A deep learning-based energy storage anomaly monitoring method and system

Abstract

This invention relates to the field of energy storage anomaly monitoring technology, and more particularly to a deep learning-based energy storage anomaly monitoring method and system. The method involves acquiring operational data from the historical stable operation of an energy storage system to construct a stable operation dataset. Joint feature modeling is then performed on this data to generate operational behavior embedding vectors and form operational behavior sequences. Furthermore, an autoencoder neural network model is used to encode the features of the operational behavior sequences, learning the behavioral change patterns under stable operating conditions of the energy storage system and constructing a normal behavior trajectory vector. Real-time operational data is then input into the model to generate the current behavior embedding vector, and its behavioral offset feature vector relative to the normal behavior trajectory vector is calculated. Anomaly determination is then performed based on the continuous offset of this feature within a continuous time window. This allows for early identification of abnormal states before operating parameters exceed threshold ranges, improving the accuracy and timeliness of energy storage system anomaly monitoring.

Description

Technical Field

[0001] This invention relates to the field of energy storage anomaly monitoring technology, and more particularly to an energy storage anomaly monitoring method and system based on deep learning. Background Technology

[0002] Energy storage systems are increasingly being used in photovoltaic power plants. These systems typically consist of numerous individual battery cells, battery modules, and a battery management system. During operation, they generate multi-dimensional operational data, including voltage, current, power, battery temperature distribution, state of charge, state of health, and charge / discharge rate. Real-time monitoring and analysis of this data allows for the timely detection of abnormal states in the energy storage system, thereby preventing safety risks such as battery overheating, capacity degradation, and overcharging / over-discharging.

[0003] In existing technologies, anomaly monitoring in energy storage systems typically employs threshold monitoring, which involves setting preset threshold ranges for operating parameters such as voltage, current, temperature, or state of charge (SOC). An alarm is triggered when the monitored parameter exceeds the preset threshold. This method is simple to implement and has low deployment costs, thus it is widely used in practical energy storage system operation monitoring. However, under different operating conditions, the operating parameters of energy storage systems often exhibit complex temporal variations, and there are significant coupling relationships between different parameters. Simply relying on fixed thresholds for anomaly detection cannot accurately reflect the overall operating status of the energy storage system. Furthermore, when the operating status of the energy storage system gradually deviates from the normal operating mode but has not yet reached the threshold range, traditional threshold monitoring methods often fail to identify such abnormal changes in a timely manner, easily leading to delayed anomaly detection or misjudgment. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method and system for monitoring energy storage anomalies based on deep learning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based method for monitoring energy storage anomalies includes the following steps: S1. Obtain voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate and BMS control command data during the historical stable operation of the energy storage system to obtain a stable operation dataset; S2. Perform joint feature modeling on the stable running dataset to obtain running behavior embedding vectors and arrange them in chronological order to obtain a running behavior sequence; S3. Based on the operating behavior sequence, the operating behavior embedding vector is input into an autoencoder neural network model to perform feature encoding processing to obtain a low-dimensional behavior representation vector, and the normal behavior trajectory vector of the energy storage system is constructed according to the change relationship of the low-dimensional behavior representation vector in the time series. S4. Obtain the real-time operation dataset during the real-time operation of the energy storage system, input the real-time operation dataset into the autoencoder neural network model to generate the current behavior embedding vector, and calculate the behavior offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector. S5. Perform anomaly detection processing based on the continuous offset of the behavior offset feature vector within a continuous time window to obtain anomaly monitoring results.

[0006] Further, S1 includes the following steps: The battery management system (BMS) and environmental monitoring devices of the energy storage system collect data on voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate, and BMS control commands during the stable operation of the energy storage system, and add a unified timestamp to obtain the original operating dataset. Time alignment processing is performed on the original running dataset, and the data is synchronously rearranged according to the timestamp identifier at a uniform sampling time interval to obtain a multidimensional running data sequence arranged in chronological order. Data cleaning and standardization are performed on the multidimensional running data sequence to remove missing data and outlier sampling points, and normalization mapping is performed on the running parameters of each dimension to obtain a stable running dataset.

[0007] Further, S2 includes the following steps: Based on the stable operation dataset, the voltage, current, power and battery temperature distribution are divided into time segments according to a preset time window, and the operation behavior segment data is constructed by combining SOC, SOH, charge and discharge rate and BMS control command status. The mean, variance, rate of change, and fluctuation amplitude of voltage, current, power, and battery temperature distribution in the operation behavior segment data are calculated respectively, and the temperature distribution gradient characteristics are calculated based on the temperature difference between battery modules to obtain the operation feature vector; The SOC, SOH, charge / discharge rate and BMS control command data are constructed into a state vector according to a time window, and the state vector and the running feature vector are concatenated according to the feature dimension to obtain the running behavior embedding vector. The embedded vector of the running behavior is processed by performing a sequence arrangement in chronological order to obtain the running behavior sequence.

[0008] Further, S3 includes the following steps: The running behavior sequence is divided into several training sample sequences according to time order, and the running behavior embedding vector of the first few time steps in each training sample sequence is input into the encoder of the autoencoder neural network model. Temporal feature compression calculation is performed on the running behavior embedding vector of the first few time steps to obtain the corresponding historical behavior representation vector. The historical behavior representation vector is input into the decoder of the autoencoder neural network model, and future behavior prediction calculation is performed on the historical behavior representation vector to obtain the predicted behavior vector for the corresponding subsequent time step. The prediction error is calculated based on the difference between the predicted behavior vector and the actual subsequent time step running behavior embedding vector. The model training loss is calculated using a loss function based on the prediction error, and backpropagation is performed on the autoencoder neural network model to update the network parameters based on the model training loss until the model converges. Based on the trained autoencoder neural network model, trajectory connection processing is performed on the relationship between the predicted behavior vector of subsequent time steps under stable operation and the change in the time series to obtain the normal behavior trajectory vector of the energy storage system.

[0009] Furthermore, the loss function is as follows: ; in, Use the loss value to train the model; t is the number of training samples; t is the index of the current time step in the training sample sequence; k is the prediction step size. The embedding vector of the actual running behavior of the i-th training sample at time t+k; The autoencoder neural network model predicts the behavior vector at time step (t+k) based on the embedded vector of the running behavior before time step t. These are the weighting coefficients for trajectory smoothing constraints.

[0010] Further, calculating the behavior offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector includes the following steps: The trajectory distance offset is calculated based on the minimum distance between the current low-dimensional behavior representation vector and the normal behavior trajectory vector, and the trajectory direction offset is calculated based on the angle between the change direction of the current low-dimensional behavior representation vector and the tangential direction of the normal behavior trajectory. The trajectory evolution velocity offset is calculated based on the change amplitude between the low-dimensional behavior representation vectors of adjacent time steps, and the trajectory distance offset, trajectory direction offset, and trajectory evolution velocity offset are combined to form a behavior offset feature vector.

[0011] Further, S5 includes the following steps: The anomaly detection feature vector is compared with the feature vectors of each anomaly pattern in the pre-built anomaly pattern feature library to obtain the feature distance between the anomaly detection feature vector and each anomaly pattern. Based on the feature distance, perform minimum distance matching processing, and determine the anomaly type corresponding to the anomaly pattern with the smallest distance as the anomaly type result of the energy storage system.

[0012] Furthermore, the abnormal pattern feature library is constructed through the following steps: Obtain operational data samples with anomaly types marked in the historical operational data of the energy storage system, and construct an embedding vector of abnormal operational behavior; The abnormal operating behavior embedding vector is input into the trained autoencoder neural network encoder, and feature encoding processing is performed on the abnormal operating behavior embedding vector to obtain the corresponding abnormal low-dimensional behavior representation vector. The abnormal low-dimensional behavior representation vectors are classified according to different abnormality types, and cluster centers are calculated for the abnormal low-dimensional behavior representation vectors corresponding to the same abnormality type to obtain the abnormal pattern feature vectors corresponding to each abnormality type. The anomaly pattern feature vectors corresponding to each anomaly type are associated and stored with the corresponding anomaly type labels to construct an anomaly pattern feature library.

[0013] Furthermore, the abnormality types include abnormal battery temperature, abnormal cell consistency, abnormal charge / discharge rate, abnormal power fluctuation, and abnormal sensor acquisition.

[0014] A deep learning-based energy storage anomaly monitoring system, applied to any of the aforementioned deep learning-based energy storage anomaly monitoring methods, includes: The data acquisition module is used to acquire voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate and BMS control command data during the historical stable operation of the energy storage system, and obtain a stable operation dataset. The behavior sequence construction module is used to perform joint feature modeling on the stable running dataset, obtain running behavior embedding vectors, and arrange them in chronological order to obtain a running behavior sequence. The model building module is used to input the embedded vector of the running behavior into the autoencoder neural network model based on the running behavior sequence to perform feature encoding processing, obtain a low-dimensional behavior representation vector, and construct the normal behavior trajectory vector of the energy storage system according to the change relationship of the low-dimensional behavior representation vector in the time series. The offset feature extraction module is used to obtain the real-time operation dataset during the real-time operation of the energy storage system, input the real-time operation dataset into the autoencoder neural network model to generate the current behavior embedding vector, and calculate the behavior offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector. The anomaly monitoring module is used to perform anomaly determination processing based on the continuous offset of the behavior offset feature vector within a continuous time window, and obtain anomaly monitoring results.

[0015] The beneficial effects of this invention are as follows: This invention acquires voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate, and BMS control command data from the historical stable operation of an energy storage system to construct a stable operation dataset. Joint feature modeling is then performed on this dataset to transform multi-dimensional operating parameters into operating behavior embedding vectors, forming an operating behavior sequence in chronological order. Furthermore, an autoencoder neural network model is used to perform feature encoding on the operating behavior sequence, learning the behavioral change patterns of the energy storage system under stable operating conditions. Based on the changing relationship of the low-dimensional behavioral representation vectors in the time series, a normal behavior trajectory vector of the energy storage system is constructed. Subsequently, real-time operating data of the energy storage system is input into the autoencoder neural network model to generate a current behavior embedding vector. By calculating the behavioral offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector, the degree of deviation of the current operating state of the energy storage system from the normal operating mode is characterized. Then, based on the continuous offset of the behavioral offset feature vector within a continuous time window, anomaly detection processing is performed. This enables the identification of the gradual deviation of the energy storage system's operating state from the normal operating trajectory, achieving early identification of abnormal states before operating parameters exceed traditional threshold ranges, thus improving the accuracy and timeliness of energy storage system anomaly monitoring. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of a deep learning-based energy storage anomaly monitoring method in this invention.

[0017] Figure 2 This is a flowchart of step S3 in this invention. Detailed Implementation

[0018] Please see Figures 1-2 As shown, this invention relates to a deep learning-based method for monitoring energy storage anomalies, comprising the following steps: S1. Obtain voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate and BMS control command data during the historical stable operation of the energy storage system to obtain a stable operation dataset; S2. Perform joint feature modeling on the stable running dataset to obtain running behavior embedding vectors and arrange them in chronological order to obtain a running behavior sequence; S3. Based on the operating behavior sequence, the operating behavior embedding vector is input into an autoencoder neural network model to perform feature encoding processing to obtain a low-dimensional behavior representation vector, and the normal behavior trajectory vector of the energy storage system is constructed according to the change relationship of the low-dimensional behavior representation vector in the time series. S4. Obtain the real-time operation dataset during the real-time operation of the energy storage system, input the real-time operation dataset into the autoencoder neural network model to generate the current behavior embedding vector, and calculate the behavior offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector. S5. Perform anomaly detection processing based on the continuous offset of the behavior offset feature vector within a continuous time window to obtain anomaly monitoring results.

[0019] In some embodiments, this solution is applied to anomaly monitoring scenarios of photovoltaic power plant-supporting energy storage systems. This energy storage system consists of several battery clusters, battery modules, individual battery cells, and a battery management system (BMS). During operation, it continuously generates voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate, and BMS control commands. Unlike existing technologies that only set thresholds for single parameters and trigger alarms when parameters exceed limits, this embodiment does not directly use the raw monitoring values ​​at each moment as the basis for anomaly judgment. Instead, it first extracts operational behavior embedding vectors that characterize the overall behavior of the energy storage system from historical stable operation data. Then, it further learns the continuous behavioral evolution law of the energy storage system under normal conditions to construct a normal behavior trajectory. The deviation of the real-time operational behavior from the normal behavior trajectory is used as the basis for anomaly identification. Specifically, in step S1, the voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate, and BMS control command data from the historical stable operation process of the energy storage system are first acquired to obtain a stable operation dataset. The "historical stable operation process" mentioned here is not arbitrary historical data, but rather data from fault-free operating periods filtered through maintenance records, BMS operation logs, and troubleshooting records. For example, in a photovoltaic energy storage power station, a period of three consecutive months without thermal alarms, overcharge / over-discharge alarms, cell imbalance alarms, and with normal manual inspection records can be selected as the stable operation sample interval. Further, the battery management system can collect data on the total voltage, cluster current, cluster power, and temperature of each module, as well as estimated cell-level SOC and SOH values. Simultaneously, the charge / discharge rate and BMS control commands at the corresponding times, such as equalization activation commands, current limiting commands, and charge / discharge switching commands, can be read. After collection, a unified timestamp is added to all types of data, and the data is synchronously rearranged according to a unified sampling period to eliminate the time misalignment between different sampling sources. Then, the rearranged multidimensional operating data sequence undergoes missing value removal, noise filtering, and normalization to obtain the stable operation dataset. Unlike existing technologies that only monitor single indicators such as voltage and temperature, this step fundamentally improves the integrity and consistency of anomaly identification by establishing a multi-dimensional stable operation dataset under a unified time benchmark. Joint feature modeling of the stable operation dataset aims to transform the original multi-dimensional parameters into operational behavior embedding vectors that characterize the overall operating state of the energy storage system.In practice, instead of directly feeding the raw parameters from a single sampling moment into the classifier as in traditional methods, the implementation first divides the voltage, current, power, and battery temperature distributions into time segments based on a preset time window. For example, an analysis window of 30 or 60 seconds is used, and the continuous sampling points within the window are organized into operational behavior segments. Then, the mean, variance, rate of change, and fluctuation amplitude of the voltage, current, power, and battery temperature distributions within the window are calculated respectively. Furthermore, the temperature distribution gradient characteristics are calculated based on the temperature difference between battery modules to characterize local heat accumulation trends and thermal imbalances between modules. Simultaneously, a state vector is constructed by combining SOC, SOH, charge / discharge rate, and BMS control commands. The state vector is then concatenated with the aforementioned operational feature vectors by adjusting the feature dimensions to obtain the operational behavior embedding vector. The chronologically ordered operational behavior embedding vectors are divided into several training sample sequences. The operational behavior embedding vectors for the first few time steps in each training sample sequence are input into the encoder of an autoencoder neural network model to perform temporal feature compression calculations, obtaining historical behavior representation vectors. Then, these historical behavior representation vectors are input into the decoder to perform prediction calculations on future behavior, obtaining predicted behavior vectors for subsequent time steps. The prediction error is calculated based on the difference between the predicted behavior vectors and the actual operational behavior embedding vectors for subsequent time steps. Backpropagation is then performed based on the loss function to update the network parameters until the model converges. It is important to note that the autoencoder neural network model in this scheme is not a traditional static autoencoder that only reconstructs input samples, but rather a temporal autoencoder prediction structure trained to "predict subsequent normal behavior changes from a preceding stable behavior sequence." In other words, this model learns how the energy storage system's behavior evolves from the current moment to future moments under stable operating conditions, rather than simply learning a compressed representation of data at a single moment. After training, trajectory connection processing is performed based on the model's understanding of the changes in the predicted behavior vectors for subsequent time steps under stable operating conditions in the time series, obtaining the normal behavior trajectory vector of the energy storage system. A real-time operational dataset of the energy storage system is acquired, and a current behavior embedding vector is constructed using the same joint feature modeling method as historical stable operational data. Subsequently, the real-time operational data is input into a trained autoencoder neural network model to generate the current behavior embedding vector, and the behavior offset feature vector relative to the normal behavior trajectory vector is calculated. Preferably, the behavior offset feature vector includes trajectory distance offset, trajectory direction offset, and trajectory evolution speed offset.The trajectory distance offset is used to characterize the degree of deviation of the current behavior from the normal behavior trajectory, reflecting the "amplitude" characteristic of the anomaly; the trajectory direction offset is characterized by the angle between the direction of change of the current low-dimensional behavior representation vector and the tangential direction of the normal behavior trajectory, reflecting the "evolution direction" characteristic of the anomaly; the trajectory evolution speed offset is calculated by the change amplitude between the low-dimensional behavior representation vectors of adjacent time steps, used to characterize the "development rate" characteristic of the anomaly. Unlike existing technologies that only judge whether the parameter value is out of bounds, this embodiment actually characterizes the current operating state of the energy storage system from three dimensions: how much the deviation is, where the deviation is, and whether the deviation is accelerating. That is to say, even if a certain anomaly is not significant in absolute value, if its direction of change has significantly deviated from the normal behavior trajectory, or its deviation speed continues to increase, it can still be identified as a noteworthy sign of anomaly evolution. Anomaly judgment processing is performed based on the continuous deviation of the behavior deviation feature vector in a continuous time window to obtain the anomaly monitoring results. Here, a single-moment trigger judgment is not used, but the persistence, accumulation, and consistency of the deviation feature in multiple consecutive time windows are examined. For example, if within five consecutive windows, the trajectory distance offset continuously increases, the trajectory direction offset remains in the same abnormal pattern offset direction, and the trajectory evolution speed offset is higher than the statistical upper bound of the stable operation phase, then the energy storage system is determined to have entered an abnormal evolution state. Furthermore, the anomaly judgment feature vector can be matched with the feature vectors of each abnormal pattern in a pre-built anomaly pattern feature library, and the anomaly type corresponding to the minimum distance is determined as the final anomaly type result. Thus, this scheme can not only provide a "whether it is abnormal" judgment, but also further output results such as abnormal battery temperature, abnormal cell consistency, abnormal charge / discharge rate, abnormal power fluctuation, or abnormal sensor acquisition. Compared with existing technologies, it no longer uses a fixed threshold as the sole criterion, but instead uses the continuous offset process as the basis for anomaly triggering, enabling the system to identify slow degradation and gradual anomalies; it no longer directly maps the current original parameters to the anomaly category, but first constructs a normal behavior trajectory, and then matches the offset features with the anomaly pattern library, making the anomaly classification based on a chain-like reasoning that is more in line with engineering mechanisms—from normal baseline to offset pattern to anomaly type. It can identify abnormal trends in advance when the parameters have not exceeded the traditional threshold, and provide more stable and reliable results on the anomaly types.

[0020] Further, S1 includes the following steps: The battery management system (BMS) and environmental monitoring devices of the energy storage system collect data on voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate, and BMS control commands during the stable operation of the energy storage system, and add a unified timestamp to obtain the original operating dataset. Time alignment processing is performed on the original running dataset, and the data is synchronously rearranged according to the timestamp identifier at a uniform sampling time interval to obtain a multidimensional running data sequence arranged in chronological order. Data cleaning and standardization are performed on the multidimensional running data sequence to remove missing data and outlier sampling points, and normalization mapping is performed on the running parameters of each dimension to obtain a stable running dataset.

[0021] Specifically, the energy storage system first collects multi-dimensional operational data during its operation through its Battery Management System (BMS) and environmental monitoring devices. The BMS acquires real-time operational parameters at the battery module and cell levels, including voltage, current, power, SOC, SOH, and charge / discharge rates. It also records the status of BMS control commands corresponding to the system's control logic, such as equalization control commands, current limiting control commands, and charge / discharge switching commands. The environmental monitoring devices collect temperature distribution data within the battery compartment. This is achieved by deploying an array of temperature sensors around the battery modules to obtain temperature information for the battery module and cell areas, thus generating temperature monitoring data reflecting the thermal distribution. During data acquisition, to ensure time consistency between different data sources, a unified timestamp is added to each sampled data point. This allows data from the BMS and environmental monitoring devices to be aligned under a unified time reference, resulting in a raw operational dataset containing multi-dimensional operational parameters. After obtaining the raw operational dataset, time alignment processing is performed on the dataset. Because the sampling periods of different types of sensors may differ—for example, current and voltage data are typically sampled at higher frequencies, while temperature sensors or environmental monitoring devices have relatively longer sampling periods—it is necessary to synchronously rearrange various types of data based on a unified timestamp. Specifically, the original operating data can be resampled according to a preset unified sampling time interval, mapping each dimension of operating parameters to a unified time series framework. For example, when the unified sampling period is set to 1 second, data with different sampling frequencies will be interpolated or rearranged based on the timestamp, ensuring that parameters such as voltage, current, power, temperature distribution, and SOC have corresponding data values ​​at the same time series position, thus forming a multidimensional operating data sequence arranged in chronological order. This step eliminates inconsistencies in sampling time among multiple data sources, enabling joint analysis of various operating parameters at a unified time scale. Subsequently, data cleaning and standardization are performed on the multidimensional operating data sequence to improve data quality and eliminate the impact of abnormal sampling on model training. During data cleaning, missing data in the data sequence is first identified and filled using time series interpolation methods, such as linear interpolation or neighborhood mean interpolation. Simultaneously, anomalous sampling points that clearly do not conform to physical laws are identified, such as sudden changes in voltage or current within a very short time that cannot be explained by the system's operating status. In these cases, a sliding window statistical method can be used for anomaly detection and the corresponding data points can be removed. After data cleaning, normalization mapping is performed on the operating parameters of each dimension. Specifically, this can be achieved by linearly normalizing each parameter based on its historical value range during stable operation, mapping data of different dimensions to the same numerical range.For example, parameters such as voltage, current, power, and temperature can be scaled according to their maximum and minimum value ranges, mapping them to a numerical range between zero and one. This avoids scale differences between parameters with different dimensions during subsequent feature modeling. Through the above processing, a stable running dataset with a uniform time structure, reliable quality, and consistent numerical scale can be obtained.

[0022] Further, S2 includes the following steps: Based on the stable operation dataset, the voltage, current, power and battery temperature distribution are divided into time segments according to a preset time window, and the operation behavior segment data is constructed by combining SOC, SOH, charge and discharge rate and BMS control command status. The mean, variance, rate of change, and fluctuation amplitude of voltage, current, power, and battery temperature distribution in the operation behavior segment data are calculated respectively, and the temperature distribution gradient characteristics are calculated based on the temperature difference between battery modules to obtain the operation feature vector; The SOC, SOH, charge / discharge rate and BMS control command data are constructed into a state vector according to a time window, and the state vector and the running feature vector are concatenated according to the feature dimension to obtain the running behavior embedding vector. The embedded vector of the running behavior is processed by performing a sequence arrangement in chronological order to obtain the running behavior sequence.

[0023] In some embodiments, a stable operating dataset is first used as input to divide the voltage, current, power, and battery temperature distribution into time segments. The preset time window is not simply an accumulation of sampling points, but rather a behavioral observation interval comprehensively set based on the operating inertia of the energy storage system, the BMS control response cycle, and the power fluctuation rhythm in a photovoltaic scenario. For example, in a photovoltaic energy storage system with a sampling period of 1 second, 30 seconds can be set as a basic analysis window, ensuring that each window covers a local charge / discharge adjustment action without being too long to mask short-term dynamic characteristics. Based on this, the voltage, current, power, and battery temperature distribution data within each time window are segmented and encapsulated. Simultaneously, the corresponding SOC, SOH, charge / discharge rate, and BMS control command status are read, thus forming the operational behavior segment data. The key here is not simply "segmenting," but rather aligning electrical quantities, heat, and control status to the same behavioral unit through a unified window, so that the subsequently extracted features no longer correspond to discrete parameter points, but rather to a complete operational behavior process. After the segment construction is completed, statistical and dynamic feature extraction processing is performed on the voltage, current, power, and battery temperature distribution in the operational behavior segment data. Taking voltage data as an example, the average voltage is calculated within each time window to reflect the overall voltage level of the cell or module within that window; the voltage variance is calculated to reflect the degree of voltage fluctuation dispersion within the window; the voltage change rate is calculated to characterize the average slope of voltage change within that window; and the voltage fluctuation amplitude is calculated to characterize the amplitude range between the maximum and minimum values. The same processing logic is applied to the current and power data to obtain characterization quantities such as current stability, load regulation severity, and power output smoothness within the corresponding window. For battery temperature distribution data, in addition to calculating the mean, variance, change rate, and fluctuation amplitude, the temperature distribution gradient feature is further calculated. This gradient feature is preferably obtained through the temperature difference matrix between each battery module. For example, for adjacent modules within the same cluster, their average temperature difference is calculated, and a temperature difference sequence is formed in a preset topological order. The mean, maximum gradient value, or gradient dispersion of this sequence is then calculated to reflect whether the heat diffusion between modules is balanced. If a module experiences localized heat accumulation due to increased internal resistance, the temperature distribution gradient characteristics will change significantly before traditional threshold alarms, even if its absolute temperature has not yet exceeded the limit. Furthermore, the SOC, SOH, charge / discharge rate, and BMS control command data are processed into state vectors within the same time window. These state vectors are not simply mechanically concatenated from the original state parameters; rather, they are organized as contextual constraints for operational behavior. Specifically, SOC represents the remaining battery capacity range corresponding to the current window, SOH represents the battery health degradation background, charge / discharge rate represents the current load intensity, and BMS control commands represent the response status of the control system to balancing, current limiting, switching, or protection logic within this window.After obtaining the operational feature vector and state vector, they are concatenated according to their feature dimensions to form an operational behavior embedding vector. This concatenation is not a simple data juxtaposition, but a unified semantic expression mechanism for behavior. Essentially, it merges the response results reflected by voltage, current, power, and temperature distributions within a window with the operating conditions reflected by SOC, SOH, rate, and control commands into the same vector structure. This ensures that each operational behavior embedding vector corresponds to the overall behavioral state of the energy storage system under specific time windows, specific operating conditions, and specific control backgrounds. In other words, the operational behavior embedding vector no longer represents a single parameter or a single device state, but rather a behavioral unit defined by multi-dimensional features. This behavioral unit can be used to characterize the dynamic response of the energy storage system under rapid fluctuations in photovoltaic power, and also to characterize different stable operating modes such as smooth nighttime discharge, low-rate compensation, or local thermal imbalance.

[0024] Further, S3 includes the following steps: The running behavior sequence is divided into several training sample sequences according to time order, and the running behavior embedding vector of the first few time steps in each training sample sequence is input into the encoder of the autoencoder neural network model. Temporal feature compression calculation is performed on the running behavior embedding vector of the first few time steps to obtain the corresponding historical behavior representation vector. The historical behavior representation vector is input into the decoder of the autoencoder neural network model, and future behavior prediction calculation is performed on the historical behavior representation vector to obtain the predicted behavior vector for the corresponding subsequent time step. The prediction error is calculated based on the difference between the predicted behavior vector and the actual subsequent time step running behavior embedding vector. The model training loss is calculated using a loss function based on the prediction error, and backpropagation is performed on the autoencoder neural network model to update the network parameters based on the model training loss until the model converges. Based on the trained autoencoder neural network model, trajectory connection processing is performed on the relationship between the predicted behavior vector of subsequent time steps under stable operation and the change in the time series to obtain the normal behavior trajectory vector of the energy storage system.

[0025] In some embodiments, firstly, the time-ordered sequence of operating behaviors is divided into several training sample sequences according to a fixed-length sliding window. Each training sample sequence consists of an embedding vector of operating behaviors for several consecutive time steps, where the first few time steps serve as the model input sequence and the subsequent time steps serve as the prediction target sequence. For example, in an energy storage system with a sampling period of 1 second and a behavior analysis window of 30 seconds, the operating behavior sequence can be divided into an input segment of 20 time steps and a prediction segment of 10 time steps. The resulting training sample sequence retains the temporal continuity of the energy storage system's operating behavior and reflects the dynamic relationship of various operating parameters changing over time under stable operating conditions. Subsequently, the embedding vector of operating behaviors for the first few time steps in each training sample sequence is input into the encoder of an autoencoder neural network model to perform temporal feature compression calculation on the embedding vector. The encoder preferably adopts a combination of a multi-layer fully connected network and a time-series mapping structure, compressing the high-dimensional operating behavior embedding vector into a low-dimensional historical behavior representation vector through layer-by-layer nonlinear mapping. This historical behavior representation vector can reduce redundant dimensions while maintaining the main dynamic feature information, thereby extracting the core features of the energy storage system's behavior changes during the stable operation phase. For example, in the stable charging phase where photovoltaic power is slowly increasing, there is usually a stable coupling relationship between the rate of change of current, average power, temperature gradient, and SOC growth rate. The encoder, by learning from a large number of stable operating samples, can compress the correlation patterns between these multidimensional parameters into a low-dimensional but information-dense behavioral representation vector. After obtaining the historical behavioral representation vector, it is input into the decoder of the autoencoder neural network model to perform future behavior prediction calculations. The decoder expands the historical behavioral representation vector into a sequence of predicted behavioral vectors for subsequent time steps through layer-by-layer reverse mapping. Specifically, based on the behavioral evolution features extracted by the encoder, the decoder predicts the changing trend of the operating behavior embedding vector within several future time steps. For example, when the energy storage system is in the stable charging phase, if the current behavioral representation vector indicates that the system is in an operating state with gradually increasing SOC, small power fluctuations, and a stable temperature gradient, then the future behavioral vector predicted by the decoder will show an evolutionary trend of continued SOC increase, stable power, and slow temperature change. Subsequently, the predicted behavioral vector is compared dimension-by-dimensionally with the actual operating behavior embedding vectors corresponding to subsequent time steps in the training sample sequence, and the prediction error is obtained by calculating the difference between the two. Prediction error can be calculated using the mean squared error method, thereby quantifying the degree of deviation between the model's predicted behavior and the actual behavior. During model training, the model training loss is calculated based on the prediction error using a preset loss function, and the parameters of the autoencoder neural network model are updated using the backpropagation algorithm. Specifically, the gradient of the loss function with respect to each network weight parameter is calculated, and the weight matrix and bias parameters in the encoder and decoder are iteratively adjusted using the gradient descent algorithm.As training samples are continuously input and network parameters are constantly updated, the model gradually converges to a state where it can stably predict the evolution trend of normal behavior. At this point, the encoder can extract behavioral features representing normal operating patterns from historical operating behavior sequences, while the decoder can infer future behavior changes based on these features. After model training is complete, the operating behavior sequences in the stable operating dataset are processed for overall prediction. Specifically, the trained autoencoder neural network model is used to predict the operating behavior embedding vectors at each time step in the stable operating phase, and the predicted behavior vectors for subsequent time steps are connected in chronological order to form a continuous behavior evolution path. This behavior evolution path is the normal behavior trajectory vector of the energy storage system under stable operating conditions. This trajectory vector essentially reflects the typical evolution law of the operating behavior embedding vector changing over time under stable operating conditions. For example, in the stable phase of the photovoltaic energy storage system absorbing photovoltaic power during the day, the normal behavior trajectory vector typically exhibits characteristics such as a gradual increase in SOC, a stable rate of change in current, a slow change in temperature distribution gradient, and small power fluctuations. When subsequent real-time operational data is mapped to the current behavior embedding vector, the degree of deviation from the normal behavior trajectory vector can be calculated to identify whether an abnormal operational trend has occurred. Through the above algorithm process, this step not only completes the learning of stable operational behavior patterns, but also transforms complex multidimensional operational data into normal behavior trajectories with clear evolutionary structures, providing a reference benchmark for abnormal deviation detection.

[0026] Furthermore, the loss function is as follows: ; in, Use the loss value to train the model; t is the number of training samples; t is the index of the current time step in the training sample sequence; k is the prediction step size. The embedding vector of the actual running behavior of the i-th training sample at time t+k; The autoencoder neural network model predicts the behavior vector at time step (t+k) based on the embedded vector of the running behavior before time step t. These are the weighting coefficients for trajectory smoothing constraints.

[0027] Specifically, during the model training phase, the training sample sequence obtained from the operational behavior sequence is input into the autoencoder neural network model. Each training sample consists of operational behavior embedding vectors from several consecutive time steps. The encoder first performs feature compression calculation on the operational behavior embedding vectors from the previous few time steps to obtain historical behavior representation vectors. Subsequently, the decoder predicts the operational behavior for future time steps based on these historical behavior representation vectors, generating corresponding predicted behavior vectors. By comparing the predicted behavior vectors with the actual operational behavior embedding vectors dimension by dimension, the difference between the two is calculated, thus obtaining the model prediction error. The first part of the loss function measures the mean square error between the predicted behavior vector and the actual behavior vector. By accumulating and averaging the errors at all prediction times in the training sample set, it reflects the model's overall predictive ability for the behavioral changes of the energy storage system. The closer the predicted behavior is to the actual behavior, the smaller this loss value, indicating that the model has learned the behavioral evolution laws of the stable operation phase more fully. To ensure the continuity and physical rationality of the predicted behavior in the time series, a trajectory smoothing constraint term is further introduced into the loss function. This constraint term limits the magnitude of change in predicted behavior over time by calculating the degree of difference in predicted behavior vectors between adjacent prediction time steps. In the actual operation of an energy storage system, operating parameters such as voltage, current, power, and temperature typically exhibit continuous changes without significant abrupt changes between adjacent time steps. If the model is trained solely on the prediction error term, it may produce discontinuous prediction results at local time points, leading to unreasonable fluctuations in the subsequently constructed normal behavior trajectory. By introducing a trajectory smoothing constraint term into the loss function, the model can minimize prediction error while maintaining smooth changes in predicted behavior over time, making the predicted behavior sequence more consistent with the dynamic changes in the actual operation of the energy storage system. The weighting coefficient is used to adjust the degree of influence between the prediction error term and the trajectory smoothing constraint term. When the weighting coefficient is set appropriately, the model can maintain high prediction accuracy while ensuring good continuity of the predicted behavior trajectory. Through the above loss function design, the autoencoder neural network model can effectively learn the coordinated change patterns of multi-dimensional operating parameters of the energy storage system under stable operating conditions.

[0028] Further, calculating the behavior offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector includes the following steps: The trajectory distance offset is calculated based on the minimum distance between the current low-dimensional behavior representation vector and the normal behavior trajectory vector, and the trajectory direction offset is calculated based on the angle between the change direction of the current low-dimensional behavior representation vector and the tangential direction of the normal behavior trajectory. The trajectory evolution velocity offset is calculated based on the change amplitude between the low-dimensional behavior representation vectors of adjacent time steps, and the trajectory distance offset, trajectory direction offset, and trajectory evolution velocity offset are combined to form a behavior offset feature vector.

[0029] In some embodiments, firstly, real-time operating data is input into the encoder of the trained autoencoder neural network model to obtain the low-dimensional behavior representation vector corresponding to the current time step, and this vector is compared with the normal behavior trajectory vector constructed from the historical stable operating phases. The normal behavior trajectory vector is formed by connecting the low-dimensional behavior representation vectors of each time step in the stable operating phase in chronological order, and it represents a continuously changing behavior trajectory in the feature space. To measure the degree of deviation of the current operating state from this trajectory, the Euclidean distance between the current low-dimensional behavior representation vector and each trajectory point of the normal behavior trajectory is first calculated, and the minimum value is selected as the trajectory distance offset. This distance reflects the deviation of the current operating state in the overall behavior characteristics. When the energy storage system is in a stable operating state, this distance is usually kept within a small range; if the system operating mode gradually deviates from the normal trajectory, for example, if abnormal cell temperature distribution or abnormal power fluctuation occurs, the distance will gradually increase. Further, after determining the trajectory position corresponding to the minimum distance, the trajectory direction offset is obtained by calculating the angle between the changing direction of the current low-dimensional behavior representation vector and the tangential direction of the normal behavior trajectory at that position. In practical implementation, the current behavioral evolution direction can be represented by calculating the difference vector between the low-dimensional behavioral representation vectors of the current time step and the previous time step. Simultaneously, the local tangential direction of the normal behavioral trajectory can be approximated by the difference vector between two adjacent trajectory points. The directional offset is then calculated using the vector angle. When the energy storage system's operating trend aligns with the normal behavioral evolution trend, this angle is typically small. If the system's operating trend gradually deviates from the normal evolution direction, such as abnormal power fluctuations or rapid temperature increases during the SOC stabilization phase, the angle will significantly increase, thus reflecting directional anomalies in the operating trend. To further characterize the changing features of the behavioral evolution rate, the trajectory evolution velocity offset is also calculated using the change amplitude between the low-dimensional behavioral representation vectors of adjacent time steps. Specifically, the Euclidean distance between the low-dimensional behavioral representation vectors of the current and previous time steps is first calculated as the current behavioral evolution velocity. Then, this evolution velocity is compared with the average evolution velocity at the corresponding position in the normal behavioral trajectory to obtain the trajectory evolution velocity offset. During the stable operation phase of an energy storage system, the rate of change of various operating parameters typically remains within a relatively stable range. For example, during the smooth charging process of a photovoltaic energy storage system, the SOC growth rate, current change rate, and temperature change rate usually exhibit smooth changes. When abnormal trends occur within the system, such as increased cell internal resistance leading to a faster temperature rise or abnormal power regulation causing increased current fluctuations, the current behavioral evolution rate will significantly deviate from the normal trajectory change rate, thus increasing the trajectory evolution rate offset. After obtaining the trajectory distance offset, trajectory direction offset, and trajectory evolution rate offset, these three are combined in a preset order to form a behavioral offset feature vector.This behavior offset feature vector comprehensively characterizes the degree of difference between the current operating state and the normal operating trajectory of the energy storage system from three dimensions: spatial distance, evolution direction, and change rate.

[0030] Further, S5 includes the following steps: The anomaly detection feature vector is compared with the feature vectors of each anomaly pattern in the pre-built anomaly pattern feature library to obtain the feature distance between the anomaly detection feature vector and each anomaly pattern. Based on the feature distance, perform minimum distance matching processing, and determine the anomaly type corresponding to the anomaly pattern with the smallest distance as the anomaly type result of the energy storage system.

[0031] In some embodiments, a pre-built anomaly pattern feature library is first invoked. This library is derived from historical operational samples labeled with anomaly types. It is obtained by constructing anomaly behavior embedding vectors for various anomaly samples, extracting low-dimensional behavior representation vectors, and further performing category classification and cluster center calculation. In other words, the anomaly pattern feature library does not store raw voltage, current, or temperature parameters, but rather anomaly pattern feature vectors that represent typical offset patterns at the behavioral level for a certain type of anomaly. For example, for battery temperature anomalies, the anomaly pattern feature vector typically shows a continuously increasing trajectory distance offset, a trajectory direction offset towards the thermal imbalance-related offset direction, and a rapid increase in trajectory evolution speed offset within a short period. For cell consistency anomalies, the anomaly pattern feature vector is more likely to show a slower but longer-lasting offset increase, and a bias towards trajectory separation caused by the accumulation of voltage and SOC differences in the behavioral direction. It is precisely because the differences in the offset feature combinations of different anomaly types can be preserved by the anomaly pattern feature library that subsequent distance matching processing has anomaly classification significance. Subsequently, distance calculations are performed on the current anomaly determination feature vector and each anomaly pattern feature vector in the anomaly pattern feature library to obtain the feature distance between the current anomaly state and each type of anomaly pattern. Preferably, Euclidean distance, Mahalanobis distance, or weighted Euclidean distance can be used for calculation. The closer the anomaly determination feature vector is to a certain anomaly pattern feature vector, the more the current energy storage system's offset feature combination matches the typical behavior pattern of this type of anomaly in historical samples. Taking Euclidean distance as an example, the anomaly determination feature vector and each anomaly pattern feature vector can be subtracted, squared, and summed along their corresponding dimensions, and then the square root can be taken to quantify their closeness in the feature space. If the differences in the contribution of different offset feature dimensions to anomaly identification are considered, different weights can be assigned to the trajectory distance offset, trajectory direction offset, and trajectory evolution speed offset to highlight feature dimensions that are more sensitive to a certain type of anomaly. For example, when identifying power fluctuation anomalies, the weight of the trajectory evolution speed offset related dimension can be appropriately increased; while when identifying cell consistency anomalies, the weight of the cumulative feature of trajectory distance offset can be increased, thereby enhancing the specificity of the matching results. After calculating the feature distances between each anomaly mode, a minimum distance matching process is performed based on these feature distances. The anomaly type corresponding to the anomaly mode with the smallest distance is determined as the anomaly type result for the energy storage system. The engineering implications of this step are that it selects the type most similar to the current anomaly state from historically known anomaly modes as the most likely anomaly attribution for the current energy storage system.For example, if a photovoltaic energy storage system experiences a sustained increase in the temperature of local modules during continuous operation, but the voltage and current have not exceeded their limits, the aforementioned steps will first convert this state into an anomaly detection feature vector. When matching it with the anomaly pattern feature library, if the distance between this feature vector and the battery temperature anomaly pattern is the smallest, the system will determine the current anomaly type as a battery temperature anomaly, and will not misjudge it as a power fluctuation anomaly or a sensor acquisition anomaly. As another example, when some cells gradually differentiate during the SOC recovery process due to varying degradation, the current anomaly detection feature vector usually has the smallest feature distance to the cell consistency anomaly pattern in the feature library, thus enabling the output of a more mechanistic classification result: cell consistency anomaly.

[0032] Furthermore, the abnormal pattern feature library is constructed through the following steps: Obtain operational data samples with anomaly types marked in the historical operational data of the energy storage system, and construct an embedding vector of abnormal operational behavior; The abnormal operating behavior embedding vector is input into the trained autoencoder neural network encoder, and feature encoding processing is performed on the abnormal operating behavior embedding vector to obtain the corresponding abnormal low-dimensional behavior representation vector. The abnormal low-dimensional behavior representation vectors are classified according to different abnormality types, and cluster centers are calculated for the abnormal low-dimensional behavior representation vectors corresponding to the same abnormality type to obtain the abnormal pattern feature vectors corresponding to each abnormality type. The anomaly pattern feature vectors corresponding to each anomaly type are associated and stored with the corresponding anomaly type labels to construct an anomaly pattern feature library.

[0033] Specifically, historical operational data samples with labeled anomaly types are first extracted from the long-term operation records of the energy storage system. These anomaly samples can originate from maintenance records, fault logs, or system alarm data, such as abnormal battery temperature, abnormal cell consistency, abnormal charge / discharge rate, or abnormal power fluctuations. For each anomaly sample, following the same feature modeling process as for stable operational data, joint feature modeling processing is performed on the voltage, current, power, and temperature distribution within its corresponding time window to construct an abnormal operational behavior embedding vector. This ensures that the abnormal sample maintains consistency in data structure with the operational behavior embedding vector during normal operation, thereby guaranteeing a unified data representation for subsequent feature extraction. Subsequently, the abnormal operational behavior embedding vector is input into the encoder of the trained autoencoder neural network model to perform feature encoding processing. Since the encoder has been trained on stable operational data, its internal parameters can effectively extract key features in the operational behavior of the energy storage system. Therefore, when the abnormal operational behavior embedding vector is input into the encoder, a corresponding low-dimensional abnormal behavior representation vector can be obtained. This low-dimensional representation vector not only retains the main information of the abnormal behavior at the operational feature level but also removes redundant features through the encoding process, making the differences between different anomaly types more apparent in the feature space. For example, in battery temperature anomaly samples, the encoded behavioral representation vectors typically show a significant shift in the temperature gradient-related dimension; while in power fluctuation anomaly samples, they are more likely to form a concentrated distribution in the power change rate and current fluctuation feature dimensions. After obtaining the low-dimensional behavioral representation vectors corresponding to each anomaly sample, a category classification process is performed based on the anomaly type label, grouping the abnormal behavioral representation vectors corresponding to the same anomaly type into the same category set. Subsequently, cluster center calculation is performed within each anomaly category set to extract typical feature vectors that can represent the abnormal behavior pattern of that type. In specific implementation, K-means clustering or density clustering algorithms can be used to aggregate features of anomaly samples of the same type, and the cluster center vector is calculated as the anomaly pattern feature vector of that anomaly type. The cluster center vector represents the typical distribution position of this type of abnormal behavior in the feature space, and it can comprehensively reflect the comprehensive characteristics of this type of anomaly in multiple dimensions such as trajectory distance offset, trajectory direction offset, and trajectory evolution speed offset. For example, for cell consistency anomalies, the cluster center vector typically shows a gradual increase in trajectory distance offset but a small directional offset, while for power fluctuation anomalies, the cluster center vector shows a significant increase in trajectory evolution speed offset. After calculating the cluster centers, the anomaly pattern feature vectors corresponding to each anomaly type are associated and stored with their corresponding anomaly type labels, and then uniformly written into the anomaly pattern feature library. This anomaly pattern feature library can be understood as a dictionary of anomaly behavior patterns, where each record contains an anomaly pattern feature vector and its corresponding anomaly type label.When the system calculates the anomaly judgment feature vector during real-time monitoring, it can perform distance matching with the feature vector in the anomaly pattern feature library to determine the anomaly pattern that is closest to the current anomaly state, and then output the corresponding anomaly type result.

[0034] Furthermore, the abnormality types include abnormal battery temperature, abnormal cell consistency, abnormal charge / discharge rate, abnormal power fluctuation, and abnormal sensor acquisition.

[0035] It should be noted that battery temperature anomalies mainly manifest as a significant increase in the temperature distribution gradient between battery modules or a sustained temperature of certain cells exceeding the average level of cells within the same group. At the behavioral characteristic level, this type of anomaly typically exhibits a continuously increasing rate of change in temperature-related feature dimensions, accompanied by a gradually increasing trajectory distance offset. For example, when a battery module experiences localized heat accumulation due to decreased heat dissipation or increased internal resistance, its temperature change trend will significantly deviate from the normal behavioral trajectory during stable operation, thus appearing as a continuously increasing temperature gradient in the behavioral offset feature vector. Cell consistency anomalies mainly refer to the gradually widening differences in voltage, SOC, or capacity change trends among different cells within the same battery module. Under stable operating conditions, the voltage and SOC changes of each cell typically maintain high consistency. However, when some cells experience capacity decay or internal resistance changes, their voltage response during charging and discharging will gradually deviate from other cells. At the behavioral characteristic level, this type of anomaly typically manifests as a gradually increasing difference in voltage fluctuation variance and SOC change, causing the operating behavior embedding vector to gradually deviate from the normal behavioral trajectory in the feature space. Charge / discharge rate anomalies typically occur when the power scheduling strategy or control system of an energy storage system malfunctions. When the charge / discharge rate exceeds the normal operating range or fluctuates frequently within a short period, the current and power change rate characteristics will show significant shifts. At the behavioral characteristic level, this type of anomaly often manifests as a simultaneous increase in trajectory direction and trajectory evolution speed offsets, indicating a significant deviation from the normal operating trajectory. Power fluctuation anomalies mainly refer to unexpected large fluctuations in the output power of the energy storage system within a short period. These anomalies are usually related to grid load fluctuations, abnormal control strategies, or changes in equipment status. At the behavioral characteristic level, this type of anomaly is mainly manifested as a significant increase in the power change rate and current fluctuation amplitude, causing the operating behavior embedding vector to exhibit rapid oscillation characteristics in the time series, thus leading to a significant increase in the trajectory evolution speed offset. Sensor acquisition anomalies are mainly caused by sensor failures, data acquisition errors, or communication anomalies, and typically manifest as a sudden change, loss, or unreasonable value of a certain operating parameter. For example, when a temperature sensor experiences an acquisition anomaly, the temperature data may experience instantaneous jumps or remain at a fixed value for a long period, resulting in abnormal fluctuations in the operating characteristic vector within a local time window. This type of anomaly typically manifests as a sudden change in trajectory distance offset at a single time step, but its evolution speed characteristics differ from those of real device anomalies. Therefore, it can be identified as a sensor acquisition anomaly during the anomaly pattern feature library matching process.

[0036] The present invention also includes a deep learning-based energy storage anomaly monitoring system, comprising: The data acquisition module is used to acquire voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate and BMS control command data during the historical stable operation of the energy storage system, and obtain a stable operation dataset. The behavior sequence construction module is used to perform joint feature modeling on the stable running dataset, obtain running behavior embedding vectors, and arrange them in chronological order to obtain a running behavior sequence. The model building module is used to input the embedded vector of the running behavior into the autoencoder neural network model based on the running behavior sequence to perform feature encoding processing, obtain a low-dimensional behavior representation vector, and construct the normal behavior trajectory vector of the energy storage system according to the change relationship of the low-dimensional behavior representation vector in the time series. The offset feature extraction module is used to obtain the real-time operation dataset during the real-time operation of the energy storage system, input the real-time operation dataset into the autoencoder neural network model to generate the current behavior embedding vector, and calculate the behavior offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector. The anomaly monitoring module is used to perform anomaly determination processing based on the continuous offset of the behavior offset feature vector within a continuous time window, and obtain anomaly monitoring results.

[0037] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A deep learning-based method for monitoring energy storage anomalies, characterized in that, Includes the following steps: S1. Obtain voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate and BMS control command data during the historical stable operation of the energy storage system to obtain a stable operation dataset; S2. Perform joint feature modeling on the stable running dataset to obtain running behavior embedding vectors and arrange them in chronological order to obtain a running behavior sequence; S3. Based on the operating behavior sequence, the operating behavior embedding vector is input into an autoencoder neural network model to perform feature encoding processing to obtain a low-dimensional behavior representation vector, and the normal behavior trajectory vector of the energy storage system is constructed according to the change relationship of the low-dimensional behavior representation vector in the time series. S4. Obtain the real-time operation dataset during the real-time operation of the energy storage system, input the real-time operation dataset into the autoencoder neural network model to generate the current behavior embedding vector, and calculate the behavior offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector. S5. Perform anomaly detection processing based on the continuous offset of the behavior offset feature vector within a continuous time window to obtain anomaly monitoring results.

2. The energy storage anomaly monitoring method based on deep learning according to claim 1, characterized in that, S1 includes the following steps: The battery management system (BMS) and environmental monitoring devices of the energy storage system collect data on voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate, and BMS control commands during the stable operation of the energy storage system, and add a unified timestamp to obtain the original operating dataset. Time alignment processing is performed on the original running dataset, and the data is synchronously rearranged according to the timestamp identifier at a uniform sampling time interval to obtain a multidimensional running data sequence arranged in chronological order. Data cleaning and standardization are performed on the multidimensional running data sequence to remove missing data and outlier sampling points, and normalization mapping is performed on the running parameters of each dimension to obtain a stable running dataset.

3. The energy storage anomaly monitoring method based on deep learning according to claim 2, characterized in that, S2 includes the following steps: Based on the stable operation dataset, the voltage, current, power and battery temperature distribution are divided into time segments according to a preset time window, and the operation behavior segment data is constructed by combining SOC, SOH, charge and discharge rate and BMS control command status. The mean, variance, rate of change, and fluctuation amplitude of voltage, current, power, and battery temperature distribution in the operation behavior segment data are calculated respectively, and the temperature distribution gradient characteristics are calculated based on the temperature difference between battery modules to obtain the operation feature vector; The SOC, SOH, charge / discharge rate and BMS control command data are constructed into a state vector according to a time window, and the state vector and the running feature vector are concatenated according to the feature dimension to obtain the running behavior embedding vector. The embedded vector of the running behavior is processed by performing a sequence arrangement in chronological order to obtain the running behavior sequence.

4. The energy storage anomaly monitoring method based on deep learning according to claim 1, characterized in that, S3 includes the following steps: The running behavior sequence is divided into several training sample sequences according to time order, and the running behavior embedding vector of the first few time steps in each training sample sequence is input into the encoder of the autoencoder neural network model. Temporal feature compression calculation is performed on the running behavior embedding vector of the first few time steps to obtain the corresponding historical behavior representation vector. The historical behavior representation vector is input into the decoder of the autoencoder neural network model, and future behavior prediction calculation is performed on the historical behavior representation vector to obtain the predicted behavior vector for the corresponding subsequent time step. The prediction error is calculated based on the difference between the predicted behavior vector and the actual subsequent time step running behavior embedding vector. The model training loss is calculated using a loss function based on the prediction error, and backpropagation is performed on the autoencoder neural network model to update the network parameters based on the model training loss until the model converges. Based on the trained autoencoder neural network model, trajectory connection processing is performed on the relationship between the predicted behavior vector of subsequent time steps under stable operation and the change in the time series to obtain the normal behavior trajectory vector of the energy storage system.

5. The energy storage anomaly monitoring method based on deep learning according to claim 4, characterized in that, The loss function is as follows: ； in, Use the loss value to train the model; t is the number of training samples; t is the index of the current time step in the training sample sequence; k is the prediction step size. The embedding vector of the actual running behavior of the i-th training sample at time t+k; The autoencoder neural network model predicts the behavior vector at time step (t+k) based on the embedded vector of the running behavior before time step t. These are the weighting coefficients for trajectory smoothing constraints.

6. The energy storage anomaly monitoring method based on deep learning according to claim 4, characterized in that, The calculation of the behavior offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector includes the following steps: The trajectory distance offset is calculated based on the minimum distance between the current low-dimensional behavior representation vector and the normal behavior trajectory vector, and the trajectory direction offset is calculated based on the angle between the change direction of the current low-dimensional behavior representation vector and the tangential direction of the normal behavior trajectory. The trajectory evolution velocity offset is calculated based on the change amplitude between the low-dimensional behavior representation vectors of adjacent time steps, and the trajectory distance offset, trajectory direction offset, and trajectory evolution velocity offset are combined to form a behavior offset feature vector.

7. The energy storage anomaly monitoring method based on deep learning according to claim 6, characterized in that, S5 includes the following steps: The anomaly detection feature vector is compared with the feature vectors of each anomaly pattern in the pre-built anomaly pattern feature library to obtain the feature distance between the anomaly detection feature vector and each anomaly pattern. Based on the feature distance, perform minimum distance matching processing, and determine the anomaly type corresponding to the anomaly pattern with the smallest distance as the anomaly type result of the energy storage system.

8. The energy storage anomaly monitoring method based on deep learning according to claim 7, characterized in that, The abnormal pattern feature library is constructed through the following steps: Obtain operational data samples with anomaly types marked in the historical operational data of the energy storage system, and construct an embedding vector of abnormal operational behavior; The abnormal operating behavior embedding vector is input into the trained autoencoder neural network encoder, and feature encoding processing is performed on the abnormal operating behavior embedding vector to obtain the corresponding abnormal low-dimensional behavior representation vector. The abnormal low-dimensional behavior representation vectors are classified according to different abnormality types, and cluster centers are calculated for the abnormal low-dimensional behavior representation vectors corresponding to the same abnormality type to obtain the abnormal pattern feature vectors corresponding to each abnormality type. The anomaly pattern feature vectors corresponding to each anomaly type are associated and stored with the corresponding anomaly type labels to construct an anomaly pattern feature library.

9. The energy storage anomaly monitoring method based on deep learning according to claim 8, characterized in that, The types of anomalies include abnormal battery temperature, abnormal cell consistency, abnormal charge / discharge rate, abnormal power fluctuation, and abnormal sensor acquisition.

10. A deep learning-based energy storage anomaly monitoring system, applied to the deep learning-based energy storage anomaly monitoring method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire voltage, current, power, battery temperature distribution, SOC, SOH, charge / discharge rate and BMS control command data during the historical stable operation of the energy storage system, and obtain a stable operation dataset. The behavior sequence construction module is used to perform joint feature modeling on the stable running dataset, obtain running behavior embedding vectors, and arrange them in chronological order to obtain a running behavior sequence. The model building module is used to input the embedded vector of the running behavior into the autoencoder neural network model based on the running behavior sequence to perform feature encoding processing, obtain a low-dimensional behavior representation vector, and construct the normal behavior trajectory vector of the energy storage system according to the change relationship of the low-dimensional behavior representation vector in the time series. The offset feature extraction module is used to obtain the real-time operation dataset during the real-time operation of the energy storage system, input the real-time operation dataset into the autoencoder neural network model to generate the current behavior embedding vector, and calculate the behavior offset feature vector of the current behavior embedding vector relative to the normal behavior trajectory vector. The anomaly monitoring module is used to perform anomaly determination processing based on the continuous offset of the behavior offset feature vector within a continuous time window, and obtain anomaly monitoring results.

Images