network architecture using the training set to obtain a trained multi-task diagnostic model, including:

[0140] S331. Initialize the network architecture parameters to obtain the initial model;

[0141] S332. Input the training set into the initial model for forward propagation iterative training. In each iteration, calculate the loss of each task branch, and then calculate the total loss by fusing the losses of each branch through dynamic weights.

[0142] S333. Backpropagate to update the model parameters until the total loss converges or the preset number of training iterations are reached, and a well-trained multi-task diagnostic model is obtained.

[0143] Specifically, the state classification branch uses the cross-entropy loss function, the trend prediction time series branch uses the MSE loss function, and the risk scoring branch uses the MSE loss function.

[0144] In this application, the trained multi-task diagnostic model is deployed on a local diagnostic node. During deployment, the model is first processed into a lightweight model file that can be deployed on an edge computing device, and then deployed on the local diagnostic node.

[0145] Optionally, when processing the model into a lightweight model file that can be deployed on edge computing devices, compression techniques such as channel pruning, parameter quantization, and knowledge distillation can be used to generate a lightweight model file suitable for deployment on edge computing devices, while outputting the model input / output definition structure and standardized configuration template.

[0146] The model input / output definition structure and standardized configuration template provide standardized guidance for model deployment and use.

[0147] Model input and output definition structure: including the format of the input data received by the model, and the meaning and format of the model output results.

[0148] Standardized configuration templates: Standardize hardware parameters (such as computing power allocation and storage usage settings) and runtime environment configurations (such as dependency library versions and inference engine parameters) when deploying models to ensure that the deployment process is standardized and reusable.

[0149] Optionally, the present invention supports an abnormal sample reflow mechanism during model training.

[0150] The abnormal sample reflux mechanism includes:

[0151] In each training iteration, the model confidence score is calculated and a preset threshold condition is used to determine whether the model confidence score is too low or to check for the existence of unknown state samples.

[0152] If the model confidence is too low or there are unknown state samples, the relevant features and their context are packaged into a mini sample set and sent back to the cloud for model retraining or parameter fine-tuning.

[0153] As another embodiment of the present invention, the method further includes:

[0154] Step S4: Trigger the corresponding local response strategy based on the predicted current health status, current failure trend, and current failure risk level.

[0155] In practical implementation, a prediction result-response rule mapping table can be preset according to the actual situation. When in use, after obtaining the predicted current health status, current fault trend, and current fault risk level, the corresponding local response strategy is triggered according to the mapping table.

[0156] The prediction result-response rule mapping table can be developed by experts in the field based on experience and actual conditions.

[0157] Local response strategies include cooling system enhancement strategies, PCS load reduction strategies, isolation discharge strategies, and alarm notification strategies.

[0158] As an exemplary embodiment of the present invention, the device semantic meta-model further includes a device semantic processing module. This device semantic processing module is used to perform standard field derivation calculations in the standard field dictionary, field addition, device registration, and semantic conflict handling.

[0159] The field derivation calculation is specifically designed for parameters that cannot be directly measured, such as efficiency and power consumption (e.g., the discharge capacity is calculated by multiplying the nominal power consumption by the state of charge (SOC)). When registering a new device model or adding a field through the device semantic meta-model, this invention automatically verifies whether the field name, unit, and derivation logic overlap or conflict with existing device models or newly added fields in the device semantic meta-model. If so, it provides prompts or merging suggestions to ensure the accuracy and consistency of the overall semantic system.

[0160] Optionally, when there are multiple local diagnostic nodes, the trained multi-task diagnostic model is deployed on each local diagnostic node.

[0161] Optionally, the present invention supports semantic drift detection for the multi-task diagnostic model deployed on each local diagnostic node.

[0162] The semantic drift detection includes:

[0163] Monitor the degree of deviation between the model input distribution and the training set sample data distribution;

[0164] Based on the degree of offset, it is determined whether to trigger a hot update of the model or parameter adjustment, so as to ensure the robustness of the model in long-term operation.

[0165] Optionally, step S3 specifically includes:

[0166] The structured feature vectors of each device in the energy storage power station are input into the trained multi-task diagnostic model deployed on each local diagnostic node.

[0167] The trained multi-task diagnostic model deployed on each local diagnostic node performs calculations on the input structured feature vector to obtain the predicted current health status, current fault trend, and current fault risk level of the energy storage power station, and sends them to the collaborative judgment module in a structured form.

[0168] The collaborative judgment module forms a candidate result set from the prediction results sent by each local diagnostic node. For the same type of abnormal event, the final prediction result is determined by majority voting or weighted voting mechanism.

[0169] In practice, all local diagnostic nodes have the same initial weight.

[0170] During use, adjustments can be made based on the accuracy of historical prediction results (i.e., historical diagnostic results) of each local diagnostic node. For example, the accuracy of each local diagnostic node's judgment can be periodically calculated (compared with actual fault verification results). Diagnostic nodes with an accuracy 10% higher than other local diagnostic nodes will have their weight increased by 0.1, while those with an accuracy 10% higher will have their weight decreased by 0.1. This adjustment continues until only one local diagnostic node has a non-zero weight, while the weights of the remaining local diagnostic nodes become zero.

[0171] In practice, when only one local diagnostic node has a non-zero weight, while the weights of all other local diagnostic nodes become zero, an alarm can be issued to remind relevant personnel to update the model deployed on the diagnostic node with the zero weight.

[0172] As some other embodiments of the present invention, the method further includes:

[0173] After the local response strategy is triggered and executed, the changes of preset target fields of each device in the energy storage power station are continuously monitored for a preset time period, the degree of improvement after the strategy response is calculated, and a feedback score is generated based on this.

[0174] The control strength of the local response strategy is adjusted based on feedback scores, forming a self-closing response optimization mechanism.

[0175] The preset target fields include temperature, current, and SOC.

[0176] Feedback scores include response performance score, latency time, and residual anomalies.

[0177] The energy storage local monitoring system for an energy storage power station provided by this invention specifically includes:

[0178] Pre-built semantic meta-models of various equipment in energy storage power stations;

[0179] A pre-built unified protocol parsing table and semantic mapping table for various devices in an energy storage power station;

[0180] Monitoring data acquisition module;

[0181] Feature extraction module;

[0182] Edge-side intelligent diagnostic module.

[0183] The device semantic meta-model includes a standard field dictionary, which defines the core parameters of various devices in an energy storage power station.

[0184] The protocol parsing table is used to structure and parse the raw fields of the data streams generated by various devices in the energy storage power station.

[0185] The semantic mapping table is used to map the parsed raw fields to unified standard fields based on the standard field dictionary of the device semantic metamodel.

[0186] The monitoring data acquisition module is used for the acquisition and unified formatting of multi-source heterogeneous monitoring data in energy storage power stations.

[0187] Specifically, the monitoring data acquisition module is configured to perform the following steps:

[0188] Real-time data streams generated by each device in the energy storage power station are collected through the device communication interface;

[0189] Real-time, based on the protocol parsing table, the raw fields in each collected data stream are parsed in a structured manner;

[0190] In real time, based on the semantic mapping table, the parsed original fields are mapped to unified standard fields according to the standard field dictionary of the device semantic meta-model, so as to obtain the standard fields of each device in the energy storage power station.

[0191] By employing a master clock anchor point and sliding time window strategy, standard fields of each device in the energy storage power station are spatiotemporally aligned and resampled to obtain and output the standard structured data of each device in the energy storage power station.

[0192] The feature extraction module is used to perform data fusion and feature extraction on the standard structured data of each device output by the monitoring data acquisition module to obtain the structured feature vectors of each device in the energy storage power station.

[0193] The edge-side intelligent diagnostic module is used to input the structured feature vectors of each device in the energy storage power station into a pre-trained multi-task diagnostic model deployed on the local diagnostic node. After model calculation, the current health status, current fault trend, and current fault risk level of the energy storage power station are predicted.

[0194] This embodiment is an example of an energy storage power station local monitoring system. This system and the energy storage power station local monitoring methods in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the energy storage power station local monitoring system, please refer to the embodiments of the energy storage power station local monitoring methods in the preceding text.

[0195] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

[0196] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0197] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for on-site monitoring of energy storage in an energy storage power station, characterized in that, The method is based on a pre-constructed device semantic meta-model of various equipment in an energy storage power station, as well as a unified protocol parsing table and semantic mapping table for various equipment in the energy storage power station. The device semantic meta-model includes a standard field dictionary. The standard field dictionary supports a scope constraint mechanism and a dynamic semantic annotation mechanism to specify the validity of a field under a specific state or condition, and can dynamically adjust the semantic interpretation of the field according to the device operating status, scenario label, or environmental variables, so that the same field has differentiated semantic expression in different contexts. The device semantic meta-model also includes a device semantic processing module. The device semantic processing module is used to implement the derivation and calculation of standard fields in the standard field dictionary, field addition, device registration, and semantic conflict handling. The methods include: S1. Acquisition and unified formatting of multi-source heterogeneous monitoring data: Real-time data streams generated by each device in the energy storage power station are collected through the device communication interface; Real-time, based on the protocol parsing table, the raw fields in each collected data stream are parsed in a structured manner; In real time, based on the semantic mapping table, the parsed original fields are mapped to unified standard fields according to the standard field dictionary of the device semantic meta-model, so as to obtain the standard fields of each device in the energy storage power station. By employing a master clock anchor point and sliding time window strategy, the standard fields of each device in the energy storage power station are spatiotemporally aligned and resampled to obtain and output the standard structured data of each device in the energy storage power station. S2. Perform data fusion and feature extraction on the standard structured data of each device output in S1 to obtain the structured feature vector of each device in the energy storage power station. S3. Edge-side intelligent diagnosis: The structured feature vectors of each device in the energy storage power station are input into the pre-trained multi-task diagnosis model deployed on the local diagnosis node. After model calculation, the current health status, current fault trend and current fault risk level of the energy storage power station are predicted. S2 includes: For the standard structured data of each device in the energy storage power station within the sliding time window, based on the collaborative relationship between the devices in the energy storage power station, the standard structured data is multidimensionally fused through the field association strategy. Then, based on the multidimensional fusion, the cross-device coupling behavior between the devices in the energy storage power station is analyzed, and key joint features are extracted from it.

2. The method for on-site monitoring of energy storage in a power station according to claim 1, characterized in that, S2 also includes: For each device in the energy storage power station, a unified sliding time window is adopted to statistically analyze the continuous characteristics and discrete state characteristics of the standard structured data of each device output in S1. Then, the key derived quantities are derived by combining physical formulas and device logic rules, and the derived key derived quantities are collected as multi-dimensional aggregated features. By constructing a feature configuration template, the multidimensional aggregated features and key joint features obtained under the sliding time window are uniformly encoded and fixed-length structured to obtain the structured feature vectors of each device in the energy storage power station.

3. The method for on-site monitoring of energy storage in an energy storage power station according to claim 2, characterized in that, The method also includes: S4. Based on the predicted current health status, current failure trend, and current failure risk level, trigger the corresponding local response strategy.

4. The method for on-site monitoring of energy storage in a power station according to claim 2, characterized in that, Methods for obtaining pre-trained multi-task diagnostic models in S3 include: S31. Construct the training set; S32. Construct the network architecture for a multi-task diagnostic model; S33. Train the network architecture using the training set to obtain a trained multi-task diagnostic model; In S31, the methods for constructing the training set include: S311. Collect historical operating data of all equipment in the energy storage power station throughout its entire life cycle; S312. Using the same method as S1 and S2, process the collected historical operation data of each device in the energy storage power station throughout its entire life cycle to obtain several historical structured feature vectors of the energy storage power station. S313. Label each historical structured feature vector. The label includes the energy storage power station health status label, fault trend label, and fault risk score label. The health status label includes "101", "102", and "103", which represent normal, slightly abnormal, and severely abnormal, respectively. The fault trend label includes "201", "202", and "203", which represent fault increase, fault decrease, and stable trend, respectively. The risk score label includes "301", "302", "303", "304", "305", and "306", which correspond to the score ranges [0,20), [20,40), [40,60), [60,80), [80,95), and [95,100], respectively, representing extremely low risk, low risk, medium risk, medium-high risk, high risk, and extremely high risk. The labeled historical structured feature vectors are collected and sorted in ascending order according to the data collection time to form a training set.

5. The method for on-site monitoring of energy storage in an energy storage power station according to claim 4, characterized in that, The network architecture of the multi-task diagnostic model includes: The input layer is used to input structured feature vectors, providing input for subsequent feature extraction; A shared hidden layer, connected to the input layer, is used to extract a global representation of the structured feature vector of the input. The task branch layer, connected to the shared hidden layer, specifically includes a state classification branch, a trend prediction time series branch, and a risk scoring branch. The state classification branch is used to predict the probability distribution of the health state categories of the energy storage power station; the trend prediction time series branch is used to predict the state change of the energy storage power station; and the risk scoring branch is used to predict the continuous distribution of the risk score of the energy storage power station. The output layer includes a first softmax function layer, a second softmax function layer, and a Linear layer. The first softmax function layer is connected to the output of the state classification branch and is used to obtain and output the health state classification corresponding to the energy storage power station based on the output of the state classification branch. The second softmax function layer is connected to the output of the trend prediction time series branch and is used to obtain and output the fault trend classification corresponding to the energy storage power station based on the output of the trend prediction time series branch. The Linear layer is connected to the output of the risk scoring branch and is used to map the output of the risk scoring branch to the corresponding risk scoring classification and output it. S33. Train the network architecture using the training set to obtain a trained multi-task diagnostic model.

6. The method for on-site monitoring of energy storage in an energy storage power station according to claim 5, characterized in that, S33 includes: S331. Initialize the network architecture parameters to obtain the initial model; S332. Input the training set into the initial model for forward propagation iterative training. In each iteration, calculate the loss of each task branch, and then calculate the total loss by fusing the losses of each branch through dynamic weights. Among them, the state classification branch adopts the cross-entropy loss function, the trend prediction time series branch adopts the MSE loss function, and the risk scoring branch adopts the MSE loss function. S333. Backpropagate to update the model parameters until the total loss converges or the preset number of training iterations are reached, and a well-trained multi-task diagnostic model is obtained.

7. The method for on-site monitoring of energy storage in an energy storage power station according to claim 1, characterized in that, When there are multiple local diagnostic nodes as described in S3, the trained multi-task diagnostic model is deployed on each local diagnostic node. S3 includes: The structured feature vectors of each device in the energy storage power station are input into the trained multi-task diagnostic model deployed on each local diagnostic node. The trained multi-task diagnostic model deployed on each local diagnostic node performs calculations on the input structured feature vector to obtain the predicted current health status, current fault trend, and current fault risk level of the energy storage power station, and sends them to the collaborative judgment module in a structured form. The collaborative judgment module forms a candidate result set from the prediction results sent by each local diagnostic node. For the same type of abnormal event, the final prediction result is determined by majority voting or weighted voting mechanism.

8. The method for on-site monitoring of energy storage in a power station according to claim 3, characterized in that, The method also includes: After the local response strategy is triggered and executed, the changes of preset target fields of each device in the energy storage power station are continuously monitored for a preset time period, the degree of improvement after the strategy response is calculated, and a feedback score is generated based on this. The control strength of the local response strategy is adjusted based on feedback scores, forming a self-closing response optimization mechanism.

9. A local monitoring system for energy storage in an energy storage power station, characterized in that, The system includes a pre-built semantic meta-model of various devices in an energy storage power station, as well as a unified protocol parsing table and semantic mapping table for various devices in the energy storage power station. The device semantic meta-model includes a standard field dictionary. The standard field dictionary supports a scope constraint mechanism and a dynamic semantic annotation mechanism to specify the validity of fields under specific states or conditions, and can dynamically adjust the semantic interpretation of fields according to the device operating status, scenario labels, or environmental variables, so that the same field has differentiated semantic expressions in different contexts. The device semantic meta-model also includes a device semantic processing module. The device semantic processing module is used to implement standard field derivation calculation, field addition, device registration, and semantic conflict handling in the standard field dictionary. The system also includes: The monitoring data acquisition module is used for the acquisition and unified formatting of multi-source heterogeneous monitoring data in energy storage power stations. The monitoring data acquisition module is configured to perform the following steps: Real-time data streams generated by each device in the energy storage power station are collected through the device communication interface; Real-time, based on the protocol parsing table, the raw fields in each collected data stream are parsed in a structured manner; In real time, based on the semantic mapping table, the parsed original fields are mapped to unified standard fields according to the standard field dictionary of the device semantic meta-model, so as to obtain the standard fields of each device in the energy storage power station. By employing a master clock anchor point and sliding time window strategy, the standard fields of each device in the energy storage power station are spatiotemporally aligned and resampled to obtain and output the standard structured data of each device in the energy storage power station. The system also includes: The feature extraction module is used to perform data fusion and feature extraction on the standard structured data of each device output by the monitoring data acquisition module to obtain the structured feature vectors of each device in the energy storage power station. The edge-side intelligent diagnostic module is used to input the structured feature vectors of each device in the energy storage power station into the pre-trained multi-task diagnostic model deployed on the local diagnostic node. After the model is calculated, the predicted current health status, current fault trend and current fault risk level of the energy storage power station are obtained. The feature extraction module is configured to perform the following steps: For the standard structured data of each device in the energy storage power station within the sliding time window, based on the collaborative relationship between the devices in the energy storage power station, the standard structured data is multidimensionally fused through the field association strategy. Then, based on the multidimensional fusion, the cross-device coupling behavior between the devices in the energy storage power station is analyzed, and key joint features are extracted from it.