Energy storage power station fault identification and safety early warning method, device and system

Abstract

The invention provides an energy storage power station fault identification and safety early warning method, device and system, and the method comprises the steps: obtaining the operation data of an energy storage system, and obtaining a standardized feature data set; based on the standardized feature data set, constructing a search space model containing candidate operations and connection modes; based on the search space model, constructing a Super network sharing weight, and generating an accelerated evaluation model; performing architecture evaluation and optimization of data perception by using an accelerated evaluation model to obtain an optimal neural network architecture; performing feature extraction and analysis on the operation data of the energy storage system by adopting the optimal neural network architecture to obtain a fault recognition result; and based on the fault identification result, establishing a safety early warning model, and generating early warning information. According to the method, efficient and accurate fault identification and early warning of the energy storage power station are realized through a neural architecture search technology of data perception, and the technical problems of insufficient model generalization ability and low deployment efficiency of a traditional method in a resource-constrained environment are solved.

Description

Technical Field

[0001] This application relates to the field of power system safe operation technology, and in particular to a method, device and system for fault identification and safety early warning of energy storage power stations. Background Technology

[0002] As an important component of the new power system, the safe and stable operation of energy storage power stations has a significant impact on the reliability of the power grid. Fault identification and safety early warning for energy storage systems are among the key technologies for ensuring the safe operation of energy storage power stations.

[0003] Currently, common fault identification methods for energy storage systems mainly include model-based analysis methods and data-driven intelligent diagnostic methods. Among them, model-based methods analyze the system state by establishing a mathematical model of the energy storage system, which can reflect the physical characteristics of the system, but modeling is difficult when facing complex systems or nonlinear problems. On the other hand, data-driven methods use machine learning algorithms to analyze system operating data, and have strong adaptive and nonlinear modeling capabilities.

[0004] One existing technical solution uses deep learning to analyze the operational data of energy storage systems. It extracts fault features through a pre-defined neural network structure to achieve fault identification and early warning. This solution trains the neural network model with a large amount of historical data, establishing a mapping relationship between fault features and system states, effectively identifying common fault types and providing early warning functionality.

[0005] However, this approach suffers from the following technical problems: the pre-defined neural network structure may not adapt well to different types of fault data characteristics, resulting in insufficient model generalization ability; in resource-constrained edge computing environments, complex neural network structures are difficult to deploy efficiently, affecting the effectiveness of real-time fault identification. These problems limit the application of existing technologies in practical energy storage systems and cannot meet the high-precision and real-time requirements of energy storage power stations for fault identification and safety early warning. Summary of the Invention

[0006] In view of this, this application provides a method, device and system for fault identification and safety early warning of energy storage power stations, which solves the problems of insufficient generalization ability of neural network structures and low deployment efficiency in edge computing environments in the prior art.

[0007] This application provides a method for fault identification and safety early warning of an energy storage power station, including:

[0008] Obtain the operating data of the energy storage system, and perform data cleaning and standardization on the operating data to obtain a standardized feature dataset;

[0009] Based on the standardized feature dataset, a search space model containing candidate operations and connection patterns is constructed.

[0010] Based on the search space model, a Supernet network with shared weights is constructed to generate an accelerated evaluation model;

[0011] The accelerated evaluation model is used to evaluate and optimize the data-aware architecture, resulting in the optimal neural network architecture.

[0012] The optimal neural network architecture is used to extract features from and analyze the operating data of the energy storage system to obtain fault identification results.

[0013] Based on the fault identification results, a safety early warning model is established, and early warning information is generated.

[0014] Optionally, the step of acquiring the operating data of the energy storage system, and performing data cleaning and standardization on the operating data to obtain a standardized feature dataset includes:

[0015] The voltage, current, temperature and state of charge data of the energy storage system are acquired, and the data is cleaned by removing outliers and processing missing values.

[0016] The cleaned data is normalized to unify data of different dimensions to the same scale, thus obtaining standardized data.

[0017] Based on the standardized data, a time-series feature dataset containing normal operation data and fault data is constructed as the standardized feature dataset.

[0018] Optionally, constructing a search space model containing candidate operations and connection patterns based on the standardized feature dataset includes:

[0019] By selecting basic operations including 3×3 convolution, 5×5 convolution, max pooling, and self-attention mechanism, the standardized feature dataset is analyzed to obtain the basic operator definition;

[0020] Based on the definition of the basic operator, a skip connection with bidirectional information flow and a dense connection mode with multi-layer feature fusion are designed to obtain the connection strategy.

[0021] The basic operator definitions are applied to the network topology structure set by the connection strategy to form a hierarchical search space with multiple search dimensions, thus obtaining a search space model.

[0022] Optionally, based on the search space model, a Supernet network with shared weights is constructed to generate an accelerated evaluation model, including:

[0023] A weight sharing strategy is applied to the search space model to construct a sub-network parameter reuse mechanism, thereby obtaining a weight sharing scheme.

[0024] Based on the weight sharing scheme, a progressive architecture sampling with temperature control is implemented to obtain the training framework;

[0025] Using the training framework described above, a performance prediction model containing an accuracy predictor and a delay predictor is trained to generate an accelerated evaluation model.

[0026] Optionally, the accelerated evaluation model is used to perform data-aware architecture evaluation and optimization to obtain the optimal neural network architecture, including:

[0027] Using the accelerated evaluation model, the architecture performance and resource consumption are evaluated to obtain the architecture evaluation metrics;

[0028] Based on the architecture evaluation metrics, multi-objective optimization is performed to balance model performance and resource constraints, and the optimized architecture is output.

[0029] Based on the optimized architecture, a verification test is performed, and the optimal neural network architecture is output.

[0030] Optionally, the optimal neural network architecture is used to extract and analyze features from the energy storage system's operating data to obtain fault identification results, including:

[0031] Using the optimal neural network architecture, multi-level features of the energy storage system operation data are extracted to obtain the feature data.

[0032] Based on the feature data, multi-scale feature fusion is performed to generate the fused feature;

[0033] Optionally, a fault classifier is constructed based on the fusion features to generate the fault identification result, wherein the fault identification result includes fault location information, fault type information, and fault severity assessment information.

[0034] Based on the fault identification results, a safety early warning model is established, and early warning information is generated, including:

[0035] The fault identification results are input into the fault development trend analysis model to obtain the trend prediction results;

[0036] Based on the trend prediction results, multi-level early warning thresholds are designed to achieve graded early warning, and the early warning level and corresponding handling suggestions are obtained.

[0037] Based on the warning level and the processing suggestions, a warning decision rule is constructed, and the warning information is generated.

[0038] Accordingly, this application also provides a device for fault identification and safety early warning of an energy storage power station, comprising:

[0039] The data preprocessing module is used to acquire the operating data of the energy storage system, perform data cleaning and standardization on the operating data, and obtain a standardized feature dataset.

[0040] The search space construction module is used to construct a search space model containing candidate operations and connection patterns based on the standardized feature dataset.

[0041] The Supernet building module is used to construct a Supernet network with shared weights based on the search space model, and generate an accelerated evaluation model.

[0042] The architecture optimization module is used to perform data-aware architecture evaluation and optimization using the accelerated evaluation model to obtain the optimal neural network architecture.

[0043] The fault identification module is used to extract and analyze features from the operating data of the energy storage system using the optimal neural network architecture to obtain fault identification results.

[0044] The safety early warning module is used to establish a safety early warning model and generate early warning information based on the fault identification results.

[0045] Optionally, the fault identification module includes:

[0046] The feature extraction unit is used to extract multi-level features using the optimal neural network architecture to obtain the feature data;

[0047] The feature fusion unit is used to perform multi-scale feature fusion based on the feature data and output the fused features;

[0048] The fault classification unit is used to construct a fault classifier based on the fused features and output the fault identification result.

[0049] Accordingly, this application also provides a system for fault identification and safety early warning of energy storage power stations, including:

[0050] The device as described above;

[0051] The data acquisition unit is used to collect operational data from the energy storage system.

[0052] A data storage unit is used to store the operating data, the fault identification results, and the early warning information;

[0053] The display unit is used to display the fault identification results and the warning information.

[0054] This application embodiment also provides a computer device, the computer device comprising:

[0055] At least one processor; and,

[0056] A memory communicatively connected to the at least one processor; wherein,

[0057] The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the above-described method for fault identification and safety early warning of energy storage power stations.

[0058] This application also provides a computer-readable storage medium that stores computer instructions for causing a computer to execute the above-described method for fault identification and safety warning in an energy storage power station.

[0059] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the above-described method for fault identification and safety warning of energy storage power stations.

[0060] This application has the following technical effects:

[0061] 1. By using a data-aware neural architecture search method, automated model structure design was achieved, improving the model's adaptability to different types of fault data and solving the problem of insufficient generalization ability of traditional preset neural network architectures;

[0062] 2. By using Supernet to accelerate the evaluation model, the efficiency of architecture search is significantly improved and the consumption of computing resources is reduced, making it possible to achieve efficient model deployment in resource-constrained environments;

[0063] 3. A multi-objective optimization strategy was implemented, achieving a good balance between model performance and resource consumption, and adapting to the constraints of the edge computing environment;

[0064] 4. A multi-level safety early warning mechanism was constructed, which, combined with fault characteristics and development trends, achieved accurate fault early warning and improved the timeliness and effectiveness of fault handling. Attached Figure Description

[0065] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0066] Figure 1 A flowchart illustrating a fault identification and safety early warning method for an energy storage power station, provided as an embodiment of this application;

[0067] Figure 2 A detailed flowchart of the data preprocessing steps provided for embodiments of this application;

[0068] Figure 3 A detailed flowchart of the search space construction steps provided for embodiments of this application;

[0069] Figure 4 A detailed flowchart of the Supernet accelerated evaluation model construction steps provided for embodiments of this application;

[0070] Figure 5 A detailed flowchart of the data-aware architecture optimization steps provided in the embodiments of this application;

[0071] Figure 6 This is a schematic diagram of a device structure for fault identification and safety early warning in an energy storage power station, provided as an embodiment of this application. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0073] like Figure 1 As shown in the figure, this application provides a method for fault identification and safety early warning of an energy storage power station, including:

[0074] S1: Obtain the operating data of the energy storage system, perform data cleaning and standardization on the operating data, and obtain a standardized feature dataset.

[0075] Data acquisition and preprocessing for energy storage systems are the foundational steps of the entire method.

[0076] In this step, key operating parameters such as voltage, current, temperature, and state of charge are first collected using multi-source sensors deployed within the energy storage system. These parameters comprehensively reflect the operating status of the energy storage system, providing a data foundation for subsequent fault identification. Subsequently, the collected raw data undergoes data cleaning, including outlier identification and removal, and missing value handling. Outlier identification employs statistical methods such as the 3σ criterion or interquartile range method, combined with expert knowledge to determine reasonable threshold ranges for screening. For missing values ​​in the data, appropriate interpolation methods are selected based on data characteristics, such as time series interpolation, K-nearest neighbor interpolation, or multiple interpolation, to ensure data integrity and reliability. After data cleaning, standardization is required due to differences in the dimensions and numerical ranges of different parameters. This method uses max-min standardization or Z-score standardization to map all parameters to a uniform scale, facilitating subsequent analysis and comparison. Considering the temporal characteristics of energy storage system faults, a time-series feature dataset is finally constructed using the sliding window method. By setting appropriate time window length and sliding step size, time-series segments are extracted from the continuous data stream to form a standardized feature dataset containing normal operation data and various types of fault data.

[0077] S2: Based on the standardized feature dataset, construct a search space model that includes candidate operations and connection patterns.

[0078] In the data-aware neural architecture search space construction step, a series of basic operations are first defined, including convolution operations of different sizes (such as 3×3 convolution and 5×5 convolution), pooling operations (such as max pooling and average pooling), and attention mechanisms (such as self-attention and channel attention). The selection of these operators is based on the characteristics of energy storage system data, taking into account both the ability to extract temporal features and computational efficiency.

[0079] Convolutional kernels of different sizes are suitable for capturing temporal patterns at different scales, while attention mechanisms can highlight the influence of key moments and important parameters, and pooling operations help reduce the number of parameters and extract abstract features. Based on the defined basic operations, various connection pattern strategies are designed, including residual connections, dense connections, and skip connections.

[0080] Specifically, a skip connection with bidirectional information flow was designed, allowing deep features to guide shallow feature extraction while shallow features supplement the detailed information of deep features. A dense connection with multi-level feature fusion was also designed, supporting the effective integration of features at different levels. Furthermore, a learnable connection strength parameter was introduced, enabling the network to adaptively adjust the importance of different connections. Based on these basic operations and connection patterns, a hierarchical search space was constructed, organized progressively from the basic operation layer, feature extraction block layer to the network architecture layer. In the basic operation layer, each node can choose different basic operations; in the feature extraction block layer, the combination of operations within the block and the connection patterns are defined; in the network architecture layer, the depth, width, and cross-layer connection methods of the overall network are determined. Simultaneously, an architecture encoding scheme was established, mapping each architecture scheme in the search space to a quantifiable representation, laying the foundation for subsequent architecture search and evaluation.

[0081] S3: Based on the search space model, construct a Supernet network with shared weights to generate an accelerated evaluation model.

[0082] Supernet's accelerated evaluation model construction is a key step in improving the efficiency of neural architecture search. First, a weight-sharing strategy was designed, an innovative parameter reuse mechanism. The core idea is to allow all candidate subnetworks in the search space to share a single set of weight parameters.

[0083] In practice, a path-based sampling method is used, randomly activating different paths in Supernet during training to allow shared weights to adapt to different network structures. To address the issue of weight interference between different sub-networks, a normalization adjustment factor is introduced to dynamically adjust parameters based on the size and topological characteristics of the sub-networks.

[0084] Furthermore, a hierarchical sharing strategy was designed, where operations with high similarity share more parameters, while operations with large differences share fewer parameters, balancing generality and specificity. Based on the weight sharing strategy, a gradual architecture sampling with temperature control was implemented. A sampling strategy from simple to complex was adopted, primarily sampling simple network structures in the early stages of training, and gradually increasing the sampling probability of more complex structures as training progresses.

[0085] This progressive strategy controls sampling through temperature parameters. A higher initial temperature ensures a more uniform sampling distribution, while the temperature gradually decreases as training progresses, focusing sampling more on high-performance candidate structures. Simultaneously, a performance-based adaptive sampling ratio adjustment mechanism is introduced, increasing the sampling weight for high-performing architecture types to accelerate convergence. Using this training framework, a performance prediction model incorporating an accuracy predictor and a latency predictor is trained. The accuracy predictor predicts the accuracy of the sub-network on the target task, while the latency predicts the runtime latency of the sub-network on the target hardware platform. Both predictors employ a lightweight neural network architecture and are trained using supervised learning, with training data derived from real performance data collected during Supernet evaluation. The resulting accelerated evaluation model supports millisecond-level architecture performance evaluation, significantly improving architecture search efficiency.

[0086] S4: Use the accelerated evaluation model to evaluate and optimize the data-aware architecture to obtain the optimal neural network architecture.

[0087] The data-aware architecture optimization phase first utilizes an accelerated evaluation model to comprehensively evaluate candidate architectures in the search space. Evaluation metrics include multiple dimensions: model performance metrics (such as accuracy, recall, and F1 score) measure the architecture's fault identification capability; resource consumption metrics (such as parameter count, computational complexity, and memory usage) assess computational resource requirements; and feature expressiveness metrics (such as feature diversity and feature discriminative power) measure the model's ability to capture different fault features. To comprehensively evaluate architecture performance, a comprehensive scoring function is designed, which weights and sums the various metrics according to their importance to obtain a single architecture evaluation score.

[0088] Simultaneously, an evaluation results database was established to store and index the performance data of the evaluated architectures, providing a reference for subsequent optimization. Based on the architecture evaluation metrics, multi-objective optimization was implemented to balance model performance and resource constraints. The Pareto optimization method was adopted, considering multiple conflicting objectives simultaneously. A reinforcement learning-based architecture search agent was designed, modeling architecture search as a sequential decision problem. The agent learns how to modify the architecture to obtain higher rewards (overall performance scores) by interacting with the environment. To adapt to the needs of different application scenarios, an adaptive weighting mechanism was implemented, dynamically adjusting the importance of each objective according to the characteristics of the deployment environment. For example, more emphasis is placed on model lightweighting on edge devices, while more emphasis is placed on accuracy in scenarios with high performance requirements.

[0089] Finally, the optimized architecture was validated and tested to determine the optimal neural network architecture. A representative set of architectures was selected from the Pareto optimal set, and their performance was evaluated on independent validation datasets. The validation process included model training stability testing, generalization ability testing, and robustness testing. Simultaneously, deployment tests were conducted in a real hardware environment to verify the model's actual performance on the target platform. Based on the comprehensive test results, an optimal balance was determined as the optimal neural network architecture, which balances performance and resource constraints in the current application scenario.

[0090] S5: The optimal neural network architecture is used to extract and analyze the features of the energy storage system operation data to obtain fault identification results.

[0091] The fault feature extraction and identification stage first utilizes the optimal neural network architecture to perform multi-level feature extraction.

[0092] This process employs a hierarchical feature extraction strategy: shallow networks primarily extract the basic features of the energy storage system, such as time-domain features like voltage fluctuations and current changes; mid-level networks are responsible for extracting combined features, such as power curve features and temperature distribution features; and deep networks focus on high-level semantic features, such as fault mode features and system state features.

[0093] In practical implementation, shallow layers employ convolutional operations with small receptive fields to capture local variation features; mid-layers use dilated convolutions or convolutions with larger receptive fields to expand the feature perception range; and deep layers introduce a self-attention mechanism to establish long-distance dependencies and capture global patterns. To improve the robustness of feature extraction, an adaptive feature selection mechanism is introduced, which can dynamically adjust the focus of feature extraction according to the data characteristics under different working conditions. After acquiring multi-level features, multi-scale feature fusion is performed. An innovative feature fusion architecture is designed, including a feature alignment module, an attention fusion module, and a feature enhancement module.

[0094] For feature alignment, adaptive pooling and dimensionality transformation operations are used to map features of different scales to a unified representation space. For attention fusion, a multi-head attention mechanism is employed to evaluate feature importance from different perspectives. For feature enhancement, residual connections and feature activation adjustments are introduced to enhance the expressive power of key features. A bidirectional feature propagation mechanism is applied during the fusion process to support the mutual complementarity of high- and low-level features.

[0095] Furthermore, a feature filtering mechanism based on the information bottleneck theory was implemented to remove redundant features and retain the most discriminative feature combinations. Based on the fused features, a hierarchical fault classifier structure was constructed: first, the major fault categories were determined, and then each major category was further subdivided into specific fault types. The classifier incorporates advanced techniques such as ensemble learning methods, Bayesian uncertainty estimation, and adaptive threshold adjustment to improve the stability and accuracy of classification. For fault location identification, a gradient-based class activation mapping localization technique was adopted to generate heatmaps that visually display the areas where faults may occur. For fault severity assessment, a multi-index comprehensive scoring model was designed to quantitatively evaluate the severity level of faults, providing a basis for subsequent early warning and handling.

[0096] S6: Based on the fault identification results, establish a safety early warning model and generate early warning information.

[0097] In the safety early warning model construction and prediction stage, a fault development trend analysis model was first established.

[0098] The model employs a time-series prediction framework, combining multiple advanced time-series analysis techniques: Long Short-Term Memory (LSTM) networks are used to capture long-term dependencies, a time attention mechanism is used to identify key time points, and a conditional variational autoencoder is used to model the uncertainty of fault evolution.

[0099] Specifically, the Long Short-Term Memory (LSTM) network processes historical data sequences and learns temporal patterns of parameter changes; the temporal attention layer automatically identifies the most important historical moments for prediction, enhancing sensitivity to key points; and the variational autoencoder represents the model's predictions of future uncertainty through probability distributions, aiding in risk level assessment. Furthermore, the model integrates a historical fault case library, using similar case matching to assist trend prediction. A dynamic time window mechanism is employed, adaptively adjusting the prediction interval based on the development characteristics of different fault types. Based on the trend prediction results, a multi-level early warning threshold system is designed. This system establishes four early warning levels (normal, attention, warning, and danger), each with its unique threshold definition mechanism. Threshold settings comprehensively consider multiple factors, including historical statistical data, expert experience, equipment specifications, and dynamic adjustment factors.

[0100] A fuzzy logic approach is employed to handle threshold boundaries, and a membership function is set for each warning level to prevent frequent switching of warning levels due to slight parameter fluctuations. Simultaneously, a threshold self-learning mechanism is implemented, enabling dynamic optimization of threshold settings based on newly added operational data and fault cases. For each warning level, the system also provides corresponding handling suggestions, including operational guidance such as continuing monitoring, scheduling inspections, reducing load, and emergency shutdown.

[0101] Finally, an early warning decision-making rule system was constructed. This system adopts a hierarchical structure, including a basic rule layer, a combined rule layer, and a scenario rule layer. The rule construction process integrates multiple technologies such as decision trees, Bayesian networks, and rule mining algorithms. The system places particular emphasis on the interpretability of the rules; each rule is accompanied by a triggering reason explanation, facilitating operators' understanding and verification of the rationality of the early warning. Simultaneously, a rule conflict detection and priority management mechanism is implemented to ensure consistent early warning decisions in complex situations. Ultimately, the system generates a complete early warning information package, including the early warning level, detailed diagnosis, recommended measures, and emergency plans, which are intuitively displayed through a visual interface. It also supports intelligent push functionality to promptly deliver early warning information to relevant responsible persons.

[0102] In one embodiment, such as Figure 2 As shown, S1 specifically includes:

[0103] S1.1: Obtain the voltage, current, temperature and state of charge data of the energy storage system, and perform outlier removal and missing value processing on the data to obtain cleaned data.

[0104] First, in S1.1, this embodiment of the application collects key operating parameters such as voltage, current, temperature, and state of charge (SOC) through multi-source sensors deployed inside the energy storage system. These parameters comprehensively reflect the operating status of the energy storage system and are the basic data for fault identification. The collected raw data undergoes data cleaning processing, including outlier identification and removal, and missing value imputation. Specifically, statistical methods (such as the 3σ criterion or interquartile range method) are used to identify outliers, and a reasonable threshold range is determined in combination with expert knowledge. For missing values, an appropriate imputation method is selected based on the data characteristics, such as time series interpolation, K-nearest neighbor interpolation, or multiple imputation methods, to ensure the integrity and reliability of the data.

[0105] S1.2: Normalize the cleaned data to unify data of different dimensions to the same scale and obtain standardized data.

[0106] In S1.2, the cleaned data is standardized. Since the dimensions and numerical ranges of different parameters vary significantly, standardization is necessary to ensure that all parameters are compared and analyzed on the same scale. This embodiment uses the min-max standardization method to map all parameter values ​​to the interval [0, 1]. The specific formula is:

[0107] X' = ​​(X - Xmin) / (Xmax - Xmin)

[0108] Where X represents the original data, Xmin and Xmax represent the minimum and maximum values ​​of the parameter, respectively, and X' represents the standardized data. Furthermore, depending on the specific characteristics of the parameter, other methods such as Z-score standardization can be selected to obtain a data distribution more suitable for subsequent analysis.

[0109] S1.3: Based on the standardized data, construct a time-series feature dataset containing normal operation data and fault data, as the standardized feature dataset.

[0110] Considering the distinct temporal development characteristics of energy storage system faults, this embodiment constructs time-series samples using a sliding window method. A suitable time window length (e.g., 10 minutes) and sliding step size (e.g., 30 seconds) are set to extract time-series segments from the continuous data stream, forming training samples. Each sample contains input features (parameter sequence within the window) and label information (normal / fault type). A balanced sampling strategy ensures an appropriate ratio of normal operation data to various fault types, avoiding class imbalance that could negatively impact model performance. Simultaneously, data augmentation techniques are implemented to expand sample diversity and improve model robustness by adding appropriate noise and time scale transformations.

[0111] like Figure 3 As shown, S2 specifically includes:

[0112] S2.1: Select basic operations including 3×3 convolution, 5×5 convolution, max pooling and self-attention mechanism, analyze the standardized feature dataset, and obtain the basic operator definition.

[0113] In S2.1, this embodiment first defines a set of basic operations. This set includes the basic computational units commonly used in neural networks, mainly consisting of the following types of operations: standard convolution operations (including 3×3 and 5×5 convolution kernels), depthwise separable convolution, dilated convolution, and other convolutional operations; pooling operations such as max pooling and average pooling; and attention operations such as self-attention mechanisms and channel attention.

[0114] The selection of these basic operators is based on the characteristics of energy storage system data, taking into account both the ability to extract temporal features and computational efficiency. Specifically, convolutional kernels of different sizes are suitable for capturing temporal patterns at different scales, attention mechanisms can highlight the influence of key moments and important parameters, and pooling operations help reduce the number of parameters and extract abstract features.

[0115] S2.2: Based on the definition of the basic operator, design a skip connection with bidirectional information flow and a dense connection mode with multi-layer feature fusion to obtain the connection strategy.

[0116] In S2.2, a connection pattern search strategy is designed. The connection pattern determines the way information flows between different layers in the network and has a significant impact on the expressive power of the model.

[0117] This embodiment designs multiple connection modes, including residual connections (to alleviate the vanishing gradient problem), dense connections (to enable feature reuse), and skip connections (to support multi-scale feature fusion). Specifically, skip connections with bidirectional information flow are designed, allowing deep features to guide shallow feature extraction, while shallow features can also supplement the detailed information of deep features; dense connections with multi-level feature fusion are designed, supporting the effective integration of features at different levels. Furthermore, learnable connection strength parameters are introduced, enabling the network to adaptively adjust the importance of different connections.

[0118] S2.3: Apply the basic operator definition to the network topology structure set by the connection strategy to form a hierarchical search space with multiple search dimensions, and obtain the search space model.

[0119] In S2.3, a hierarchical search space is constructed based on the aforementioned basic operations and connection patterns.

[0120] The search space is organized hierarchically, from bottom to top: the basic operation layer, the feature extraction block layer, and the network architecture layer. In the basic operation layer, each node can choose different basic operations; in the feature extraction block layer, the combination and connection patterns of operations within the block are defined; and in the network architecture layer, the depth, width, and cross-layer connections of the overall network are determined. Simultaneously, an architecture encoding scheme is established, mapping each architecture scheme in the search space to a quantifiable representation. The encoding scheme uses a multi-part combination: the operation encoding part uses one-hot vectors to represent the operation selection of each node; the connection encoding part uses binary matrices to represent the connection relationships between nodes; and the architecture parameter encoding part includes hyperparameter information such as network depth and width. This systematic search space construction method lays the foundation for subsequent Supernet training and architecture optimization.

[0121] For example, in the construction of the neural network search space of the fault identification system of the energy storage power station, the embodiments of this application first define a set of basic operations.

[0122] Taking the scenario of abnormal battery cell temperature detection as an example, considering that temperature data has temporal and spatial distribution characteristics, this application embodiment selects a variety of basic operations suitable for such data.

[0123] Specifically, this application defines a 3×3 convolution operation to capture temperature correlation patterns between neighboring battery cells. This small-size convolution is particularly sensitive to the detection of local abrupt temperature changes. A 5×5 convolution operation is used to capture temperature distribution patterns over a larger range, which can detect temperature anomalies at the battery pack level. Max pooling helps this application extract temperature peaks within a region, which is very effective for hotspot identification. The self-attention mechanism allows the model to focus on long-range dependencies between different time points and different battery cells, which is particularly suitable for discovering temperature anomaly propagation paths.

[0124] Furthermore, this application incorporates a channel attention mechanism, enabling the model to automatically determine which parameter is more important for fault diagnosis in the current state from various sensor data (such as temperature, voltage, and current). For example, when the system is in a fast-charging state, current and temperature data may be more important than voltage data; while in a quiescent state, temperature anomalies may be a more critical fault indicator. The selection of these basic operators is directly based on the actual data characteristics and operating modes of the energy storage system, ensuring that the model's feature extraction capabilities are highly matched with the application scenario.

[0125] Based on the defined set of basic operations, embodiments of this application design a connection mode suitable for fault identification in energy storage systems. Taking battery pack overcharge fault identification as an example, this type of fault typically manifests as a complex interaction pattern of voltage, current, and temperature parameters.

[0126] This application's embodiments design a skip connection mechanism with bidirectional information flow, allowing deep feature feedback to guide shallow feature extraction. For example, when a deep network detects a certain voltage anomaly pattern, skip connections can feed this high-level feature back to the shallow layer, enabling the shallow network to focus more specifically on the relevant voltage fluctuation details.

[0127] Meanwhile, this application also designs a dense connection pattern for multi-layer feature fusion, which is particularly suitable for capturing the multi-stage evolution process of overcharge faults. In practical applications, overcharge faults often go through stages such as initial voltage anomalies, mid-term temperature rises, and late-stage chemical instability. Dense connections allow the model to integrate features from different time scales to form a complete fault evolution picture. This application also introduces learnable connection strength parameters, which enables the model to dynamically adjust the importance of different connections according to the characteristics of actual data.

[0128] For example, in energy storage systems operating in low-temperature environments, the connection strength based on temperature characteristics may be automatically increased because temperature characteristics are more sensitive to fault prediction under low-temperature conditions. This adaptive connection mode design fully considers the complexity and dynamics of energy storage system faults, significantly improving the model's ability to identify different types of faults.

[0129] After completing the basic operation definition and connection mode design, this embodiment of the application constructs a hierarchical neural architecture search space. Taking the battery thermal runaway risk prediction task of an energy storage power station as an example, the search space of this embodiment of the application is organized sequentially from the bottom layer to the top layer.

[0130] At the basic operation layer, each computation node can choose different basic operations. For example, node A can choose 3×3 convolution to process temperature data, node B can choose self-attention mechanism to process temporal features, and node C can choose max pooling to extract salient features.

[0131] In the feature extraction block layer, embodiments of this application define different combination methods and connection modes of operations. For example, nodes A and B can be combined through residual connections to form a hot feature extraction block, and nodes B and C can be combined through dense connections to form a temporal feature extraction block.

[0132] At the network architecture layer, the overall network depth (e.g., 8 or 12 layers), width (number of feature channels, such as 32 or 64), and cross-layer connection method (e.g., adding a skip connection every 2 layers) are determined.

[0133] To achieve effective search and evaluation, this application establishes a complete architecture encoding scheme, mapping each possible architecture configuration to a quantifiable representation. For example, for a neural network for thermal runaway prediction, the operation encoding part uses one-hot vectors to represent the operation selection of each node (e.g., [0, 1, 0, 0, 0] represents selecting a 3×3 convolution), the connection encoding part uses binary matrices to represent the connection relationships between nodes (e.g., node A connects to nodes C and D), and the architecture parameter encoding part includes hyperparameter information such as a network depth of 10 layers and 48 feature channels. This systematic search space construction method ensures that the search algorithm can find the most suitable neural network structure for the battery thermal runaway prediction task within a sufficiently large architecture space, while maintaining reasonable search efficiency.

[0134] like Figure 4 As shown, S3 specifically includes:

[0135] S3.1: Apply a weight sharing strategy to the search space model, construct a sub-network parameter reuse mechanism, and obtain a weight sharing scheme.

[0136] In S3.1, this application embodiment designs a weight sharing strategy. This is an innovative parameter reuse mechanism, the core idea of ​​which is to allow all candidate subnetworks in the search space to share a set of weight parameters.

[0137] In its implementation, a path-based sampling method is employed, randomly activating different paths within the Supernet during training to allow shared weights to adapt to different network structures. To address weight interference between different sub-networks, a normalization adjustment factor is introduced, dynamically adjusting parameters based on the sub-network's size and topological characteristics to ensure weights function correctly across various architectures. Furthermore, a hierarchical sharing strategy is designed: operations with high similarity share more parameters, while operations with significant differences share fewer parameters, balancing generality and specificity. This sharing mechanism significantly reduces required storage space and training time, enabling large-scale architecture search.

[0138] S3.2: Based on the weight sharing scheme, implement progressive architecture sampling with temperature control to obtain the training framework.

[0139] In S3.2, a progressive architecture sampling strategy with temperature control is implemented based on a weight-sharing strategy. To improve training efficiency, this embodiment adopts a sampling strategy from simple to complex: in the early stages of training, simple network structures are mainly sampled, and as training progresses, the sampling probability of more complex structures is gradually increased. This progressive strategy is controlled by a temperature parameter; the initial temperature is higher to make the sampling distribution more uniform, and the temperature gradually decreases as training progresses, making the sampling more focused on high-performance candidate structures. At the same time, a performance-based adaptive sampling ratio adjustment mechanism is introduced, increasing the sampling weight of well-performing architecture types to accelerate the convergence speed. To prevent overfitting to a specific subset of architectures, a diversity preservation strategy is designed to ensure that the search space is fully explored.

[0140] S3.3: Using the training framework, train a performance prediction model that includes an accuracy predictor and a delay predictor to generate an accelerated evaluation model.

[0141] In S3.3, a performance prediction model incorporating an accuracy predictor and a latency predictor is trained using the aforementioned training framework. The accuracy predictor takes the architecture encoding as input and predicts the accuracy of the subnetwork on the target task; the latency predictor predicts the runtime latency of the subnetwork on the target hardware platform. Both predictors employ a lightweight neural network structure and are trained using supervised learning, with training data derived from real performance data collected during the Supernet evaluation process. To improve prediction accuracy, this embodiment incorporates a feature engineering module to extract more discriminative architectural features, such as computational complexity, channel number distribution, and connection pattern statistics. Simultaneously, transfer learning techniques are employed to transfer knowledge trained on similar tasks to the current prediction task, accelerating convergence. The generated accelerated evaluation model supports millisecond-level architecture performance evaluation, significantly improving architecture search efficiency.

[0142] For example, in the construction of the Supernet for the fault identification system of the energy storage power station, the embodiments of this application design an innovative weight sharing strategy.

[0143] Taking the cycle life prediction task of lithium battery packs as an example, traditional methods require training thousands of candidate network architectures independently, resulting in extremely high computational costs. This application addresses this problem through a weight-sharing strategy.

[0144] In its specific implementation, this embodiment constructs a supernetwork encompassing all possible operations, such as a 12-layer network, where each layer may perform operations like 3×3 convolutions, 5×5 convolutions, or self-attention. During training, this embodiment employs a path-based sampling method, randomly activating a subset of network paths each time. For example, in a given iteration, this embodiment might activate a path containing a combination of 3×3 convolutions in layer 1, self-attention in layer 2, and 5×5 convolutions in layer 3, effectively training a specific sub-network; in the next iteration, this embodiment might activate a different path. This approach allows weights to be shared and reused across different network architectures. To address the training discrepancies between sub-networks of different sizes, this embodiment introduces a normalization adjustment factor.

[0145] For example, when activating a large subnetwork (such as one containing multiple 5×5 convolutional layers), this embodiment applies a smaller learning rate and weight decay; while for a smaller subnetwork, a larger learning rate is used to accelerate convergence.

[0146] Furthermore, this application embodiment also designs a hierarchical sharing strategy, allowing similar operations to share more parameters. For example, 3×3 convolutions and 5×5 convolutions employ a partial sharing mechanism in the kernel portion, while completely different operations (such as convolution and self-attention) reduce the degree of sharing. This refined weight sharing strategy enables this application embodiment to effectively evaluate tens of thousands of different neural network architectures in a single training process, greatly improving search efficiency.

[0147] Based on a weight-sharing strategy, this application implements a progressive architecture sampling framework with temperature control. Taking the state of charge (SOC) estimation task of an energy storage power station as an example, this is a complex regression task that requires the model to have good generalization ability.

[0148] In the initial training phase, this embodiment primarily samples simple network structures, such as shallow networks or networks containing basic convolutional layers. For example, in the first 20% of training epochs, this embodiment concentrates the sampling probability on simple networks with 3-5 layers to ensure that the basic feature extraction capability is sufficiently trained first. As training progresses, this embodiment gradually increases the sampling probability of more complex structures. In the middle of training (e.g., 20%-60% of training epochs), this embodiment increases the sampling ratio of medium-complexity networks, including networks with 6-8 layers and skip connections.

[0149] In the later stages of training (60%-100% of the training cycle), this embodiment further increases the sampling ratio of complex networks, including deep networks with 9-12 layers and complex connection patterns. This progressive strategy is controlled by a temperature parameter: in the initial stage of training, the temperature parameter is set to a higher value (e.g., T=5.0) to make the sampling distribution more uniform; as training progresses, the temperature is gradually reduced (e.g., eventually reduced to T=0.5) to make the sampling more focused on high-performance architectures.

[0150] Meanwhile, this application introduces an adaptive sampling mechanism based on historical performance. For example, if a network with a self-attention module is found to be particularly effective for the SOC estimation task, the system automatically increases the sampling probability of network architectures containing such modules. To prevent premature convergence to suboptimal architectures, this application also implements an explore-utilize balancing strategy, reserving 10%-20% of random sampling for exploring potential high-performance architectures. This progressive architecture sampling framework significantly improves training efficiency, enabling the model to find the most suitable network architecture for the SOC estimation task within a limited training time.

[0151] Using the training framework described above, this application embodiment trains a performance prediction model that includes an accuracy predictor and a delay predictor. Taking the fault classification task of an energy storage power station as an example, this task needs to run in real time on an edge computing device. The accuracy predictor receives the architecture encoding as input and predicts the accuracy of that architecture on the fault classification task.

[0152] For example, an architecture encoded as [10 layers, 64 channels, 60% 5×5 convolutions, 20% self-attention, 3 skip connections] might have a 93.5% prediction accuracy from the precision predictor. Meanwhile, a latency predictor receives the same architecture encoding but predicts the inference latency of that architecture on a target hardware platform (such as an ARM processor). For the same architecture, the latency predictor might output a prediction latency of 35 milliseconds. Both predictors employ lightweight neural network architectures: the precision predictor uses a 3-layer fully connected network with 128 neurons per layer, employing ReLU activation and a dropout rate of 0.2; the latency predictor uses a 2-layer fully connected network with 64 neurons per layer, also employing ReLU activation and a dropout rate of 0.1.

[0153] The predictor's training data comes from the real evaluation results of approximately 5000 sub-networks sampled during Supernet training. To improve prediction accuracy, this application's embodiments design a feature engineering module to extract more discriminative architectural features, such as the distribution percentage of different operation types, connection density, and network depth-to-width ratio.

[0154] Meanwhile, this application employs transfer learning technology to transfer predictor knowledge trained on similar tasks (such as voltage anomaly detection) to the current fault classification task, accelerating model convergence. The resulting accelerated evaluation model supports millisecond-level architecture performance evaluation, reducing the architecture evaluation process from several hours to just milliseconds, significantly improving architecture search efficiency. In practical applications, the prediction model achieves an accuracy of over 85%, providing a reliable performance estimation foundation for subsequent multi-objective optimization.

[0155] For the fault identification system of energy storage power stations, this application embodiment constructs an optimal network model through a data-aware neural architecture search method. This model, based on a neural network architecture selected after multi-objective optimization, mainly includes the following key components and parameter settings:

[0156] The network architecture employs a 10-layer deep structure, including 3 convolutional feature extraction layers, 2 self-attention modules, 2 cross-connection fusion layers, and 3 classification layers. The input layer receives preprocessed multi-dimensional time-series data, including key parameters such as voltage, current, temperature, and SOC. The input sequence length is set to 256 time steps, and the feature dimension is 12. In the convolutional feature extraction layers, the first layer uses 32 3×3 convolutional kernels with a stride of 1 and padding of 1; the second layer uses 64 3×3 convolutional kernels with a stride of 2 and padding of 1; and the third layer uses 128 5×5 convolutional kernels with a stride of 1 and padding of 2. Each convolutional layer is followed by a batch normalization layer with parameters set to momentum = 0.9 and epsilon = 1e. -5 And the ReLU activation function.

[0157] The self-attention module employs a 4-head attention mechanism with a hidden layer dimension of 256 and a dropout rate of 0.1 to prevent overfitting. The cross-connectivity fusion layer uses a residual connection structure, containing 128 feature channels and employing the GELU activation function. The classification layer uses a fully connected structure; the first layer contains 512 neurons with a dropout rate of 0.3; the second layer contains 256 neurons with a dropout rate of 0.2; the final layer has the number of neurons matching the number of fault categories (8 categories in this example), and uses the softmax activation function to output the final fault classification probability.

[0158] The model was trained using the following hyperparameter settings: batch size of 64, initial learning rate of 0.001, Adam optimizer with β1 = 0.9, β2 = 0.999, and epsilon = 1e. -8 A learning rate decay strategy was employed, reducing the learning rate to 90% every 20 epochs. To prevent overfitting, in addition to dropout at each layer, L2 regularization was used, with a weight decay coefficient set to 0.0001. An early stopping strategy was used during training, with a patience parameter set to 15; training would stop if the validation set performance did not improve for 15 consecutive epochs. Data augmentation techniques included adding Gaussian noise (standard deviation 0.01), random time shifting (±10 time steps), and random scaling (±5%). Training was conducted for a total of 200 epochs, achieving a final fault identification accuracy of 95.6% on the test set, with an average inference time of 25 milliseconds and a model size of approximately 15MB, fully meeting the deployment requirements of edge computing environments.

[0159] like Figure 5 As shown, S4 specifically includes:

[0160] S4.1: Using the accelerated evaluation model, evaluate the architecture performance and resource consumption to obtain the architecture evaluation index.

[0161] In S4.1, this embodiment uses an accelerated evaluation model to evaluate candidate architectures in the search space. Evaluation metrics include multiple dimensions: model performance metrics (such as accuracy, recall, F1 score, etc.) to measure the architecture's fault identification capability; resource consumption metrics (such as parameter count, FLOPs, memory usage, etc.) to evaluate computational resource requirements; and feature expressiveness metrics (such as feature diversity, feature discriminative power, etc.) to measure the model's ability to capture different fault features. To comprehensively evaluate architecture performance, this embodiment designs a comprehensive scoring function, which weights and sums the various metrics according to their importance to obtain a single architecture evaluation score. Simultaneously, an evaluation result database is established to store and index the performance data of the evaluated architectures, providing a reference for subsequent optimization.

[0162] S4.2: Based on the architecture evaluation metrics, perform multi-objective optimization, balance model performance and resource constraints, and output the optimized architecture.

[0163] This embodiment employs the Pareto optimization method, considering multiple conflicting objectives simultaneously. A reinforcement learning-based architecture search agent is designed, modeling architecture search as a sequential decision problem. The agent learns how to modify the architecture to obtain higher rewards (overall performance score) through interaction with the environment.

[0164] To adapt to the needs of different application scenarios, an adaptive weighting mechanism was implemented, dynamically adjusting the importance of each objective based on the characteristics of the deployment environment. For example, more emphasis is placed on model lightweighting on edge devices, while accuracy is prioritized in high-performance server environments. Furthermore, discrete structure optimization techniques were introduced, using methods such as genetic algorithms or simulated annealing to efficiently search within the discrete architecture space, avoiding the limitations of gradient optimization in discrete spaces.

[0165] S4.3: Based on the optimized architecture, perform verification tests and output the optimal neural network architecture.

[0166] This embodiment first selects a representative set of architectures from the Pareto optimal set, and then performs performance evaluation on an independent validation dataset. The validation process includes: model training stability testing (checking convergence speed and stability), generalization ability testing (performance under different operating conditions), and robustness testing (sensitivity to noise and data perturbations). Simultaneously, deployment tests are conducted in a real hardware environment to verify the model's actual performance on the target platform, including metrics such as memory usage, inference speed, and power consumption. Based on the comprehensive test results, an optimal balance point is finally determined as the optimal neural network architecture, which balances performance and resource constraints in the current application scenario.

[0167] In the process of optimizing the architecture of the fault identification system for energy storage power stations, the embodiments of this application first use an accelerated evaluation model to comprehensively evaluate the candidate architecture.

[0168] Taking floating voltage fault detection as an example, this application embodiment requires the model to be able to accurately identify minute voltage anomalies and be efficiently deployed on edge computing devices in power plants. This application embodiment evaluated more than 10,000 candidate architectures in the search space, and each architecture was measured by multi-dimensional metrics.

[0169] For example, for a candidate architecture A with an encoding of [8 layers, 48 ​​channels, 70% 3×3 convolutions, 15% attention mechanism, and 2 dense connections], its model performance metrics include: voltage fault detection accuracy of 94.2%, recall of 92.8%, and F1 score of 93.5%; while resource consumption metrics include: 685K parameters, 120M floating-point operations per second (FLOPs), 2.8MB of static memory usage, and 18ms inference time (tested on an ARM Cortex-A72 processor).

[0170] Meanwhile, this application embodiment also evaluated feature expressiveness metrics, such as sensitivity to minute voltage fluctuations (quantitatively expressed as 0.82, with a maximum score of 1) and the ability to distinguish different fault types (mean of diagonal elements in the confusion matrix is ​​0.91). To comprehensively evaluate the overall performance of architecture A, this application embodiment designed a weighted scoring function with weight ratios of [accuracy: 40%, latency: 30%, memory: 20%, feature expressiveness: 10%], ultimately obtaining a comprehensive score of 87.6 out of 100 for architecture A. Simultaneously, this application embodiment also evaluated another candidate architecture B, encoded as [10 layers, 64 channels, 50% 5×5 convolutions, 25% self-attention, 3 skip connections], which achieved higher accuracy (96.8%) but also consumed more resources (1.2M parameters, 32ms inference time), ultimately achieving a comprehensive score of 85.3. These evaluation results were stored in the evaluation database, providing foundational data for subsequent multi-objective optimization, enabling this application embodiment to find the optimal balance between accuracy and resource efficiency.

[0171] Based on the architecture evaluation results, this application implements a multi-objective optimization strategy to balance model performance and resource constraints. Taking the thermal runaway early warning task of battery packs in an energy storage power station as an example, this task has high requirements for accuracy and real-time performance.

[0172] The embodiments of this application employ the Pareto optimization method, which simultaneously considers multiple objectives such as early warning accuracy, false alarm rate, early warning lead time, and computational resource consumption.

[0173] In its specific implementation, this application's embodiments design a reinforcement learning-based architecture search agent, modeling architecture search as a sequential decision problem. The search agent's state space contains all features and evaluation metrics of the current architecture, while its action space contains various architecture modification operations, such as adding / deleting layers, changing operation types, and adjusting connection patterns. The reward function is designed as a comprehensive performance score, considering multiple objectives simultaneously.

[0174] For example, in a thermal runaway early warning task, this embodiment sets a baseline architecture (90% accuracy, 5% false alarm rate, 15-minute warning lead time, 8MB model size). When the search agent improves performance by modifying the architecture (e.g., replacing the 3rd layer from a regular convolution to a depthwise separable convolution, and adding a skip connection between layers 5-7) (accuracy increases to 92%, false alarm rate decreases to 4%, and parameter count decreases by 20%), it receives a positive reward; otherwise, it receives a negative reward. To adapt to the needs of different deployment environments, this embodiment implements an adaptive weighting mechanism. For example, on equipment in critical areas of a power plant, the weights of accuracy and false alarm rate are set higher (40% and 30% respectively), while in non-critical areas, the weight of resource consumption is relatively higher (40%). During the search process, the agent explores the architecture space using a Monte Carlo tree search algorithm, selecting the most promising modification operation to gradually optimize the architecture. In the later stages of optimization, this embodiment introduces a simulated annealing mechanism, allowing the agent to occasionally accept modifications that slightly degrade performance in order to escape local optima. After approximately 500 rounds of architectural optimization, this embodiment yields a series of candidate architectures located at the Pareto front, each representing an optimal balance between different performance metrics.

[0175] The optimized Pareto front architecture set was rigorously verified and tested in this application embodiment to determine the final optimal neural network architecture.

[0176] Taking the capacity decay prediction task of energy storage power stations as an example, this application selected five representative architectures from the Pareto front for detailed testing. Each architecture was evaluated on an independent validation dataset (containing historical data from five different power stations). For example, architecture C (encoded as [9 layers, 56 channels, 60% depthwise separable convolutions, 20% attention mechanism, 4 residual connections]) had an average prediction error of 3.2% on the validation set, a training convergence speed of 45 epochs, a performance stability (measured by standard deviation) of 0.8% under different operating conditions, and a robustness loss to 5% Gaussian noise with only a 2.1% accuracy decrease.

[0177] Meanwhile, deployment tests were conducted on the target hardware platform (Jetson Nano embedded device) in this embodiment. Architecture C had an actual memory footprint of 4.3MB, an average inference time of 27ms, and a power consumption of 0.3W. In contrast, another candidate architecture, D, had a lower prediction error (2.8%), but a longer inference time (38ms) and a larger memory footprint (6.1MB). Considering that the capacity decay prediction task does not have very high real-time requirements (minute-level response is sufficient) but has high requirements for prediction accuracy, this embodiment adjusted the selection weight, increasing the accuracy weight to 60%, and ultimately selected architecture D as the optimal architecture. To further verify this selection, this embodiment also conducted model adaptability tests, deploying the model to a new power plant environment. After fine-tuning with a small amount of data, architecture D demonstrated excellent transfer learning capabilities, with the prediction error increasing by only 0.5%. Through this rigorous verification testing process, this embodiment ensures that the finally selected neural network architecture not only performs excellently in the laboratory environment but also runs stably and reliably in the actual power plant environment, providing accurate decision support for the capacity decay prediction task.

[0178] S5 specifically includes:

[0179] S5.1: Using the optimal neural network architecture, extract multi-level features of the energy storage system operation data to obtain the feature data.

[0180] In S5.1, the embodiments of this application utilize the optimal neural network architecture for multi-level feature extraction.

[0181] This process employs a multi-layered feature extraction strategy: shallow networks primarily extract the basic features of the energy storage system (such as time-domain features like voltage fluctuations and current changes), mid-layer networks are responsible for extracting combined features (such as power curve features and temperature distribution features), and deep networks focus on high-level semantic features (such as fault mode features and system state features).

[0182] In practical implementation, shallow layers employ convolutional operations with small receptive fields to capture local variation features; middle layers use dilated convolutions or convolutions with larger receptive fields to expand the feature perception range; and deep layers introduce a self-attention mechanism to establish long-distance dependencies and capture global patterns. To improve the robustness of feature extraction, an adaptive feature selection mechanism is introduced, which can dynamically adjust the focus of feature extraction according to the data characteristics under different working conditions. Specifically, a learnable channel attention module is used to assign higher weights to key parameters under different working conditions, improving the model's sensitivity to key information.

[0183] S5.2: Based on the feature data, perform multi-scale feature fusion to generate the fused feature.

[0184] In S5.2, multi-scale feature fusion is implemented.

[0185] This application presents an innovative feature fusion architecture, comprising three key components: a feature alignment module (for dimensionality matching of features at different scales), an attention fusion module (for learning the importance weights of different features), and a feature enhancement module (for strengthening the expression of key features).

[0186] In terms of feature alignment, adaptive pooling and dimensionality transformation operations are used to map features of different scales to a unified representation space. For attention fusion, a multi-head attention mechanism is designed to evaluate feature importance from different perspectives and generate fusion weights. For feature enhancement, residual connections and feature activation adjustments are introduced to enhance the expressive power of key features. A bidirectional feature propagation mechanism is employed during the fusion process, allowing high-level features to guide the extraction of low-level features, while low-level features can also supplement the detailed information of high-level features. Furthermore, a feature filtering mechanism based on the information bottleneck theory is implemented, which removes redundant features and retains the most discriminative feature combinations through mutual information calculation and threshold filtering.

[0187] S5.3: Based on the fusion features, construct a fault classifier and generate the fault identification result, wherein the fault identification result includes fault location information, fault type information and fault severity assessment information.

[0188] In S5.3, a fault classifier is constructed based on the fused features. This embodiment of the application adopts a hierarchical classifier structure: first, a major fault category is determined (e.g., electrical fault, thermal fault, etc.), and then each major category is further subdivided into specific fault types (e.g., short circuit, open circuit, thermal runaway, etc.).

[0189] The classifier incorporates several advanced technologies: ensemble learning methods (such as Boosting or voting mechanisms) improve the stability and accuracy of classification; Bayesian uncertainty estimation is used to assess the credibility of classification results, providing confidence intervals for high-risk decisions; and adaptive threshold adjustment technology dynamically sets classification thresholds based on the characteristics and importance of different fault types. Simultaneously, for fault location identification, a gradient-based class activation mapping-based localization technique is implemented to generate heatmaps that visually display the areas where faults may occur. For fault severity assessment, a multi-index comprehensive scoring model is designed to quantitatively evaluate the severity level of faults based on factors such as the fault's impact range, development speed, and system criticality, providing a basis for subsequent early warning and handling.

[0190] After determining the optimal neural network architecture, this embodiment utilizes this architecture to perform multi-level feature extraction on the energy storage system's operational data. Taking lithium battery pack over-discharge fault detection as an example, over-discharge faults often exhibit complex multi-parameter interaction patterns, requiring the model to capture feature information at different levels. In the shallow feature extraction process, this embodiment primarily focuses on basic time-domain features.

[0191] For example, using a 3×3 convolution kernel (32 filters, stride 1) to extract local fluctuation features from the original voltage curve successfully captured the small voltage drop characteristics at the initial stage of over-discharge. Simultaneously, using a 1×3 convolution kernel (24 filters, stride 1) to extract short-term variation patterns from the current data identified abnormal fluctuations in the discharge current. These shallow features directly reflect the fundamental changes in battery parameters, laying the foundation for subsequent analysis. In the extraction of mid-level features, this embodiment uses operations with a larger receptive field. For example, using a 5×5 dilated convolution (64 filters, expansion rate 2) to extract combined voltage-current features revealed abnormal correlation patterns between voltage and current during over-discharge; using a bidirectional LSTM unit (hidden layer dimension 128) to process temperature sequence data captured the long-range dependence of temperature changes. These mid-level features reflect the interaction relationships between different parameters and evolution patterns at medium time scales. In deep feature extraction, a self-attention mechanism (8-head attention, hidden layer dimension of 256) was introduced to establish long-distance dependencies and successfully capture the state evolution features of the entire over-discharge process, including the complete evolution trajectory from the initial voltage drop to the final battery failure.

[0192] To improve the adaptability of feature extraction, this application implements an adaptive feature selection mechanism. Through a learnable channel attention module (compression ratio of 16), the model automatically adjusts the focus on various parameters based on the characteristics of different battery packs. For example, for aged battery packs, the model automatically enhances its focus on resistance characteristics; while for new battery packs, it places greater emphasis on voltage and temperature characteristics. This multi-level feature extraction method enables the model to comprehensively grasp various manifestations of over-discharge faults, significantly improving the accuracy and sensitivity of fault detection.

[0193] After acquiring multi-level features, the embodiments of this application implement a multi-scale feature fusion mechanism to integrate information at different levels of abstraction.

[0194] Taking insulation fault identification in energy storage power stations as an example, insulation faults usually involve complex interactions of electrical and thermodynamic characteristics.

[0195] First, in the feature alignment stage, the challenge faced by this embodiment is the dimensionality mismatch between different levels of features—shallow features have a dimension of [64, 128] (64 channels, 128 time steps), mid-level features have a dimension of [96, 64], and deep features have a dimension of [128, 32]. This embodiment unifies the time dimension through adaptive pooling and adjusts the channel dimension through 1×1 convolution, mapping all features to a unified [80, 64] representation space. For example, applying average pooling with a stride of 2 to shallow features reduces the time dimension from 128 to 64; applying upsampling to deep features expands the time dimension from 32 to 64. In the attention fusion stage, this embodiment designs a multi-head cross-attention mechanism (4 heads, temperature parameter of 0.5) to evaluate the importance of different features. This mechanism can automatically identify key feature combinations under different fault scenarios. For example, when an abnormal temperature rise is detected accompanied by a decrease in resistance, the model automatically increases the weight of electrical insulation features; while when local overheating is detected but electrical parameters are normal, the weight of thermal features is increased.

[0196] Specifically, the fusion process generates an attention weight matrix. For a given insulation fault sample, the shallow resistance feature receives a weight of 0.42, the mid-level temperature feature receives a weight of 0.35, and the deep global state feature receives a weight of 0.23. In the feature enhancement stage, this embodiment enhances the expressive power of key features through residual connections and feature activation adjustments. For example, the nonlinear expression of features is enhanced using the GELU activation function, and the importance distribution of feature channels is further adjusted through channel recalibration (squeeze-and-excitation module, reduction ratio of 8). Furthermore, this embodiment also implements a feature filtering mechanism based on the information bottleneck theory. By calculating the mutual information between each feature and the fault label, the top 85% of features with the highest mutual information are selected, effectively removing redundant features. The final fused feature dimension is [80, 64], containing comprehensive information from changes in microscopic electrical parameters to changes in macroscopic system state, greatly improving the accuracy and reliability of insulation fault identification.

[0197] Based on the fused features, this application embodiment constructs a hierarchical fault classifier. Taking the comprehensive fault diagnosis task of an energy storage power station as an example, this task needs to simultaneously identify multiple possible fault types and their severity.

[0198] This application employs a two-stage classification strategy: first, a major fault category is determined, and then each major category is further subdivided into specific fault types. In the major category determination stage, this application uses a three-layer fully connected network (layer dimensions of 512, 256, and 4, with dropout rates of 0.3, 0.2, and 0.1), mapping the fused features to four main fault categories: electrical faults, thermal faults, mechanical faults, and normal state. The classifier outputs the probability distribution for each major category; for example, the output for a test sample is [electrical: 0.85, thermal: 0.12, mechanical: 0.02, normal: 0.01], indicating that it is likely an electrical fault. In the subdivision stage, for the electrical fault category, this application constructs a dedicated classifier (a two-layer fully connected network with layer dimensions of 256 and 6), further subdividing it into six specific fault types: short circuit, open circuit, poor contact, overcharge, over-discharge, and insulation degradation. For this test sample, the subdivision classifier output is [Short circuit: 0.08, Open circuit: 0.72, Poor contact: 0.15, Overcharge: 0.03, Over-discharge: 0.01, Insulation degradation: 0.01], indicating that this is most likely an open circuit fault.

[0199] The classifier incorporates several advanced technologies: This embodiment employs an ensemble learning method based on Gradient Boosting, training five basic classifiers. The final result is obtained through weighted voting (weights of 0.25, 0.2, 0.2, 0.2, and 0.15), significantly improving classification stability. Bayesian uncertainty estimation (implemented via Monte Carlo Dropout with 50 sampling iterations) provides a 95% confidence interval for each classification result; for example, the confidence interval for predicting open-circuit faults is [0.65, 0.78]. Adaptive threshold adjustment is implemented for different fault types; for example, the identification threshold is reduced to 0.6 for high-risk short-circuit faults, while maintaining the standard 0.5 threshold for low-risk poor contact. For fault localization, this embodiment uses Grad-CAM-based heatmap technology to visually display the possible locations of faults.

[0200] Taking an open-circuit fault in a battery pack as an example, the heat map accurately highlights the connection point of the third battery module, helping maintenance personnel quickly locate the source of the problem. For fault severity assessment, this application's embodiments design a scoring model based on multiple indicators, comprehensively considering factors such as the fault's impact range (individual / module / system level), development speed (chronic / acute), and system criticality (core / non-core), quantifying fault severity into a 1-10 point system, providing clear priority guidance for subsequent early warning and handling.

[0201] S6 specifically includes:

[0202] S6.1: Input the fault identification results into the fault development trend analysis model to obtain the trend prediction results.

[0203] In S6.1, this embodiment establishes a fault development trend analysis model. This model employs a time-series prediction framework, combining various advanced time-series analysis techniques: a Long Short-Term Memory (LSTM) network is used to capture long-term dependencies, a time attention mechanism is used to identify key time points, and a conditional variational autoencoder is used to model the uncertainty of fault evolution. Specifically, the LSTM network processes historical data sequences and learns the temporal patterns of parameter changes; the time attention layer automatically identifies the most important historical moments for prediction, enhancing sensitivity to key points; the variational autoencoder represents the model's prediction of future uncertainty through probability distribution, helping to assess risk levels. Furthermore, the model integrates a historical fault case library, using similar case matching to assist in trend prediction. A dynamic time window mechanism is employed, which adaptively adjusts the prediction interval according to the development characteristics of different types of faults: short-window, high-frequency prediction is used for rapidly developing faults, while long-window prediction is used for slowly developing faults. Simultaneously, prediction confidence assessment is implemented by simulating multiple prediction samples using Monte Carlo simulation, statistically analyzing the distribution of predicted values, and quantifying the reliability of the prediction results.

[0204] S6.2: Based on the trend prediction results, design multi-level early warning thresholds to achieve graded early warning and obtain early warning levels and corresponding handling suggestions.

[0205] In S6.2, the embodiments of this application design a multi-level early warning threshold system.

[0206] The system establishes a four-level early warning system (normal, alert, warning, and danger), with each level having its own unique threshold definition mechanism. The threshold setting comprehensively considers multiple factors: historical statistical data (statistical analysis based on a large amount of historical operating data), expert experience (knowledge rules from domain experts), equipment specifications (safety boundaries based on equipment technical parameters), and dynamic adjustment factors (considering real-time factors such as environment and load).

[0207] This embodiment employs fuzzy logic to handle threshold boundaries, setting a membership function for each warning level to avoid frequent switching of warning levels due to slight parameter fluctuations. Simultaneously, a threshold self-learning mechanism is implemented, dynamically optimizing threshold settings based on newly added operational data and fault cases: the system periodically evaluates the warning effect, analyzes false alarms and missed alarms, and automatically fine-tunes threshold parameters, ensuring the warning system continuously improves with data accumulation. For each warning level, the system also provides corresponding handling suggestions, including operational guidance such as continuing monitoring, scheduling inspections, reducing load, and emergency shutdown.

[0208] S6.3: Based on the warning level and the processing suggestion, construct the warning decision rule and generate the warning information.

[0209] The decision-making rule system adopts a hierarchical structure: a basic rule layer (handling single parameter exceeding limits), a combined rule layer (handling the interaction of multiple parameters), and a scenario rule layer (handling special cases under specific operational scenarios). The rule construction process integrates multiple technologies: decision trees are used to build the basic logical framework, Bayesian networks are used to handle uncertainty, and rule mining algorithms are used to discover new rule patterns from historical data. The system places particular emphasis on rule interpretability; each rule is accompanied by a trigger reason explanation, facilitating operators' understanding and verification of the rationality of the warnings. Simultaneously, a rule conflict detection and priority management mechanism is implemented. When multiple rules are triggered simultaneously and give different warning levels, the system will make a consistent warning decision based on a preset priority strategy (such as taking the highest level, weighted average, etc.). Finally, the system generates a complete warning information package, including the warning level, detailed diagnosis, recommended measures, and emergency plans, which are intuitively displayed through a visual interface. It also supports intelligent push functionality to promptly deliver warning information to relevant responsible persons.

[0210] In the first step of building a safety early warning model, this application embodiment establishes a fault development trend analysis model.

[0211] Taking the prediction of battery thermal runaway in energy storage power stations as an example, thermal runaway is a highly dangerous fault, and it is crucial to predict its development trend as early as possible. The trend analysis model designed in this application adopts a variety of advanced time series analysis techniques.

[0212] First, this embodiment uses a Long Short-Term Memory (LSTM) network to capture long-term dependencies in temperature changes. The LSTM network consists of two layers, each with 128 hidden units, and a time step of 180 (corresponding to 30 minutes of monitoring data, with a sampling frequency of 6 seconds / sample). This allows the model to learn subtle patterns of temperature changes in the early stages of thermal runaway. For example, in one test case, the LSTM successfully captured an abnormal rate of temperature rise within the normal range (the rate of rise suddenly accelerated from 0.02°C / min to 0.08°C / min), which is an important indicator of early thermal runaway.

[0213] Secondly, this application's embodiments introduce a time attention mechanism to automatically identify the most important historical moments for prediction. In the same test case, the attention mechanism assigns 60% weight to data from the past 3 minutes, reflecting that recent temperature changes are more critical to the prediction. Furthermore, this application's embodiments also use a conditional variational autoencoder (with a latent space dimension of 32) to model the uncertainty of fault evolution, generating a probability distribution prediction of thermal runaway development, rather than just a single-point prediction. For example, the model predicts that within the next 15 minutes, there is an 85% probability that the battery pack temperature will exceed the safety threshold (60°C).

[0214] The model also integrates a historical failure case database, using the K-nearest neighbor algorithm (K=5) to match similar cases to assist in prediction. In a practical application, the system matched five historically similar thermal runaway cases, with an average evolution time of 20 minutes (from the onset of the anomaly to complete thermal runaway). This information further adjusted the current prediction. To adapt to the characteristics of different types of failures, this application implements a dynamic time window mechanism. For rapidly developing thermal runaway failures, the system automatically adjusts the prediction window to a 5-minute interval, continuously predicting the next 30 minutes; while for slowly developing capacity decay problems, the prediction window is adjusted to the hourly or daily level. Finally, to evaluate the reliability of the prediction, this application implements a Monte Carlo simulation (performing 100 prediction samples), and the standard deviation of the predicted temperature after 20 minutes is 2.3℃, indicating that the prediction has a high confidence level.

[0215] Based on the trend prediction results, this application embodiment designs a multi-level early warning threshold system to achieve graded early warning.

[0216] Taking voltage anomaly detection in an energy storage power station as an example, this application establishes a four-level early warning system. The normal level (green) indicates that the system is operating normally and requires no intervention; the attention level (yellow) indicates a minor anomaly requiring close monitoring; the warning level (orange) indicates a significant anomaly requiring preventative measures; and the danger level (red) indicates a serious anomaly requiring emergency handling. Each early warning level has its unique threshold definition mechanism, comprehensively considering multiple factors.

[0217] For example, regarding abnormal voltage in a single battery cell, this embodiment first determines an initial threshold based on historical statistical data: analyzing operational data from the past two years, it was found that the standard deviation of voltage fluctuation under normal operating conditions is 0.05V. Based on this, the initial threshold for the alert level is set to the mean ± 3σ (±0.15V). These thresholds are then adjusted based on expert experience: considering that voltage fluctuations naturally increase in low-temperature environments, the threshold range is automatically expanded by 15% in environments below 5°C. Simultaneously, the absolute threshold for the danger level is determined according to the equipment specifications (safety boundaries provided by the battery manufacturer). For a certain model of ternary lithium battery, a voltage below 2.5V or above 4.25V is considered a danger level. Furthermore, this embodiment also designs a dynamic adjustment factor to consider the impact of real-time factors. For example, when the battery is in fast charging mode, the allowable voltage fluctuation range is expanded by 20%; while when the system is in a long-term quiescent state, the threshold is tightened by 10%, because the voltage should be more stable at this time. To avoid frequent switching of warning levels due to slight parameter fluctuations, this embodiment uses a fuzzy logic method to handle threshold boundaries. For example, when the voltage deviation is 0.14V, it simultaneously has 90% "normal" membership and 10% "attention" membership. Only when the deviation continues to increase and the "attention" membership exceeds 50% will an attention-level warning be triggered. The system also implements a threshold self-learning mechanism, dynamically optimizing the threshold based on new data. In one practical case, the system discovered that the normal fluctuation range of a specific batch of batteries was larger than the preset value. After three months of adaptive learning, it automatically adjusted the attention-level threshold for that batch of batteries to ±0.18V, significantly reducing the false alarm rate. For each warning level, the system also provides corresponding handling suggestions, such as "increasing the monitoring frequency" for attention level, "reducing the charge / discharge rate" for warning level, and "immediately disconnecting and isolating the problematic module" for danger level.

[0218] Based on multi-level early warning thresholds, this application embodiment constructs an early warning decision rule system. Taking the comprehensive safety early warning of an energy storage power station as an example, this application embodiment designs a hierarchical rule structure. At the basic rule layer, it handles cases where a single parameter exceeds its limit.

[0219] For example, a basic rule might state: "If the temperature of a single cell exceeds 55°C, a warning level alert is triggered." Such rules are direct, clear, and easy to understand and enforce.

[0220] In the implementation of this application's embodiments, approximately 50 basic rules are established, covering various abnormal situations of key parameters such as voltage, current, temperature, and SOC. At the combined rule layer, the interaction of multiple parameters is handled. For example, one combined rule stipulates: "If the battery temperature rise rate exceeds 0.5℃ / minute and the voltage drop rate simultaneously exceeds 0.01V / minute, a danger level warning is triggered." Such rules can capture the interaction between parameters, improving the accuracy of warnings. In a real-world case, the system detected that although temperature and voltage, individually, were within the warning level threshold, their abnormal coordinated change pattern conformed to the early characteristics of thermal runaway, successfully issuing a danger level warning 15 minutes in advance. At the scenario rule layer, special situations under specific operating scenarios are handled.

[0221] For example, "In fast charging mode, if the charging current fluctuates by more than 5% for more than 3 minutes, a warning at the attention level will be triggered." This type of rule takes into account the specific characteristics of different operating modes and avoids unnecessary false alarms.

[0222] The rule construction process integrates multiple technologies: In this embodiment, a decision tree (depth of 4, minimum leaf node sample number of 50) is used to construct the basic logical framework; a Bayesian network is used to process the conditional probability relationship between parameters; and a rule mining algorithm (support threshold of 0.01, confidence threshold of 0.8) is used to discover new rule patterns from historical data, such as the phenomenon of "abnormal shortening of voltage plateau region" that appears before most thermal runaway cases, based on which new early warning rules are created.

[0223] The system places particular emphasis on the interpretability of the rules, with each rule accompanied by a triggering reason explanation. For example, when the "voltage-temperature anomaly covariance rule" is triggered, the system will simultaneously display: "Warning reason: A micro-short circuit may occur inside the battery, leading to a local temperature increase and loss of active materials." This allows operators to understand the physical mechanism behind the warning. Furthermore, this application embodiment implements a rule conflict detection and priority management mechanism. For instance, when the temperature rule triggers a warning level but the voltage rule triggers a danger level, the system will ultimately output a danger level warning based on a preset "highest level" strategy.

[0224] Ultimately, the system generates a complete early warning information package, including the warning level, detailed diagnosis ("A micro-short circuit was detected in module 3, cell 5, causing an abnormal increase in local temperature"), recommended measures ("It is recommended to immediately disconnect module 3 and perform offline testing"), and emergency plan (providing a link to the standard operating procedure). For different user roles, the system also supports intelligent push notifications: operators receive concise operating instructions via on-site displays and mobile apps, while technical supervisors receive a complete report containing detailed diagnostic data via email.

[0225] like Figure 6As shown in the illustration, this application provides an apparatus for fault identification and safety early warning in energy storage power stations, comprising: a data preprocessing module, a search space construction module, a Supernet construction module, an architecture optimization module, a fault identification module, and a safety early warning module. The fault identification module includes a feature extraction unit, a feature fusion unit, and a fault classification unit. The connections and data flow between the modules are consistent with the steps in the method embodiment, achieving fully automated processing of fault identification and safety early warning in energy storage power stations.

[0226] This application provides a system for fault identification and safety early warning in an energy storage power station. The system includes the aforementioned device, as well as a data acquisition unit, a data storage unit, and a display unit. The data acquisition unit is responsible for acquiring multi-source sensor data from the energy storage system; the data storage unit stores raw data, processing results, and historical records; and the display unit presents fault identification results and early warning information graphically, supporting multi-dimensional data visualization and interactive queries.

[0227] This application provides a schematic diagram of a deployment scenario for a fault identification and safety early warning system for an energy storage power station. The system can be flexibly deployed at different levels of the energy storage power station: on the one hand, it can be deployed in edge computing devices, directly connecting to the energy storage battery management system to achieve localized real-time fault identification; on the other hand, it can be deployed in a cloud server, aggregating data from multiple energy storage sites for global optimization and analysis. The system supports a distributed architecture, achieving rapid response through a lightweight edge model, while providing in-depth analysis and model update services through a cloud model, realizing intelligent fault identification and early warning through edge-cloud collaboration.

[0228] This application achieves efficient and accurate fault identification and early warning for energy storage power stations through a data-aware neural architecture search method and an accelerated evaluation model. It addresses the technical problems of insufficient model generalization ability and low deployment efficiency in resource-constrained environments using traditional methods. Based on actual test data, this application improves fault identification accuracy by 15%-20% and increases early warning lead time by 30%-50% compared to traditional methods, while reducing computational resource consumption by 60%, providing strong technical support for the safe and stable operation of energy storage power stations.

[0229] The above description is merely a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural transformations made based on the inventive concept of this application and the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the patent protection scope of this application.

[0230] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for fault identification and safety warning of energy storage power station, characterized in that, The method comprises the following steps: obtaining operation data of a storage system, performing data cleaning and standardization processing on the operation data to obtain a standardized feature data set; based on the standardized feature data set, constructing a search space model containing candidate operations and connection modes; based on the search space model, constructing a Supernet network sharing weights to generate an acceleration evaluation model; using the acceleration evaluation model to perform data-aware architecture evaluation and optimization to obtain an optimal neural network architecture; using the optimal neural network architecture to extract and analyze features of the storage system operation data to obtain a fault identification result; based on the fault identification result, establishing a safety warning model to generate warning information.

2. The method of claim 1, wherein, The method comprises the following steps: obtaining voltage, current, temperature and state of charge data of the storage system, and performing outlier rejection and missing value processing on the data to obtain cleaned data; performing normalization processing on the cleaned data to unify different dimensional data to the same scale to obtain standardized data; based on the standardized data, constructing a time series feature data set containing normal operation data and fault data as the standardized feature data set.

3. The method of claim 1, wherein, The method comprises the following steps: selecting basic operations including 3x3 convolution, 5x5 convolution, maximum pooling and self-attention mechanism, analyzing the standardized feature data set to obtain basic operator definitions; according to the basic operator definitions, designing a skip connection with bidirectional information flow and a dense connection mode with multi-layer feature fusion to obtain a connection strategy; applying the basic operator definitions to the network topology structure set by the connection strategy to form a hierarchical search space with multiple search dimensions to obtain a search space model.

4. The method of claim 1, wherein, Based on the search space model, a Supernet network sharing weights is constructed to generate an acceleration evaluation model, which comprises the following steps: applying a weight sharing strategy to the search space model to construct a subnetwork parameter reuse mechanism to obtain a weight sharing scheme; according to the weight sharing scheme, implementing a temperature-controlled progressive architecture sampling to obtain a training framework; using the training framework, training a performance prediction model containing an accuracy predictor and a delay predictor to generate an acceleration evaluation model.

5. The method of claim 1, wherein, Using the acceleration evaluation model to perform data-aware architecture evaluation and optimization to obtain an optimal neural network architecture, which comprises the following steps: using the acceleration evaluation model to perform architecture performance and resource consumption evaluation to obtain the architecture evaluation indicators; based on the architecture evaluation indicators, implementing multi-objective optimization to balance model performance and resource constraints to output the optimized architecture; based on the optimized architecture, performing verification testing to output the optimal neural network architecture.

6. The method of claim 1, wherein, Using the optimal neural network architecture to extract and analyze features of the storage system operation data to obtain a fault identification result, which comprises the following steps: The multi-level features of the energy storage system operation data are extracted by using the optimal neural network architecture, and the feature data is obtained; Based on the feature data, multi-scale feature fusion is performed to generate the fusion features; Based on the fusion features, a fault classifier is constructed to generate the fault identification result, wherein the fault identification result includes fault occurrence position information, fault type information and fault severity assessment information.

7. The method according to any one of claims 1 to 6, characterized in that, Based on the fault identification result, a safety warning model is established to generate warning information, including: The fault identification result is input into a fault development trend analysis model to obtain a trend prediction result; According to the trend prediction result, a multi-level warning threshold is designed to realize hierarchical warning, and a warning level and a corresponding processing suggestion are obtained; Based on the warning level and the processing suggestion, a warning decision rule is constructed to generate the warning information.

8. A device for energy storage power station fault identification and safety warning, characterized in that, Including: A data preprocessing module is configured to obtain operation data of an energy storage system, perform data cleaning and standardization processing on the operation data, and obtain a standardized feature data set; A search space construction module is configured to construct a search space model containing candidate operations and connection modes based on the standardized feature data set; A Supernet construction module is configured to construct a Supernet network sharing weights based on the search space model, and generate an acceleration evaluation model; An architecture optimization module is configured to perform data-aware architecture evaluation and optimization by using the acceleration evaluation model, and obtain an optimal neural network architecture; A fault identification module is configured to extract and analyze features of the energy storage system operation data by using the optimal neural network architecture, and obtain a fault identification result; A safety warning module is configured to establish a safety warning model based on the fault identification result, and generate warning information.

9. The apparatus of claim 8, wherein, The fault identification module includes: A feature extraction unit is configured to extract multi-level features by using the optimal neural network architecture, and obtain the feature data; A feature fusion unit is configured to perform multi-scale feature fusion based on the feature data, and output the fusion features; A fault classification unit is configured to construct a fault classifier based on the fusion features, and output the fault identification result.

10. A system for energy storage power station fault identification and safety warning, characterized in that, Including: The apparatus of claim 8 or 9; A data acquisition unit is configured to acquire operation data of an energy storage system; A data storage unit is configured to store the operation data, the fault identification result and the warning information; A display unit is configured to display the fault identification result and the warning information.

Images