Abnormity detection and interpretation method, device and equipment for energy storage system and medium

Abstract

The invention discloses an energy storage system-oriented anomaly detection and interpretation method, apparatus and device, and a medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: acquiring multi-source time sequence data of the energy storage system, and performing multi-channel time alignment processing on the multi-source time sequence data to obtain a multi-channel observation vector sequence; determining an anomaly score through an anomaly detection model based on the multi-channel observation vector sequence; determining an abnormal event under the condition that the abnormal score meets a preset alarm condition, and constructing a structured abnormal evidence set corresponding to the abnormal event; and inputting the structured exception evidence set into a language model to perform causal analysis on the exception event, and generating exception reason interpretation and execution suggestions corresponding to exception reasons. According to the invention, the overall operation and maintenance efficiency and the processing consistency can be improved.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and in particular relates to a method, device, equipment and medium for anomaly detection and interpretation in energy storage systems. Background Technology

[0002] With the continuous development of new energy applications, energy storage systems have been widely deployed in scenarios such as photovoltaic-energy storage microgrids, charging stations, and user-side energy storage. These systems typically require the coordination of multiple operations, including photovoltaic power generation, energy storage charging and discharging, load consumption, and grid interaction. Their operational status is affected by various factors such as time, electricity price, weather conditions, and dispatch strategies, resulting in a dynamic operating mode. The resulting operational data is complex in both time and indicator dimensions, involving multiple types of time-series data such as voltage, current, power, state of charge, electricity price, load, and environmental parameters. Furthermore, with the application of automated control and intelligent algorithms in energy storage systems, the system's operating status is no longer limited to static conditions. Abnormal manifestations have gradually evolved from single indicator exceeding limits to multi-indicator coupled changes, trend deviations, or anomalies related to operational strategies. Under complex operating conditions and long-term operation, the difficulty of identifying and analyzing abnormal states continues to increase.

[0003] In related technologies, the methods for anomaly monitoring and analysis of energy storage systems mostly rely on static threshold rules, manual experience judgment, or single detection models. These methods have limited adaptability to changes in operating conditions and strategy adjustments. Anomaly information is usually presented in the form of simple alarms, which are difficult to reflect the time characteristics of the anomaly and the process of related data changes. Further investigation by manual means is still required, and the overall operation and maintenance efficiency and processing consistency need to be improved. Summary of the Invention

[0004] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes a method, apparatus, equipment, and medium for anomaly detection and interpretation in energy storage systems, in order to improve overall operation and maintenance efficiency and processing consistency.

[0005] Firstly, this application provides a method for anomaly detection and interpretation in energy storage systems, the method comprising: Acquire multi-source time-series data of the energy storage system, and perform multi-channel time alignment processing on the multi-source time-series data to obtain a multi-channel observation vector sequence; Based on the multi-channel observation vector sequence, an anomaly score is determined using an anomaly detection model; If the anomaly score meets the preset alarm conditions, an anomaly event is identified, and a structured set of anomaly evidence corresponding to the anomaly event is constructed. The structured set of anomaly evidence is input into a language model to perform causal analysis on the anomaly events, generate explanations of the anomaly causes, and corresponding execution suggestions.

[0006] Secondly, this application provides an anomaly detection and interpretation device for energy storage systems, the device comprising: The data acquisition module is used to acquire multi-source time-series data of the energy storage system and perform multi-channel time alignment processing on the multi-source time-series data to obtain a multi-channel observation vector sequence; Anomaly detection module is used to determine anomaly score based on the multi-channel observation vector sequence using an anomaly detection model; An anomaly interpretation module is used to determine an anomaly event when the anomaly score meets preset alarm conditions, and to construct a structured set of anomaly evidence corresponding to the anomaly event. The anomaly explanation module is further configured to input the structured anomaly evidence set into a language model to perform causal analysis on the anomaly event, generate anomaly cause explanations and corresponding execution suggestions.

[0007] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the anomaly detection and interpretation method for energy storage systems as described in the first aspect above.

[0008] Fourthly, this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the anomaly detection and interpretation method for energy storage systems as described in the first aspect above.

[0009] Fifthly, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the anomaly detection and interpretation method for energy storage systems as described in the first aspect.

[0010] In a sixth aspect, this application provides a computer program product, including a computer program that, when executed by a processor, implements the anomaly detection and interpretation method for energy storage systems as described in the first aspect above.

[0011] The anomaly detection and interpretation method, device, electronic equipment, non-transitory computer-readable storage medium, chip, and computer program product provided in this application for energy storage systems perform multi-channel time alignment on multi-source time-series data of the energy storage system and construct a multi-channel observation vector sequence. Based on an anomaly detection model, the operating status of the energy storage system is quantitatively evaluated to determine anomaly scores. When the anomaly scores meet the alarm conditions, a structured set of anomaly evidence corresponding to the anomaly event is constructed. The structured set of anomaly evidence is then input into a language model for causal analysis. Thus, while achieving accurate detection of anomaly events in the energy storage system, it can automatically generate anomaly cause explanations and actionable handling suggestions corresponding to the formation mechanism of the anomaly event, upgrading anomaly handling from merely providing alarm information to evidence-based explanation and decision support.

[0012] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0013] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a schematic diagram illustrating an application scenario of the anomaly detection and interpretation method for energy storage systems provided in the embodiments of this application; Figure 2 This is one of the flowcharts illustrating the anomaly detection and interpretation method for energy storage systems provided in this application embodiment; Figure 3 This is the second flowchart of the anomaly detection and interpretation method for energy storage systems provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of the anomaly detection and interpretation device for energy storage systems provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0014] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0015] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0016] With the development of new energy technologies, energy storage systems have been widely applied in scenarios such as photovoltaic-energy storage microgrids, charging stations, user-side energy storage, and integrated energy management. These energy storage systems typically require coordination between multiple business units, including photovoltaic power generation, energy storage charging and discharging, load consumption, and grid power purchase and sale. Their operational status is affected by various factors such as time, weather conditions, electricity pricing mechanisms, and dispatch strategies, exhibiting significant dynamic changes. The resulting operational data is complex in both time and index dimensions, involving multiple types of time-series data such as voltage, current, power, state of charge, electricity price, load, and ambient temperature.

[0017] In practical applications, energy storage systems no longer operate solely as standalone hardware devices. Instead, they are integrated with automated control and intelligent algorithms to achieve functions such as peak-valley scheduling, energy optimization, and coordinated control of operational status. As control strategies and operating modes continue to evolve, the operating status and abnormal manifestations of energy storage systems also change. This means that abnormal states are no longer limited to exceeding the limits of a single indicator, but may manifest as changes in a combination of multiple indicators, long-term trend deviations, or deviations related to the operating strategy.

[0018] The following are the main methods used in related technologies for monitoring and analyzing anomalies in the operating status of energy storage systems: One type monitors operational indicators based on pre-set static thresholds or empirical rules. An alarm is triggered when the monitored indicator exceeds the preset range. This method is simple to implement, but it is highly sensitive to changes in operating conditions and struggles to adapt to changes in state caused by different seasons, load levels, and strategy adjustments. Furthermore, the threshold maintenance cost is high.

[0019] Another approach relies on the manual experience of operations and maintenance personnel to analyze and judge operational data. While this method can incorporate on-site experience to some extent, it is highly dependent on personnel, making it difficult to maintain consistency in large-scale deployment scenarios, and its response speed to new anomalies or policy changes is limited.

[0020] Another type involves some solutions that attempt to introduce data-driven models to detect abnormal states. However, these solutions focus primarily on the anomaly determination itself and fail to adequately characterize the anomaly formation process, the correlation between relevant operational data, and the connection between anomalies and operational strategies. The experience accumulated from historical operational data is difficult to reuse effectively, leading to the recurrence of similar problems at different times or in different systems.

[0021] In view of this, this application provides an anomaly detection and interpretation method for energy storage systems, aiming to solve the problems of insufficient accuracy in anomaly detection, difficulty in automatically explaining the causes of anomalies, and reliance on human experience in anomaly handling in related technologies under complex operating scenarios. By uniformly aligning and characterizing multi-source time-series operating data of batteries and related equipment, a multi-model collaborative time-series anomaly detection framework is constructed. Based on joint learning of reconstruction and prediction, accurate scoring of instantaneous and trend anomalies is achieved. Furthermore, the anomaly detection results, key data fragments, and historical similar cases are structured into anomaly evidence. Language models and knowledge graphs in the energy storage field are introduced for retrieval and causal reasoning, generating natural language explanations and strategy suggestions corresponding to the anomaly events. Therefore, this application can achieve automatic detection and interpretable analysis of multiple types of anomalies, upgrading anomaly alarms from simple numerical prompts to well-supported, verifiable, and executable handling solutions, thereby improving the safety, reliability, intelligence level, and adaptability to new scenarios in the operation and maintenance of energy storage systems.

[0022] The following description, in conjunction with the accompanying drawings, details the anomaly detection and interpretation method for energy storage systems provided in this application through specific embodiments and application scenarios.

[0023] The anomaly detection and interpretation method for energy storage systems provided in this application embodiment can be applied to integrated energy systems containing energy storage units, such as residential energy storage systems, industrial and commercial energy storage systems, photovoltaic-energy storage microgrid systems, charging station energy storage systems, grid-side or independent energy storage power stations, and is used to detect and interpret anomalies in the operating status of energy storage systems under different operating modes and scheduling conditions.

[0024] Figure 1 This is a schematic diagram illustrating an application scenario of the anomaly detection and interpretation method for energy storage systems provided in some embodiments of this application. For example... Figure 1 As shown, the application environment may include, for example, an energy storage system 102, a network, and electronic devices 104.

[0025] The energy storage system 102 includes, but is not limited to, energy storage batteries, battery management-related equipment, and energy conversion and control equipment connected to new energy power generation units and electrical loads. For example... Figure 1 The residential energy storage scenario shown can be connected to distributed photovoltaic power generation units and user-side loads to achieve energy storage and release.

[0026] Multi-source time-series data generated during the operation of the energy storage system can be transmitted to electronic device 104 via network. The multi-source time-series data includes, but is not limited to, the operating status data of the energy storage battery, power and energy data, and control and event data related to the operation of the energy storage system.

[0027] Electronic device 104 can be a server or an edge computing node, used to analyze and process received multi-source time-series data, and output detection and interpretation results related to abnormal operation of the energy storage system. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, and big data and artificial intelligence platforms.

[0028] It should be noted that, Figure 1 The application scenarios shown are merely examples, and the specific structural form, deployment scale, and deployment location of the computing devices of the energy storage system do not constitute a limitation on this application.

[0029] The anomaly detection and interpretation method for energy storage systems provided in this application embodiment can be executed by an electronic device or a functional module or entity in an electronic device that can implement the method.

[0030] The following uses an electronic device as the execution subject to illustrate the anomaly detection and interpretation method for energy storage systems provided in the embodiments of this application.

[0031] Figure 2 This is a flowchart illustrating an anomaly detection and interpretation method for energy storage systems provided in some embodiments of this application. For example... Figure 2 As shown, the method includes steps 210 to 240.

[0032] Step 210: Obtain multi-source time-series data of the energy storage system, and perform multi-channel time alignment processing on the multi-source time-series data to obtain a multi-channel observation vector sequence.

[0033] Multi-source time-series data refers to the data sequence that is continuously collected by different functional modules, devices or subsystems according to their respective sampling periods and changes over time during the operation of the energy storage system.

[0034] In some embodiments, multi-source time-series data includes, but is not limited to, one or more of the following: battery operation data (e.g., battery voltage, current, state of charge, and temperature data) collected by the battery management system; power and energy scheduling data (e.g., charging and discharging power and grid-connected power) collected by the energy management system; environmental data (e.g., ambient temperature and equipment operating status) collected by the environmental and auxiliary systems; and control command data and event log data collected by the control and communication systems.

[0035] In some embodiments, the multi-source time-series data also includes input data related to the energy storage system's operating strategy. Strategy-related inputs refer to data that constrains or guides the scheduling and control of the energy storage system. Strategy-related inputs include at least one or more of the following: electricity price sequences, photovoltaic and load forecast data, grid connection constraints, power constraints, minimum state of charge (SOC) constraints, and actual control command data issued during each operating cycle. The electricity price sequence refers to a time-varying price data sequence generated according to time-of-use pricing or real-time pricing mechanisms. Photovoltaic and load forecast data refers to data obtained by predicting photovoltaic output and load demand over a future period. Control command data refers to charging / discharging or power regulation commands actually executed by the control system during each operating cycle.

[0036] For example, multi-source time-series data can be transmitted from the energy storage system to electronic devices via fieldbus, industrial Ethernet, or wireless communication networks.

[0037] Since data from different sources may differ in sampling frequency, sampling start time, and timestamp accuracy, this application embodiment performs multi-channel time alignment processing on the acquired multi-source time series data. In this way, multi-source time series data from different data sources can be mapped onto a unified time axis, forming a multi-channel observation vector sequence composed of multiple channel data at each time point.

[0038] The multi-channel observation vector sequence consists of multiple multi-channel observation vectors arranged sequentially along the time dimension. Each multi-channel observation vector is used to characterize the observed value of each channel data of the energy storage system at the corresponding time.

[0039] Multi-channel time alignment processing refers to matching data from different channels according to their timestamps, using a preset unified time base as a reference, so that each channel corresponds to the state value of the same operating moment at the same time point. A channel refers to a time-series data unit used to carry the same physical quantity or the same logical variable, such as a battery voltage channel, a charge / discharge power channel, or an ambient temperature channel.

[0040] Step 220: Based on the multi-channel observation vector sequence, determine the anomaly score through the anomaly detection model.

[0041] Electronic devices take a multi-channel observation vector sequence as input and feed it into an anomaly detection model for processing. In some embodiments, the anomaly detection model performs feature extraction and state evaluation on the input multi-channel observation vector sequence to generate a numerical result, which is used to characterize the degree to which the current operating state deviates from the normal operating mode, as an anomaly score.

[0042] The anomaly score can be a continuous or discrete value, used to reflect the degree of anomaly in the operating status of the energy storage system within a corresponding time period.

[0043] In some embodiments, the anomaly detection model can be constructed using various implementation methods to adapt to the operating scale, data complexity, and deployment conditions of different energy storage systems.

[0044] For example, anomaly detection models can be built on feedforward neural networks or convolutional neural networks, or they can be anomaly detection models based on traditional machine learning algorithms, such as support vector machines, isolated forests, probabilistic statistical models, or clustering models, to analyze and process multi-channel observation vector sequences and output anomaly scores to characterize the degree of abnormality in the operating state.

[0045] In practice, anomaly scores can be generated for each time point in the multi-channel observation vector sequence, or for data within a preset time range, to represent the overall operating status of the energy storage system at a given time point or within a given time range.

[0046] Step 230: If the anomaly score meets the preset alarm conditions, identify the anomaly event and construct a structured set of anomaly evidence corresponding to the anomaly event.

[0047] Among them, the preset alarm conditions refer to the judgment rules used to determine whether an abnormal event is triggered, which can be set based on the relationship between the abnormal score and the preset threshold.

[0048] After receiving the anomaly score, the electronic device determines whether the score meets preset alarm conditions. When the anomaly score exceeds the corresponding anomaly score threshold for at least one consecutive moment or time period, the electronic device can determine that an anomaly event has occurred in the energy storage system within the corresponding time period. An anomaly event refers to an operating state of the energy storage system that deviates significantly from its normal operating state during operation.

[0049] For example, if the abnormal score is If M consecutive time windows (usually M=2~3) are passed, the electronic equipment determines that the energy storage system has an anomaly and meets the preset alarm conditions. Among these, the anomaly scoring threshold... It can be determined from historical normal data, such as taking its 80th, 90th, or 99th percentile.

[0050] After identifying an abnormal event, the electronic device constructs a structured set of abnormal evidence corresponding to the abnormal event. This structured set of abnormal evidence refers to a data set formed by organizing various types of operational data related to the abnormal event according to a preset data field structure. It is used to uniformly represent the occurrence time, associated data, and changes of the abnormal event.

[0051] In some embodiments, the structured anomaly evidence set includes at least one or more of the following: anomaly time period, anomaly type corresponding to anomaly event, anomaly score, and multi-channel runtime data fragments related to the anomaly event.

[0052] In some embodiments, the structured anomaly evidence set also includes the AI ​​strategy executed at the current moment or during the anomaly period, and the corresponding strategy intent prompt. The AI ​​strategy refers to the operational decision automatically generated by the artificial intelligence decision-making model in the energy storage system based on multi-source operational data, predictive information, and constraints. This strategy guides the energy storage system to perform charge / discharge control, power allocation, operating mode switching, or other scheduling control behaviors at the corresponding moment. The strategy intent prompt provides a brief explanation of the decision background and objectives of the AI ​​strategy.

[0053] By introducing policy intent cues, AI policies can be associated with their corresponding operational purposes. This enables the subsequent anomaly interpretation process to distinguish whether the anomaly is the result of the policy actively executing according to the predetermined goal, or due to the policy deviating from the expected goal or execution anomaly. This provides contextual semantic support for anomaly cause analysis and avoids misjudging strategic behavior as system anomalies.

[0054] In a specific example, the energy storage system is an industrial and commercial energy storage system, and its multi-channel observation vector sequence includes battery charging and discharging power channels, battery state of charge channels, battery temperature channels, and grid-connected power channels.

[0055] If, during a given runtime period, the anomaly detection model outputs anomaly scores higher than a preset anomaly score threshold at multiple consecutive time points (e.g., anomaly scores of 0.82, 0.87, and 0.85 over three consecutive sampling periods, while the threshold is 0.8), the electronic device determines that a preset alarm condition is met and identifies the time period comprised of these consecutive time points as the anomaly time period corresponding to the anomaly event. Then, based on the anomaly time period, the electronic device extracts channel data segments within the corresponding time range from the multi-channel observation vector sequence. For example, it extracts battery temperature change data, battery state of charge change data, and charge / discharge power change data within the anomaly time period. Simultaneously, it analyzes the changing trend of the anomaly scores within the anomaly time period. Subsequently, the electronic device aligns the aforementioned anomaly score sequence with each channel data segment along a unified time axis and organizes them according to a preset data field structure, such as constructing data records in the form of "anomaly time period - anomaly score sequence - channel data segment," thereby forming a structured set of anomaly evidence corresponding one-to-one with the anomaly event.

[0056] Through the above processing, when an abnormal event is identified, the operational data scattered across different time points and channels can be centrally organized in a structured form to obtain a structured set of abnormal evidence corresponding to the abnormal event, providing a unified data foundation for subsequent processing.

[0057] Step 240: Input the structured set of abnormal evidence into the language model to perform causal analysis on the abnormal events, generate explanations of the causes of the abnormalities and corresponding execution suggestions.

[0058] A language model is a model capable of natural language understanding and generation. It can perform semantic parsing and reasoning on structured input information and output natural language text results related to the input information. For example, language models include, but are not limited to, Large Language Models (LLMs). Large language models are built on deep neural networks and possess natural language understanding and generation capabilities. They can perform semantic parsing and reasoning on a set of structured anomaly evidence to generate corresponding explanations of the anomalies and implementation suggestions. For example, language models can be deployed on local computing devices, edge computing nodes, or cloud computing platforms within energy storage systems.

[0059] After constructing the structured set of anomaly evidence, the electronic device inputs this set into a language model. The language model first parses the structured set of anomaly evidence to identify the timing, severity, and key data change characteristics related to the anomaly. Based on this, the language model infers and analyzes the relationships between the data involved in the anomaly to determine at least one anomaly-related cause.

[0060] After identifying the cause of the anomaly, the language model generates a corresponding explanation of the cause, which describes the relevant factors and their relationships in natural language. Furthermore, the language model also generates action suggestions corresponding to the cause of the anomaly, indicating the appropriate handling or remedial measures to be taken in response to the anomaly.

[0061] In a specific example, after parsing the structured anomaly evidence set, the language model concluded that during the abnormal time period, the battery temperature channel showed a continuous upward trend, the state of charge decreased at an abnormally rapid rate, while the charging and discharging power remained at a high level. After parsing the structured anomaly evidence set, the language model identified the correlation between the aforementioned channel data. Based on this correlation, the language model performed causal analysis on the anomaly, determining that the anomaly might be related to insufficient heat dissipation of the battery under high-power charging and discharging conditions, and took this analysis result as the cause of the anomaly. Subsequently, the language model generated an explanation of the anomaly cause based on the cause, such as describing the changes in relevant operating parameters during the formation of the anomaly in natural language. Furthermore, the language model generated corresponding implementation suggestions based on the anomaly cause, such as suggesting reducing the charging and discharging power, checking the operating status of the heat dissipation device, or adjusting the operating strategy in subsequent operations.

[0062] According to the anomaly detection and interpretation for energy storage systems provided in this application, multi-channel time alignment is performed on multi-source time-series data of the energy storage system to construct a multi-channel observation vector sequence. Based on the anomaly detection model, the operating status of the energy storage system is quantitatively evaluated to determine the anomaly score. When the anomaly score meets the alarm conditions, a structured anomaly evidence set corresponding to the anomaly event is constructed. The structured anomaly evidence set is then input into a language model for causal analysis. Thus, while accurately detecting anomaly events in the energy storage system, it can automatically generate anomaly cause explanations and actionable handling suggestions corresponding to the formation mechanism of the anomaly event. This upgrades anomaly handling from merely providing alarm information to evidence-based interpretation and decision support.

[0063] Before performing anomaly detection on the operating data of the energy storage system, it is necessary to perform unified time alignment and preprocessing on the multi-source time series data, and organize the processed data into a multi-channel observation vector sequence with consistent structure and continuous time, so as to provide a stable and usable input data foundation for the subsequent anomaly detection model.

[0064] Therefore, in some embodiments, step 210 includes: Step 211: Collect multi-source time-series data from multiple data sources and obtain strategy data corresponding to the energy storage system operation strategy to form an original multi-source time-series data set; Step 212: Align each data point in the original multi-source time series data set to a unified time reference to obtain time-aligned multi-source time series data; Step 213: Based on time-aligned multi-source time series data, perform missing data processing, outlier sampling point suppression, and scale unification processing respectively to obtain multi-channel time series data with consistent dimensions. Step 214: Based on multi-channel time-series data with consistent dimensions, slide window slicing is performed according to a preset window length to construct a multi-channel observation vector sequence for input to the anomaly detection model.

[0065] In this embodiment, the electronic device first collects multi-source time-series data of the energy storage system from multiple data sources. Exemplarily, the data sources include, but are not limited to, one or more of the following: battery management system, energy management system, environmental and auxiliary system, communication and event log, and control and communication system.

[0066] Simultaneously, electronic devices acquire strategy data corresponding to the energy storage system's operation strategy, such as one or more of the following: charging and discharging strategy, power scheduling strategy, electricity price-based charging strategy, photovoltaic and load forecasting strategy, or other operational constraint information, to form an original multi-source time-series data set.

[0067] Subsequently, the electronic device aligns each data point in the original multi-source time series data set to a unified time base according to the timestamp, resulting in time-aligned multi-source time series data.

[0068] After time alignment is completed, the electronic equipment processes any missing data: for short-term missing data, linear interpolation is used to fill in the missing data; for long-term missing data, the corresponding time period is marked and the marked data is skipped in subsequent processes. Based on the time-aligned multi-source time-series data, missing data processing, outlier suppression, and scale unification are performed on each channel's data in sequence.

[0069] Subsequently, the electronic device performs anomaly suppression and scaling on the time-aligned multi-channel data. This includes smoothing outliers in each channel using a moving median filter, labeling significantly outlier data points based on statistical distribution, and reducing the weight of outlier data points in subsequent training phases. Furthermore, the electronic device can perform standardization on each channel's data, including zero-mean unit variance standardization or preset interval normalization, to eliminate dimensional differences between different physical quantities.

[0070] After completing the above processing, the electronic device performs sliding window slicing on the multi-channel time-series data with consistent dimensions using a preset window length, thereby constructing a continuous multi-channel observation vector sequence. Among them, the multi-channel observation vector sequence For multi-channel observation vectors within multiple sliding windows The data set that makes up the data It refers to t Multi-channel observation vector at any given time.

[0071] The above technical solution enables the unified mapping of multi-source time-series data from different data sources and time scales to a consistent time base and data structure, forming a multi-channel observation vector sequence with consistent dimensions and continuous time. This reduces the impact of data heterogeneity and time inconsistency on anomaly detection results. Furthermore, by constructing model input data using a sliding window approach, the anomaly detection model can analyze the system's operating status based on continuous time periods, which helps improve the stability and reliability of the subsequent anomaly detection process and provides a solid data foundation for the accurate identification of anomalies in energy storage systems.

[0072] During the operation of energy storage systems, anomalies may manifest as a deviation between the instantaneous operating state and the normal mode, or as a sustained deviation of key operating indicators from their historical statistical characteristics over a period of time. Relying solely on a single-dimensional anomaly detection method makes it difficult to effectively identify both sudden and gradual anomalies simultaneously.

[0073] Therefore, it is necessary to comprehensively evaluate the operating status of energy storage systems from different perspectives based on multi-channel time-series data. By constructing an anomaly detection model that can simultaneously characterize instantaneous difference features and trend deviation features, a unified quantitative score can be given to the system operating status, thereby improving the comprehensiveness and stability of anomaly detection results.

[0074] Therefore, in some embodiments, step 220 specifically includes steps 221 to 225: Step 221: Input the multi-channel observation vector sequence into the dual-domain attention encoder for encoding to obtain dual-domain fusion features; Step 222: Input the dual-domain fusion features into the decoder to obtain the reconstructed sequence; Step 223: Determine the residual strength based on the difference between the reconstructed sequence and the multi-channel observation vector sequence; Step 224: Determine the recent offset intensity based on the statistical offset of the channel data corresponding to the preset key indicators in the multi-channel observation vector sequence within the preset time range; Step 225: Weighted fusion of residual intensity and recent offset intensity to obtain anomaly score.

[0075] In this embodiment, the electronic device inputs a multi-channel observation vector sequence into a dual-domain attention encoder for feature encoding. The dual-domain attention encoder is used to jointly model the multi-channel observation vector sequence in both the channel and time dimensions to extract dual-domain fusion features that reflect the correlation between different data channels and the characteristics of data changes over time.

[0076] Subsequently, the electronic device inputs the dual-domain fused features into the decoder, which generates a reconstruction sequence corresponding to the multi-channel observation vector sequence. The reconstruction sequence is used to characterize the reconstruction result of the normal operating state of the energy storage system under the current model conditions.

[0077] Based on this, the electronic device determines the residual intensity according to the difference between the reconstructed sequence and the original multi-channel observation vector sequence. The residual intensity is used to characterize the degree of anomaly of the energy storage system over a short time period.

[0078] Simultaneously, the electronic equipment selects channel data corresponding to preset key indicators from the multi-channel observation vector sequence, and analyzes the statistical offset of the key channel data within a preset time range to determine the recent offset intensity. The recent offset intensity is used to characterize the degree of anomaly of the energy storage system over a recent long-term period.

[0079] Finally, the electronic device performs a weighted fusion of the residual intensity and the recent offset intensity to obtain an anomaly score that characterizes the overall anomaly degree of the energy storage system.

[0080] For example, an electronic device can calculate an anomaly score according to the following formula:

[0081] In the formula, Characterization t Anomaly scores at specific times are used to comprehensively characterize the energy storage system during this period. t The overall degree of anomaly at any given moment; These are weighting coefficients used to adjust the relative weights of residual strength and recent offset strength in anomaly scoring; they are typically set to... Residual strength Used to characterize the instantaneous deviation between the current operating state of the energy storage system and the normal mode depicted by the model; recent migration intensity. It is used to characterize the continuous deviation of key operating indicators from their normal statistical range within the most recent L time windows. The two are complementary in anomaly characterization and are helpful in distinguishing and identifying sudden anomalies (i.e., "sudden anomalies") and gradual deviation anomalies (i.e., "gradual deviation").

[0082] In the above embodiments, by using the same multi-channel observation vector sequence, the operating status of the energy storage system is evaluated from two complementary perspectives: instantaneous difference and time trend. Different types of abnormal features are uniformly mapped to abnormal scoring results. This avoids the risk of misjudgment caused by relying on a single abnormal criterion, improves the adaptability and stability of abnormal detection to complex operating conditions, and provides a reliable basis for subsequent abnormal event judgment and interpretation analysis.

[0083] In the analysis of operational data from energy storage systems, traditional anomaly detection methods typically focus on modeling single-dimensional features, such as the independent changes in each data channel or the evolution of data over time. This makes it difficult to simultaneously characterize the correlations between different operational indicators and the dependencies formed by changes in operational status over time. In multi-source, multi-indicator, and strongly coupled energy storage systems, this approach can easily lead to insufficient characterization of complex anomaly patterns, thus affecting the accuracy of anomaly detection results.

[0084] Based on this, the embodiments of this application introduce a dual-domain attention encoder in the anomaly detection model to model the energy storage system operation data from different dimensions.

[0085] The dual-domain attention encoder includes a feature domain encoder and a temporal domain encoder. The former is used to model the correlation between different operating metrics, while the latter is used to model the dependence of operating states in the time dimension, thereby providing a more comprehensive feature representation for subsequent anomaly detection.

[0086] Accordingly, in some embodiments, step 221 specifically includes: encoding the correlation between different channels in the multi-channel observation vector sequence in the feature domain encoder to obtain feature domain attention features; encoding the dependency of the multi-channel observation vector sequence along the time dimension in the time domain encoder to obtain time domain attention features; and fusing the channel domain attention features and the time domain attention features to obtain dual-domain fusion features.

[0087] In the feature domain encoder, for multi-channel observation vectors at the same time, the observation value corresponding to each channel is used as an independent feature input. The correlation between different channels is evaluated through attention calculation, and the features of each channel are weighted and combined according to the correlation to obtain feature domain attention features that can reflect the correlation between different operating indicators. For example, when there are significant coupled changes between indicators such as power, voltage, and state of charge, the feature domain encoder can assign higher correlation weights to these channels.

[0088] In the time-domain encoder, for multi-channel observation vectors at multiple consecutive time points, the observation results at different time points are correlated and modeled along the time dimension. An attention mechanism is used to calculate the degree of dependency between time points, and historical observation information is weighted and aggregated based on the degree of dependency, thereby obtaining time-domain attention features that reflect the temporal evolution characteristics of the energy storage system's operating state. For example, when there is a clear and continuous trend of change in the system's operating state between adjacent time points, the time-domain encoder can enhance the observation features within the corresponding time period.

[0089] After obtaining the attention features in the feature domain and the attention features in the time domain, the dual-domain attention encoder fuses them to form a dual-domain fused feature that simultaneously contains channel correlation information and time dependency information. This dual-domain fused feature serves as input to the subsequent processing of the anomaly detection model, supporting the reconstruction of the energy storage system's operating state and anomaly assessment.

[0090] In the above embodiments, by simultaneously characterizing the channel correlation and temporal dependence of multi-channel observation vector sequences within the same anomaly detection model, and forming a more comprehensive feature representation through dual-domain fusion, the modeling capability of the anomaly detection model for the complex operating characteristics of energy storage systems can be improved. Compared to encoding based on only a single dimension, this scheme is beneficial for enhancing the perception of multi-source coupling anomalies and temporal evolution anomalies, thereby improving the accuracy and stability of anomaly detection results.

[0091] In the multi-channel operation data of energy storage systems, the types and ranges of physical quantities corresponding to different channels vary. Anomalies often manifest as deviations in the operating status of some or more channels from the normal mode. Without a unified measurement of the deviations of each channel, it is difficult to form a quantitative indicator of anomalies that can reflect the overall operating status.

[0092] Therefore, in some embodiments, step 223 specifically includes: for each channel data in the multi-channel observation vector sequence, calculating the difference between each channel data and the reconstructed value of the corresponding channel in the reconstruction sequence; performing absolute value processing on each difference value, and performing aggregation operation on the difference values ​​after absolute value processing in all channels to obtain the residual intensity.

[0093] exist t At time 1, the electronic device takes the actual observation value of each channel in the multi-channel observation vector at that time and compares it with the reconstructed value at the same time and channel position in the reconstruction sequence, and calculates the difference between the two.

[0094] Subsequently, the electronic device performs absolute value processing on the difference values ​​to eliminate the sign effect caused by the actual observation values ​​being higher or lower than the reconstructed values, so that the difference only reflects the magnitude of the deviation.

[0095] After processing the absolute values ​​of the differences for each channel, the electronic device then performs a unified aggregation of the absolute difference values ​​for all channels, for example, by averaging them across the channel dimensions, to obtain the result. t The residual strength at time t.

[0096] For example, for a multi-channel observation vector sequence and its corresponding reconstruction sequence Electronic devices can calculate based on the following formula. t residual strength at time :

[0097] In the formula, n Represents a multi-channel observation vector sequence The number of channels; i Indicates the channel index, used to indicate the first... i One channel; Indicates at time t The i The residual values ​​for each channel. The residual value is the difference between the actual observed value and the corresponding reconstructed value for that channel. For example, the electronic device can calculate the residual values ​​based on the following formula:

[0098] In other words, electronic devices can assess the overall deviation of the actual observations of each data channel of the energy storage system from the model reconstruction results by calculating the residual intensity. For example, by taking the absolute value of the residuals of each channel and averaging them, the instantaneous anomaly intensity of the system's operating state can be quantified.

[0099] Through the above implementation methods, the embodiments of this application can uniformly map the reconstruction deviation of each channel into a single residual intensity index in a multi-channel operating data scenario, thereby avoiding the problem that it is difficult to directly compare the degree of anomaly of different channels, thus providing a consistent and quantifiable input basis for the subsequent calculation of anomaly scores.

[0100] During the operation of energy storage systems, some anomalies do not manifest as sudden deviations at a single moment, but rather as a trend of key operating indicators deviating from their normal fluctuation range over a period of time. Judging solely based on instantaneous differences makes it difficult to identify such gradually evolving anomalies in a timely manner.

[0101] Therefore, it is necessary to analyze the changes in the statistical characteristics of representative key operating indicators in the system's operating state within a certain time range, and to quantify the degree of deviation of the current observation value relative to its recent statistical characteristics in order to form a recent deviation intensity that can characterize trend anomalies.

[0102] Accordingly, in some embodiments, step 224 specifically includes: selecting multiple key channel data from the multi-channel observation vector sequence to characterize preset key operating indicators; calculating the statistical features corresponding to each key channel data within a preset time range; for any key channel data, calculating the normalized offset result corresponding to the key channel data based on the difference between the key channel data and the corresponding statistical mean, and in combination with the corresponding statistical dispersion; and aggregating the normalized offset results of each key channel data to obtain the recent offset intensity.

[0103] The preset time range is a time window consisting of the most recent consecutive time points, with the current time as the end time and going back backward.

[0104] The electronic equipment first selects multiple key channel data from the multi-channel observation vector sequence to characterize the critical operating states of the energy storage system. These key channel data correspond to operating indicators including, but not limited to, total system power, bus voltage, state of charge, or representative temperature, which reflect the core operating characteristics of the energy storage system and can be adjusted according to the actual system conditions or mission requirements.

[0105] For each key channel data, the electronic device acquires the observation values ​​of that channel at multiple consecutive moments within a preset time range, and calculates the corresponding statistical characteristics based on the observation values.

[0106] The statistical characteristics include at least the statistical mean of the key channel data over the time range and the statistical dispersion used to reflect the degree of fluctuation.

[0107] After obtaining the statistical characteristics, for any key channel data, the electronic device compares the observed value of that channel at the current moment with its corresponding statistical mean, and normalizes the difference between the two by combining the corresponding statistical dispersion, to obtain the normalized offset result of the key channel data at the current moment. The normalized offset result is used to characterize the degree of deviation of the key channel from its recent normal fluctuation range.

[0108] Based on this, the electronic equipment performs unified aggregation processing on the normalized offset results corresponding to all key channel data, such as averaging on the key channel dimension, to obtain the recent offset intensity corresponding to the current moment, which is used to characterize the overall offset trend of the energy storage system in the time dimension.

[0109] For example, the electronic device can select m A vector is composed of key indicators (such as total power, bus voltage, SOC, representative temperature, etc.). Calculate its rolling mean and rolling standard deviation over the most recent L time windows:

[0110]

[0111] In the formula, Represents the index vector of the first i Key indicators t The observed value at time; s This represents the time index, used to iterate through each moment within the time window; This indicates that within the most recent L time windows, the th iRolling average of key indicators; This indicates that within the most recent L time windows, the th i The rolling standard deviation of a key indicator relative to its rolling mean is used to characterize the normal range of fluctuation of that indicator.

[0112] Furthermore, the electronic device can calculate the recent offset intensity according to the following formula. :

[0113] In the formula, To prevent extremely small constants from being divided by zero, set .

[0114] The above formula is used to characterize the overall deviation of key operating indicators of an energy storage system from its normal statistical range over a recent period at the current moment. By averaging the standardized offsets of each key indicator, the strength of the trend anomaly in the system's operating status is quantified.

[0115] In the above embodiments, by quantitatively describing the trend deviation of the energy storage system's operating status based on the statistical characteristics changes of multiple key operating indicators over a period of time, it can reflect the degree to which key indicators continuously deviate from the normal fluctuation range. This is beneficial for identifying gradual anomalies and chronic degradation problems, and complements the anomaly measurement based on instantaneous differences. It provides a more comprehensive input basis for the comprehensive evaluation of anomaly scoring, thereby improving the accuracy and stability of anomaly detection results.

[0116] In the aforementioned embodiments, the anomaly detection model processes multi-channel observation vector sequences based on the trained model parameters during the application phase to determine anomaly scores and identify anomalous events. However, the detection performance of the anomaly detection model during the application phase depends on the quality of its learning of the normal operating modes of the energy storage system during the training phase.

[0117] If a model is trained solely on a single reconstruction task, it is prone to overfitting local features and struggles to simultaneously model both the current operational state and subsequent evolutionary trends. Furthermore, in multi-channel scenarios, different channels contribute differently to anomaly formation. Without constraints on attention distribution, the model may exhibit scattered attention across channels, negatively impacting the prominence of anomaly-related channels.

[0118] In this embodiment, the anomaly detection model comprises an encoder and a decoder. The encoder is a dual-domain attention encoder, used to extract features from multi-channel data sample sequences. The dual-domain attention encoder includes a feature domain encoder and a time domain encoder, which are used to model the correlation between different data channels and the dependencies formed by data changes over time, respectively.

[0119] The decoder includes a reconstruction decoder head and a prediction decoder head. The reconstruction decoder head is used to reconstruct the input sample sequence based on the features output by the encoder, and the prediction decoder head is used to predict the sample data at least one time after the current time step based on the features.

[0120] Based on the above model structure, the training process of the anomaly detection model includes the following steps: Step 310: Input the multi-channel data sample sequence from the training phase into the dual-domain attention encoder to be trained to obtain the training sample features; Step 320: Decode the features of the training samples using the reconstruction decoding head to obtain the reconstructed sample sequence, and determine the reconstruction deviation between the multi-channel data sample sequence and the reconstructed sample sequence at any given time. Step 330: Decode the features of the training samples using the prediction decoding head to obtain the prediction sample sequence, and determine the prediction deviation between the multi-channel data sample sequence and the prediction sample sequence at least one time after the corresponding time. Step 340: Based on the reconstruction bias, prediction bias and the attention matrix parameters at the corresponding time, construct the joint reconstruction-prediction loss function; wherein, the attention matrix parameters satisfy the group sparsity regularization constraint, so that the attention weights corresponding to a preset number of channels at the same time exceed the weight threshold, and the attention weights corresponding to the other channels do not exceed the weight threshold. Step 350: Train the anomaly detection model based on the reconstruction-prediction joint loss function to obtain the trained anomaly detection model.

[0121] During training, the electronic device can construct multi-channel data sample sequences from historical operating data and input these sequences into a dual-domain attention encoder in batches. The dual-domain attention encoder jointly models the relationships between different channels in the input samples and the evolution of the samples over time using both a feature domain encoder and a time domain encoder, thereby outputting the corresponding training sample features at each time step.

[0122] Subsequently, the electronic device simultaneously feeds the training sample features into both the reconstruction decoder and the prediction decoder. The reconstruction decoder, based on the training sample features, generates the current time step (…). t A reconstructed sample sequence (time step) corresponds one-to-one with the input sample in the time dimension, used to recover the observation values ​​of the input sample in each channel. The prediction decoder head then generates a sequence representing at least one time step after the current time step (e.g., time step 1) based on the features of the same training sample. t + h A predicted sample sequence (at any given time) is used to characterize the future trend of the sample over time.

[0123] After generating the reconstructed sample sequence and the predicted sample sequence, the electronic device calculates the reconstruction deviation between the input sample and the reconstructed sample, and the prediction deviation between the input sample and the predicted sample at each time step. The reconstruction deviation reflects the model's fit to the current operating state, while the prediction deviation reflects the model's ability to characterize the temporal evolution of the operating state.

[0124] Based on this, reconstruction bias and prediction bias are jointly incorporated into the training objective, and combined with the attention matrix parameters obtained during training, a joint reconstruction-prediction loss function for model training is constructed.

[0125] For example, the reconstruction-prediction joint loss function can be expressed as:

[0126] In the formula, For the reconstruction-prediction joint loss function; , and This represents the preset weighting coefficients, used to adjust the relative contributions of different loss terms in the joint loss function; This indicates that the difference is measured using the norm form; h This represents the step index, used to indicate the step number after the current time. h A future moment; k This represents the number of future moments involved in the calculation of the prediction deviation; Indicates in t The attention matrix parameters generated by the dual-domain attention encoder at each time step, This indicates that a group sparse regularization constraint is imposed on the attention matrix parameters.

[0127] The purpose of applying a set of sparse regularization constraints to the attention matrix parameters is to suppress the disordered diffusion of attention across multiple channels and to concentrate attention on a small number of channels with significant influence. In other words, this constraint allows only a subset of channels to be assigned significant attention weights at any given time, while compressing the attention weights of the remaining channels. For example, only the attention weights of some channels are kept at non-zero or high levels, while the attention weights of the remaining channels are limited to a preset range. This helps suppress interference from irrelevant channels in multi-channel scenarios, guiding the encoder to prioritize key channels that contribute significantly to anomaly detection during the modeling process, thus providing a more stable and distinguishable basis for subsequent anomaly scoring.

[0128] Finally, the electronic device updates the parameters of the anomaly detection model in reverse based on the reconstruction-prediction joint loss function, and obtains the trained anomaly detection model through multiple rounds of iterative training, which is then used for anomaly detection in subsequent application stages.

[0129] For example, based on the loss value calculated using this loss function, the model parameters of the dual-domain attention encoder, reconstruction decoder, and prediction decoder in the anomaly detection model are updated in reverse. This process of sample input, feature extraction, decoding, and parameter updating is repeated continuously during multiple training iterations until the joint loss function converges or a preset training termination condition is met, thus obtaining a trained anomaly detection model. Training termination conditions include, but are not limited to, minimizing the loss value, reaching a preset training duration, reaching a preset number of training rounds, or achieving the expected training effect verification metrics.

[0130] By introducing a reconstruction term, the anomaly detection model can accurately learn the normal operating mode of the multi-channel observation vector sequence at the current moment during training. By introducing a prediction term, the model simultaneously constrains its ability to characterize the evolution trend of the operating state over time during training. This allows the model to not only focus on the fitting of the current observation value, but also to perceive the changing characteristics of the operating state in the time dimension, thereby improving its ability to perceive subtle anomalies in the early stages.

[0131] The above training method can simultaneously constrain the anomaly detection model's ability to model the current operating state and operating trend during the training phase, enabling the model to take into account both static and dynamic features when learning normal operating modes, thereby improving the accuracy and reliability of anomaly detection.

[0132] In the process of anomaly detection in energy storage systems, simply identifying the anomaly itself is insufficient to support subsequent anomaly interpretation, cause analysis, and response decisions. If anomaly-related information is scattered across different data sources or time segments, it not only increases the cost of manual investigation but also hinders the direct application of anomaly detection results to intelligent interpretation and strategy recommendation.

[0133] Therefore, after an anomaly is identified, it is necessary to uniformly extract and structure the key data, model output results, and operational strategy information related to the anomaly to form an information set that can fully characterize the context of the anomaly, so as to serve as the basic input for subsequent causal analysis and interpretation.

[0134] Based on this, in some embodiments, step 230 specifically includes: determining the anomaly type and anomaly time period corresponding to the anomaly event; determining the changing trends of the anomaly score and the residual intensity and recent offset intensity used to generate the anomaly score within the anomaly time period; extracting key channel data related to the anomaly event from the multi-channel observation vector sequence within the anomaly time period, and determining the reconstruction sequence corresponding to the key channel data; extracting the operation strategy and operation strategy prompts executed within the anomaly time period from the energy storage system operation strategy; and organizing the anomaly type, anomaly score and its changing trend, residual intensity and its changing trend, recent offset intensity and its changing trend, key channel data and corresponding reconstruction sequences, operation strategies and operation strategy prompts according to the preset evidence field structure to form a structured anomaly evidence set corresponding to the anomaly event.

[0135] In this embodiment, the electronic device first determines the abnormal time period corresponding to the abnormal event, such as the abnormal score. If there are M consecutive time windows, then those M time windows are considered an abnormal time period. Based on the anomaly scoring method, the electronic device categorizes abnormal events into one or more preset anomaly types.

[0136] In some embodiments, anomaly types can be classified according to the anomaly manifestation, the affected object, and the cause dimension, and the same anomaly event can belong to a combination of multiple anomaly types at the same time.

[0137] For example, from the perspective of the form of abnormality, the types of abnormality can include, but are not limited to, sudden abnormalities and trend abnormalities. Sudden abnormalities are characterized by a rapid increase in the abnormality score in a short period of time, while trend abnormalities are characterized by a continuous shift in the abnormality score over a longer period of time.

[0138] From the perspective of the affected objects, the types of anomalies may include, but are not limited to, individual battery cell anomalies, battery cluster anomalies, power conversion unit anomalies, environmental and auxiliary machine anomalies, or system-level anomalies. Among them, individual battery cell anomalies may manifest as voltage or temperature anomalies in individual channels, while system-level anomalies may manifest as multiple key indicators deviating from the normal range simultaneously.

[0139] From the perspective of operational metrics, anomaly types can include, but are not limited to, power anomalies, voltage anomalies, current anomalies, state of charge anomalies, temperature anomalies, or revenue anomalies. Among them, revenue anomalies are used to characterize the deviation between the execution results of the operational strategy and the expected revenue.

[0140] From the perspective of operational strategy, anomaly types can also include strategy deviation anomalies, such as deviations in the execution of charging and discharging strategies, power constraint triggering anomalies, or scheduling timing anomalies, which reflect situations where the energy storage system deviates from the established operational strategy during actual operation.

[0141] In practical applications, the anomaly type of an abnormal event can be jointly determined based on one or more of the above dimensions and used as one of the constituent fields in the structured anomaly evidence set.

[0142] Subsequently, the electronic device extracts the sequence of anomaly scores changing over time during the anomaly period, and simultaneously extracts the changes in residual intensity and recent offset intensity used to generate the anomaly scores during that period, in order to form a trend reflecting the evolution of the anomaly.

[0143] Based on this, the electronic device filters key channel data with high correlation to the abnormal event from the multi-channel observation vector sequence within the abnormal time period. For example, when the anomaly is mainly triggered by the power and state of charge channels, only the observation sequence of the corresponding channel within the abnormal time period is extracted. At the same time, the reconstructed sequence corresponding one-to-one with the key channel data is extracted from the anomaly detection model output, which is used to compare the actual observed values ​​with the expected values ​​in the normal mode of the model.

[0144] Furthermore, the electronic device extracts information on the actual operating strategies executed during abnormal time periods from the energy storage system's operating strategy, such as charging and discharging strategies, power limiting strategies, or scheduling constraint information, and extracts operating strategy prompts corresponding to the operating strategies to characterize the intent or constraint background when formulating the strategies.

[0145] Finally, the electronic device organizes the anomaly type, anomaly time period, anomaly score and its trend, residual strength and its trend, recent offset strength and its trend, key channel data and corresponding reconstruction sequence, operation strategy and operation strategy prompts in a unified manner according to the preset evidence field structure, forming a structured anomaly evidence set corresponding to the anomaly event.

[0146] Through the above embodiments, abnormal information scattered across multiple data sources and multiple time dimensions can be uniformly integrated and abnormal events can be fully characterized in a structured form. This provides a clear and reusable evidence basis for subsequent abnormal cause analysis and interpretation, avoids the reliance on manual item-by-item investigation in the abnormal handling process, and improves the efficiency and consistency of the abnormal handling process.

[0147] After obtaining a structured set of anomaly evidence, the key to improving the intelligence level of anomaly detection systems lies in transforming this evidence into a reasonable explanation of the anomaly formation mechanism and actionable handling suggestions. Analysis based solely on rules or simple matching methods fails to fully utilize historical knowledge and complex causal relationships, and is also detrimental to generating targeted handling solutions for different operational scenarios.

[0148] Based on this, in some embodiments, step 240 specifically includes: inputting a structured set of anomaly evidence into a language model, so that the language model constructs query conditions for knowledge retrieval based on the structured set of anomaly evidence, and retrieves mechanism path information and historical anomaly case information related to the anomaly event from a pre-built knowledge graph of the energy storage system based on the query conditions; performing causal relationship inference based on the mechanism path information and historical anomaly case information to obtain the cause of the anomaly; generating an explanation of the cause of the anomaly based on the cause of the anomaly and the mechanism path information associated with the cause of the anomaly; and generating an execution suggestion corresponding to the anomaly event based on the cause of the anomaly, the explanation of the cause of the anomaly, and the operation strategy and operation strategy prompts corresponding to the anomaly event.

[0149] In some embodiments, the electronic device encodes the structured anomaly evidence set into a structured anomaly description and inputs it into a language model. The language model then constructs query conditions for knowledge retrieval based on anomaly type, anomaly score and its constituent information, anomaly time range, key channel data sets, and AI strategy intent prompts. These query conditions may include anomaly type identifiers, key operational indicator names, operational scenario information where the anomaly occurred, and feature descriptions related to operational benefits or strategy execution.

[0150] Subsequently, based on the query conditions, the electronic device retrieves mechanistic path information and historical anomaly case information related to the abnormal event from a pre-built knowledge graph of the energy storage system. For example, when anomaly evidence shows that the temperature channel is consistently high and the power is limited, it can retrieve mechanistic paths related to "temperature increase - internal resistance change - efficiency decrease" and similar historical cases, as well as historical anomaly case information corresponding to the mechanistic path, and use the retrieval results as external knowledge input to the language model.

[0151] After acquiring information on the mechanistic pathways and historical anomaly cases, the language model infers causal relationships based on this information, identifying at least one possible cause of the anomaly. For example, it might determine that the anomaly was caused by decreased heat dissipation capacity or abnormal ambient temperature.

[0152] In some embodiments, the language model generates a set of candidate anomaly causes based on the structured set of anomaly evidence and external knowledge input. For any candidate anomaly cause, the language model can further provide a set of key variables or devices corresponding to that anomaly cause and generate condition items for verifying the anomaly cause. The condition items may include inspection requirements for runtime data, event logs, or configuration parameters, and generate corresponding confidence scores for each candidate anomaly cause.

[0153] Subsequently, based on the cause of the anomaly and the mechanistic path information associated with it, the language model generates an explanation of the cause of the anomaly to describe the process of its formation, thereby describing the causal chain of the anomaly in natural language.

[0154] Finally, combining the cause of the anomaly, its explanation, and the corresponding operational strategies and policy hints during the anomaly, the language model generates execution suggestions for the anomaly. For example, it might suggest checking the cooling system, adjusting the power limit, or temporarily modifying the scheduling strategy.

[0155] In some embodiments, the system pre-maintains a verifiable capability mapping table to indicate whether various operational data, log fields, and configuration parameters can be automatically obtained through the system interface. For each candidate anomaly cause output by the language model, the electronic device maps the condition items to machine-readable verification items and determines the corresponding verification method based on the verifiable capability mapping table: when all verification items corresponding to a candidate anomaly cause can be automatically obtained, the candidate anomaly cause is marked as automatically verified, and the corresponding data verification and rule judgment are triggered; when at least one verification item requires manual confirmation, the candidate anomaly cause is marked as requiring manual verification or a combination of automatic and manual verification, and the verification steps that can be completed automatically and the suggested manual inspection steps are listed in the execution suggestions respectively. Subsequently, based on the candidate anomaly causes, anomaly cause explanations, and verification method determination results, the language model generates anomaly cause explanations to illustrate the anomaly formation mechanism, and further generates execution suggestions corresponding to the anomaly event, thereby explaining the cause of the anomaly, the evidence, and the subsequent executable verification and handling solutions to the operation and maintenance personnel in natural language.

[0156] Therefore, through the above embodiments, it is possible to perform interpretable causal analysis on abnormal events based on structured anomaly evidence and combined with knowledge of the energy storage system domain, and generate execution suggestions that match the actual operation strategy. This upgrades anomaly handling from simple alarms to an "interpretable, verifiable, and executable" intelligent decision support process, which helps to improve the operation and maintenance efficiency and handling accuracy of energy storage systems.

[0157] The following description, in conjunction with the accompanying drawings, illustrates a specific example of the anomaly detection and interpretation method for energy storage systems provided in this application.

[0158] Figure 3 This is a flowchart illustrating the anomaly detection and interpretation process for energy storage systems provided in some embodiments of this application. For example... Figure 3 As shown, this invention provides an anomaly detection and interpretation method for multi-source time-series data of energy storage systems. It constructs a complete processing chain from anomaly identification and evidence organization to causal interpretation and report generation, enabling automatic anomaly discovery, traceable analysis, and the output of actionable remediation suggestions. The goal is to achieve end-to-end anomaly detection and attribution interpretation. The method includes steps one through six.

[0159] Step 1: Data Acquisition and Unified Preprocessing In this step, the electronic device collects multi-source time-series data from multiple data sources, including the battery management system, energy management system, environment and auxiliary equipment, communication and event logs, and control commands. At the same time, it collects input data related to the energy storage system operation strategy, including electricity price series, photovoltaic and load forecast information, grid connection constraints, power constraints, guaranteed SOC constraints, and actual control commands corresponding to each cycle, to form the original multi-source time-series data set.

[0160] Subsequently, multi-source time-series data from different data sources are aligned to a unified time base. Short-term missing data are imputed using linear interpolation, while long-term missing data are marked and skipped during the training phase. Simultaneously, a moving median filter is used to suppress sampling spikes, and significant outliers are labeled based on Z-scores to reduce their weighting impact during training. Furthermore, each data channel is standardized to meet zero mean, unit variance, or a preset normalization interval. Then, the processed multi-source data is sliced ​​using a sliding window of preset length to form a multi-channel observation vector sequence, and a prediction step size is set for subsequent modeling.

[0161] Through the above processing, a multi-channel observation vector sequence with consistent dimensions that can be directly input into the model is obtained.

[0162] Step 2: Modeling Temporal Relationships Using a Dual-Domain Attention Mechanism In this step, a dual-domain attention encoder is constructed to model the multi-channel observation vector sequence. Specifically, the correlation between different data channels is modeled in the feature domain, and the long-term and short-term dependencies of the same channel along the time dimension are modeled in the time domain. A fused representation (i.e., dual-domain fused features) is obtained through mutual attention and gating fusion. Based on the fused representation, the decoder simultaneously outputs the reconstructed sequence at the current time step and the future sequence. k The predicted sequence at each time point.

[0163] During the training phase, the model is optimized by constructing a joint reconstruction-prediction loss function. The reconstruction term constrains the model's learning accuracy for the current normal operating mode, while the prediction term enhances the model's ability to represent operating trends. Combined with group sparse regularization constraints applied to the attention matrix, this highlights a small number of variables related to anomalies, providing a foundation for subsequent localization.

[0164] Step 3: Unsupervised anomaly representation learning and anomaly scoring This step constructs a uniform outlier score. It consists of two parts: one is the residual term, which measures the degree of inconsistency between the current observation and the model's expectations; the other is the recent offset term, which measures the deviation of key indicators from the most recent normal level.

[0165] Therefore, without the need for manual labeling of abnormal samples, the normal behavior patterns of energy storage systems during historical operation can be learned, and the degree of deviation between current observations and normal patterns can be measured. This enables a unified characterization of "sudden anomalies" and "gradual shifts," avoiding dependence on fixed thresholds or single anomaly forms.

[0166] Step 4: Evidence Construction and Packaging Upon detecting an anomaly, the electronic device structures and organizes information related to the anomaly over the time period in which it occurred, forming an anomaly evidence set. This evidence includes not only the anomaly type, anomaly score, and its changes over time, but also segments of key operational variables highly correlated with the anomaly and their corresponding model reconstruction results, thus providing a clear picture of the source of the anomaly deviation.

[0167] At the same time, the electronic equipment also incorporates the operating strategies executed during the anomaly and their strategic intent, and introduces information on the deviation of benefits or risks compared with the baseline operating plan, so that the evidence of anomalies can reflect both the operational status problem and the impact at the strategic level.

[0168] The aforementioned evidence was constructed with consistent inclusion of data source and timestamp information to ensure its traceability and verifiability.

[0169] Step 5: Knowledge Retrieval and Causal Inference Electronic devices construct query conditions based on anomaly type and evidence summary, and retrieve mechanistic paths and similar historical cases related to anomalies from a pre-built knowledge graph of energy storage systems. For example, electronic devices can retrieve battery internal resistance change mechanisms related to temperature anomalies, or profit / loss paths related to strategy execution deviations. Based on this, and combined with current anomaly evidence, causal inferences are made to obtain a ranking of possible anomaly causes. At the same time, observation points that need further verification are marked to support the credibility of subsequent explanations of "why this judgment is made".

[0170] Step Six: Generate Explanation Report Electronic devices input causal inference results and a set of abnormal evidence into a language model to generate an anomaly explanation report tailored to operational scenarios. The explanation report summarizes the abnormal event in natural language and, combined with specific evidence, provides an analysis of the cause of the anomaly, the corresponding confidence level, and actionable verification steps and handling suggestions. In cases where operational benefits deviate, the report can further provide impact assessment results under different execution strategies and offer traceable correlation information.

[0171] For example, first summarize "what happened" in natural language, then present "key evidence and quantitative values" one by one, followed by "possible causes and their confidence levels", "executable verification steps and handling suggestions", and when there is a profit deviation, attach "expected profit improvement by executing alternative instructions" to support operation and maintenance personnel in understanding, verifying and making decisions on anomalies.

[0172] The core of this application's embodiments lies in constructing a multi-model collaborative framework for time-series anomaly detection and intelligent interpretation. At the anomaly detection level, an integrated architecture combining self-supervised representation learning and unsupervised anomaly scoring is adopted. Through unified data alignment, missing data imputation, resampling, normalization, and feature processing of multi-source time-series data, multi-modal time-series data such as electricity price, power, voltage, temperature, state of charge, and event logs are mapped to a shared representation space. Furthermore, based on a joint learning approach combining reconstruction and prediction tasks, the normal operation mode of the energy storage system is modeled, thereby uniformly quantifying and evaluating the residual and statistical features generated during operation, forming anomaly scoring results with a consistent scale, and providing a confidence basis for anomaly judgment.

[0173] At the level of anomaly interpretation and handling, this application further introduces a reasoning mechanism that combines a large language model with a knowledge graph in the energy storage field. The anomaly scores, key data fragments, and historical similar anomaly cases obtained during the anomaly detection process are organized in a structured form as anomaly evidence. Based on the anomaly evidence, retrieval conditions are constructed, and the mechanism path and historical case information related to the anomaly event are obtained from the knowledge graph. On this basis, the large language model performs causal reasoning to generate traceable and auditable anomaly cause explanations. At the same time, it provides verification steps and handling or operation strategy suggestions corresponding to the anomaly cause, upgrading the anomaly alarm from merely indicating anomaly phenomena to a decision plan with sufficient evidence and direct execution, thereby significantly shortening the time for anomaly diagnosis and handling.

[0174] Through the above methods, the embodiments of this application can not only achieve automated detection of anomalies in energy storage systems, but also output highly reliable and targeted explanations of the causes of anomalies and strategic recommendations, and have the ability to adapt to different operating scenarios and new operating conditions. Compared with traditional solutions, the embodiments of this application have significant improvements in the accuracy of anomaly detection, the quality of anomaly explanation, and the overall scalability of the system.

[0175] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0176] The anomaly detection and interpretation method for energy storage systems provided in this application can be executed by an anomaly detection and interpretation device for energy storage systems. This application uses the example of an anomaly detection and interpretation device for energy storage systems executing the anomaly detection and interpretation method for energy storage systems to illustrate the anomaly detection and interpretation device for energy storage systems provided in this application.

[0177] like Figure 4 As shown, the anomaly detection and interpretation device for energy storage systems includes: a data acquisition module 401, an anomaly detection module 402, and an anomaly interpretation module 403.

[0178] The data acquisition module 401 is used to acquire multi-source time-series data of the energy storage system and perform multi-channel time alignment processing on the multi-source time-series data to obtain a multi-channel observation vector sequence.

[0179] Anomaly detection module 402 is used to determine anomaly scores based on multi-channel observation vector sequences and anomaly detection models.

[0180] The anomaly interpretation module 403 is used to determine the anomaly event when the anomaly score meets the preset alarm conditions, and to construct a structured set of anomaly evidence corresponding to the anomaly event.

[0181] The anomaly explanation module 403 is also used to input a structured set of anomaly evidence into a language model to perform causal analysis on the anomaly event, generate anomaly cause explanations and corresponding execution suggestions.

[0182] According to the anomaly detection and interpretation device for energy storage systems provided in this application embodiment, multi-channel time alignment is performed on multi-source time-series data of the energy storage system to construct a multi-channel observation vector sequence. Based on the anomaly detection model, the operating status of the energy storage system is quantitatively evaluated to determine the anomaly score. When the anomaly score meets the alarm conditions, a structured anomaly evidence set corresponding to the anomaly event is constructed. The structured anomaly evidence set is then input into a language model for causal analysis. Thus, while accurately detecting anomaly events in the energy storage system, it can automatically generate anomaly cause explanations and actionable handling suggestions corresponding to the formation mechanism of the anomaly event, upgrading anomaly handling from merely providing alarm information to evidence-based interpretation and decision support.

[0183] In some embodiments, the anomaly detection module is further configured to input the multi-channel observation vector sequence into a dual-domain attention encoder for encoding to obtain dual-domain fusion features; input the dual-domain fusion features into a decoder to obtain a reconstructed sequence; determine the residual intensity based on the difference between the reconstructed sequence and the multi-channel observation vector sequence; determine the recent offset intensity based on the statistical offset of the channel data corresponding to the preset key indicators in the multi-channel observation vector sequence within a preset time range; and perform weighted fusion of the residual intensity and the recent offset intensity to obtain an anomaly score.

[0184] In some embodiments, the dual-domain attention encoder includes a feature domain encoder and a time domain encoder; the anomaly detection module is further configured to encode the correlation between different channels in the multi-channel observation vector sequence in the feature domain encoder to obtain feature domain attention features; to encode the dependency of the multi-channel observation vector sequence along the time dimension in the time domain encoder to obtain time domain attention features; and to fuse the channel domain attention features and the time domain attention features to obtain dual-domain fused features.

[0185] In some embodiments, the anomaly detection module is further configured to calculate the difference between each channel data and the reconstructed value of the corresponding channel in the reconstructed sequence for each channel data in the multi-channel observation vector sequence; perform absolute value processing on each difference value; and perform aggregation operation on the difference values ​​after absolute value processing in all channels to obtain the residual intensity.

[0186] In some embodiments, the anomaly detection module is further configured to select multiple key channel data from the multi-channel observation vector sequence to characterize preset key operating indicators; calculate statistical features corresponding to each key channel data within a preset time range; the statistical features include statistical mean and statistical dispersion; for any key channel data, calculate the normalized offset result corresponding to the key channel data based on the difference between the key channel data and the corresponding statistical mean, and in combination with the corresponding statistical dispersion; and aggregate the normalized offset results of each key channel data to obtain the recent offset intensity.

[0187] In some embodiments, the decoder includes a reconstruction decoding head and a prediction decoding head; the above apparatus further includes a model training module, used to input the multi-channel data sample sequence of the training phase into the dual-domain attention encoder to be trained to obtain training sample features; decode the training sample features through the reconstruction decoding head to obtain a reconstruction sample sequence, and determine the reconstruction deviation between the multi-channel data sample sequence and the reconstruction sample sequence at any time; decode the training sample features through the prediction decoding head to obtain a prediction sample sequence, and determine the prediction deviation between the multi-channel data sample sequence and the prediction sample sequence at at least one time after the corresponding time; construct a joint reconstruction-prediction loss function based on the reconstruction deviation, the prediction deviation, and the attention matrix parameters at the corresponding time; wherein, the attention matrix parameters satisfy the group sparsity regularization constraint, so that the attention weights corresponding to a preset number of channels at the same time exceed the weight threshold, and the attention weights corresponding to the remaining channels do not exceed the weight threshold; train the anomaly detection model based on the joint reconstruction-prediction loss function to obtain a trained anomaly detection model.

[0188] In some embodiments, the data acquisition module is further configured to collect multi-source time-series data from multiple data sources and acquire strategy data corresponding to the energy storage system operation strategy to form an original multi-source time-series data set; align each data in the original multi-source time-series data set to a unified time reference to obtain time-aligned multi-source time-series data; based on the time-aligned multi-source time-series data, perform missing data processing, outlier sampling point suppression, and scale unification processing respectively to obtain dimensionally consistent multi-channel time-series data; based on the dimensionally consistent multi-channel time-series data, perform sliding window slicing according to a preset window length to construct a multi-channel observation vector sequence for anomaly detection model input.

[0189] In some embodiments, the anomaly interpretation module is further configured to determine the anomaly type and anomaly time period corresponding to the anomaly event; determine the anomaly score and the changing trends of the residual intensity and recent offset intensity used to generate the anomaly score within the anomaly time period; extract key channel data related to the anomaly event from the multi-channel observation vector sequence within the anomaly time period, and determine the reconstruction sequence corresponding to the key channel data; extract the operation strategy and operation strategy prompts executed within the anomaly time period from the energy storage system operation strategy; and organize the anomaly type, anomaly score and its changing trend, residual intensity and its changing trend, recent offset intensity and its changing trend, key channel data and corresponding reconstruction sequences, operation strategies and operation strategy prompts according to a preset evidence field structure to form a structured anomaly evidence set corresponding to the anomaly event.

[0190] In some embodiments, the anomaly explanation module is further configured to input a structured set of anomaly evidence into a language model, so that the language model constructs query conditions for knowledge retrieval based on the structured set of anomaly evidence, and retrieves mechanism path information and historical anomaly case information related to the anomaly event from a pre-built energy storage system knowledge graph based on the query conditions; performs causal relationship inference based on the mechanism path information and historical anomaly case information to obtain the cause of the anomaly; generates an explanation of the cause of the anomaly based on the cause of the anomaly and the mechanism path information associated with the cause of the anomaly; and generates an execution suggestion corresponding to the anomaly event based on the cause of the anomaly, the explanation of the cause of the anomaly, and the operation strategy and operation strategy prompts corresponding to the anomaly event.

[0191] The anomaly detection and interpretation device for energy storage systems in this application embodiment can be an electronic device or a component within an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal or other devices besides a terminal, such as a server.

[0192] The anomaly detection and interpretation device for energy storage systems provided in this application embodiment can realize all the processes implemented in the above-described anomaly detection and interpretation method embodiment for energy storage systems. To avoid repetition, these processes will not be described again here.

[0193] In some embodiments, such as Figure 5 As shown, this application embodiment also provides an electronic device 500, including a processor 501, a memory 502, and a computer program stored in the memory 502 and executable on the processor 501. When the program is executed by the processor 501, it implements the various processes of the above-described embodiment of the anomaly detection and interpretation method for energy storage systems and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0194] This application provides a non-transitory computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the above-described embodiment of the anomaly detection and interpretation method for energy storage systems and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0195] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-storable media, such as computer read-only memory (ROM), random-access memory (RAM), magnetic disks, or optical disks.

[0196] The computer-readable storage medium may include: read-only memory (ROM), random-access memory (RAM), magnetic disk or optical disk, etc.

[0197] This application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described anomaly detection and interpretation method for energy storage systems.

[0198] This application provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-described embodiments of the anomaly detection and interpretation method for energy storage systems, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0199] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0200] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0201] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the related technology, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0202] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

[0203] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0204] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for anomaly detection and interpretation in energy storage systems, characterized in that, include: Acquire multi-source time-series data of the energy storage system, and perform multi-channel time alignment processing on the multi-source time-series data to obtain a multi-channel observation vector sequence; Based on the multi-channel observation vector sequence, an anomaly score is determined using an anomaly detection model; If the anomaly score meets the preset alarm conditions, an anomaly event is identified, and a structured set of anomaly evidence corresponding to the anomaly event is constructed. The structured set of anomaly evidence is input into a language model to perform causal analysis on the anomaly events, generate explanations of the anomaly causes, and corresponding execution suggestions.

2. The anomaly detection and interpretation method for energy storage systems according to claim 1, characterized in that, The process of determining anomaly scores based on the multi-channel observation vector sequence using an anomaly detection model includes: The multi-channel observation vector sequence is input into a dual-domain attention encoder for encoding to obtain dual-domain fusion features; The dual-domain fusion features are input into the decoder to obtain the reconstructed sequence; The residual strength is determined based on the difference between the reconstructed sequence and the multi-channel observation vector sequence; Based on the statistical offset of the channel data corresponding to the preset key indicators in the multi-channel observation vector sequence within a preset time range, the recent offset intensity is determined. The residual intensity and recent offset intensity are weighted and fused to obtain an anomaly score.

3. The anomaly detection and interpretation method for energy storage systems according to claim 2, characterized in that, The dual-domain attention encoder includes a feature-domain encoder and a temporal-domain encoder; The step of inputting the multi-channel observation vector sequence into a dual-domain attention encoder for encoding to obtain dual-domain fusion features includes: The correlation between different channels in the multi-channel observation vector sequence is encoded in the feature domain encoder to obtain the feature domain attention feature; Temporal attention features are obtained by encoding the dependencies of the multi-channel observation vector sequence along the time dimension in the temporal encoder. The channel domain attention features and the time domain attention features are fused to obtain dual-domain fused features.

4. The anomaly detection and interpretation method for energy storage systems according to claim 2 or 3, characterized in that, Determining the residual strength based on the difference between the reconstructed sequence and the multi-channel observation vector sequence includes: For each channel data in the multi-channel observation vector sequence, calculate the difference between each channel data and the reconstructed value of the corresponding channel in the reconstructed sequence; The absolute values ​​of each difference are processed, and the difference values ​​after absolute value processing are aggregated across all channels to obtain the residual intensity.

5. The anomaly detection and interpretation method for energy storage systems according to claim 2 or 3, characterized in that, The determination of recent offset intensity based on the statistical offset of channel data corresponding to preset key indicators in the multi-channel observation vector sequence within a preset time range includes: From the multi-channel observation vector sequence, select multiple key channel data to characterize preset key operating indicators; Within a preset time range, calculate the statistical characteristics corresponding to the data of each key channel; the statistical characteristics include the statistical mean and the statistical dispersion. For any key channel data, based on the difference between the key channel data and the corresponding statistical mean, and combined with the corresponding statistical dispersion, calculate the normalized offset result corresponding to the key channel data. The normalized offset results of each key channel data are aggregated to obtain the recent offset intensity.

6. The anomaly detection and interpretation method for energy storage systems according to claim 2, characterized in that, The decoder includes a reconstruction decoder head and a prediction decoder head; The training steps of the anomaly detection model include: The multi-channel data sample sequence during the training phase is input into the dual-domain attention encoder to be trained to obtain the training sample features; The training sample features are decoded by the reconstruction decoding head to obtain the reconstructed sample sequence, and the reconstruction deviation between the multi-channel data sample sequence and the reconstructed sample sequence at any time is determined. The training sample features are decoded by a prediction decoding head to obtain a prediction sample sequence, and the prediction deviation between the multi-channel data sample sequence and the prediction sample sequence is determined at least one time after the corresponding time. Based on the reconstruction bias, prediction bias and attention matrix parameters at the corresponding time, a joint reconstruction-prediction loss function is constructed; wherein, the attention matrix parameters satisfy group sparsity regularization constraints, so that the attention weights corresponding to a preset number of channels at the same time exceed the weight threshold, and the attention weights corresponding to the remaining channels do not exceed the weight threshold. The anomaly detection model is trained based on the reconstruction-prediction joint loss function to obtain a trained anomaly detection model.

7. The anomaly detection and interpretation method for energy storage systems according to claim 1, characterized in that, The step of performing multi-channel time alignment processing on the multi-source time series data to obtain a multi-channel observation vector sequence includes: Multi-source time-series data are collected from multiple data sources, and strategy data corresponding to the energy storage system's operation strategy are obtained to form an original multi-source time-series data set; Align each data point in the original multi-source time series data set to a unified time reference to obtain time-aligned multi-source time series data. Based on the time-aligned multi-source time-series data, missing data processing, outlier sampling point suppression, and scale unification processing are performed respectively to obtain multi-channel time-series data with consistent dimensions. Based on the dimensionally consistent multi-channel time-series data, a sliding window slice is performed according to a preset window length to construct a multi-channel observation vector sequence for input to the anomaly detection model.

8. The anomaly detection and interpretation method for energy storage systems according to claim 1, characterized in that, Construct a set of structured anomaly evidence corresponding to the anomalous event, including: Determine the anomaly type and time period corresponding to the anomaly event; Determine the changing trends of the abnormal score and the residual strength and recent offset strength used to generate the abnormal score within the abnormal time period; During the abnormal time period, key channel data related to the abnormal event are extracted from the multi-channel observation vector sequence, and the reconstruction sequence corresponding to the key channel data is determined. Extract the operating strategies and operational strategy prompts executed during the abnormal time period from the energy storage system's operating strategy; According to the preset evidence field structure, the anomaly type, anomaly score and trend, residual strength and trend, recent offset strength and trend, key channel data and corresponding reconstruction sequence, operation strategy and operation strategy prompts are organized to form a structured anomaly evidence set corresponding to the anomaly event.

9. The anomaly detection and interpretation method for energy storage systems according to claim 1 or 8, characterized in that, The step of inputting the structured set of anomaly evidence into a language model to perform causal analysis on the anomaly event, generating explanations of the anomaly causes and corresponding execution suggestions, includes: The structured set of abnormal evidence is input into a language model, so that the language model can construct query conditions for knowledge retrieval based on the structured set of abnormal evidence, and retrieve mechanism path information and historical abnormal case information related to the abnormal event from a pre-built knowledge graph of energy storage system based on the query conditions. Based on the aforementioned mechanism path information and historical anomaly case information, causal relationship inference is performed to obtain the cause of the anomaly; Based on the cause of the anomaly and the mechanistic path information associated with the cause of the anomaly, an explanation of the cause of the anomaly is generated; Based on the cause of the anomaly, the explanation of the cause of the anomaly, and the corresponding operating strategy and operating strategy prompts, an execution suggestion corresponding to the anomaly is generated.

10. An anomaly detection and interpretation device for energy storage systems, characterized in that, The device includes: The data acquisition module is used to acquire multi-source time-series data of the energy storage system and perform multi-channel time alignment processing on the multi-source time-series data to obtain a multi-channel observation vector sequence; Anomaly detection module is used to determine anomaly score based on the multi-channel observation vector sequence using an anomaly detection model; An anomaly interpretation module is used to determine an anomaly event when the anomaly score meets preset alarm conditions, and to construct a structured set of anomaly evidence corresponding to the anomaly event. The anomaly explanation module is further configured to input the structured anomaly evidence set into a language model to perform causal analysis on the anomaly event, generate anomaly cause explanations and corresponding execution suggestions.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the anomaly detection and interpretation method for energy storage systems as described in any one of claims 1-9.

12. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the anomaly detection and interpretation method for energy storage systems as described in any one of claims 1-9.

Images