An artificial intelligence-based energy storage system fault prediction and diagnosis method

Abstract

This invention discloses an artificial intelligence-based method for fault prediction and diagnosis of energy storage systems, comprising the following steps: collecting operational data of the energy storage system; preprocessing the operational data to generate a preprocessed multidimensional feature sequence; performing feature representation transformation on the multidimensional feature sequence to extract low-dimensional latent feature representations; constructing a symbolic regression network; performing symbolic expression search, applying dynamic weight penalties during the search process, and outputting a target expression; predicting the future operating state of the energy storage system and generating fault warning information; analyzing the contribution of each variable and operator in the target expression, determining the key parameters leading to the fault, and identifying the corresponding fault category. This invention achieves high-precision prediction and interpretable identification of energy storage system faults by introducing energy storage mechanism constraint operators and a game-theoretic adaptive penalty mechanism driven by residual distribution.

Description

Technical Field

[0001] This invention relates to the field of energy storage technology, and in particular to an artificial intelligence-based method for predicting and diagnosing faults in energy storage systems. Background Technology

[0002] With the continuous growth of new energy power generation, energy storage systems are increasingly being used in scenarios such as grid peak shaving and frequency regulation, renewable energy consumption, and emergency backup. Energy storage systems typically consist of large-scale battery clusters, power conversion devices, and cooling and monitoring equipment. Their operating environment is complex, and load dynamics fluctuate significantly, making them highly susceptible to faults such as overheating, capacity decay, power anomalies, and insufficient heat dissipation. Once a fault occurs, it may lead to a decrease in the operating efficiency of the energy storage system or even cause safety accidents. Therefore, how to accurately predict the operating status of energy storage systems and diagnose faults has become a critical issue that urgently needs to be addressed.

[0003] In existing technologies, fault prediction for energy storage systems largely relies on methods based on statistical modeling or traditional machine learning, such as threshold-triggered alarms, support vector machine classification, or long short-term memory network prediction. While these methods can capture trends in operational data to some extent, they suffer from two prominent problems: First, the prediction results are often "black box" numerical outputs, lacking interpretability and failing to clearly reveal the physical mechanism of the fault; second, most common methods employ fixed thresholds or single-feature discrimination, which are insufficiently sensitive to early latent anomalies, easily leading to missed or false alarms.

[0004] Therefore, how to provide an artificial intelligence-based method for fault prediction and diagnosis of energy storage systems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an artificial intelligence-based method for fault prediction and diagnosis of energy storage systems. This invention introduces energy storage mechanism constraint operators into a symbolic regression network and combines a dynamic weight penalty mechanism driven by residual distribution and adaptive game theory. This ensures that the expression search results conform to operational data patterns and satisfy physical constraints such as energy conservation, capacity decay, and temperature rise boundaries. Simultaneously, by replacing the traditional threshold triggering method with symbolic substructure pattern matching, interpretable diagnosis of faults such as overheating, capacity decay, power anomalies, and insufficient heat dissipation is achieved. Therefore, this invention can identify potential latent anomalies in energy storage systems in advance, reduce the risk of false alarms and missed alarms, improve the reliability of prediction results and the interpretability of diagnosis, and has significant engineering application value.

[0006] According to an embodiment of the present invention, a method for fault prediction and diagnosis of an energy storage system based on artificial intelligence includes the following steps:

[0007] Collect operational data from energy storage systems;

[0008] The running data is preprocessed to generate a preprocessed multidimensional feature sequence;

[0009] The multidimensional feature sequence is subjected to feature representation transformation to extract low-dimensional latent feature representations;

[0010] Based on the low-dimensional latent feature representation, a symbolic regression network is constructed, wherein the operator set of the symbolic regression network includes conventional mathematical operators and energy storage mechanism constraint operators;

[0011] Symbolic expression search is performed based on the symbolic regression network. During the search process, dynamic weight penalties are applied to candidate expressions that do not meet the energy conservation constraint, capacity decay constraint, or temperature rise boundary constraint, and the target expression is output.

[0012] The future operating status of the energy storage system is predicted based on the target expression, and a fault warning message is generated when the prediction result exceeds a preset threshold.

[0013] The contribution of each variable and operator in the target expression is analyzed to determine the key parameters that cause the fault and to identify the corresponding fault category.

[0014] Furthermore, the operational data includes:

[0015] At the battery cluster level, the terminal voltage, surface temperature, state of charge of individual cells, and overall charging and discharging current of the battery cluster are collected.

[0016] On the electrical side, the input power, output power, and power factor parameters of the energy storage system are collected.

[0017] On the thermal management side, the airflow speed of the cooling fan, the temperature of the coolant, and the temperature distribution of the heat sink are collected.

[0018] On the environmental side, data on ambient temperature, humidity, and external power fluctuations are collected.

[0019] The data collection adopts a time synchronization mechanism, which uses a global clock to uniformly timestamp the data collected by each sensor.

[0020] Furthermore, the preprocessing includes:

[0021] All types of operational data are time-aligned according to a unified sampling period and timestamp;

[0022] Outlier removal is performed on data with noise interference, and the outlier removal adopts a combination of statistical threshold detection and sliding window mean correction.

[0023] Imputation processing is performed on missing data, which includes two methods: temporal neighborhood interpolation and regression prediction based on relevant parameters.

[0024] The parameters of different dimensions are normalized by a combination of interval normalization and Z-score standardization.

[0025] Furthermore, the feature representation transformation includes:

[0026] The preprocessed multidimensional feature sequence is divided into several sub-sequences according to the time window;

[0027] Dimensionality reduction is performed on each subsequence, including extracting the main linear components using principal component analysis and extracting nonlinear latent features using an autoencoder.

[0028] Linear components are combined with nonlinear latent features to form a low-dimensional latent feature representation;

[0029] The low-dimensional latent features represent both the preservation of the temporal evolution characteristics of the operational data and the retention of the spatial correlation between parameters of the battery cluster, electrical side, thermal management side, and environmental side.

[0030] Furthermore, the construction of the symbolic regression network includes:

[0031] Define a symbolic regression expression search tree on the input space of low-dimensional latent feature representations;

[0032] The search tree includes a set of predefined operators, which includes conventional mathematical operators and energy storage mechanism constraint operators. The conventional mathematical operators include addition, subtraction, multiplication, division, exponential, and logarithmic operators. The energy storage mechanism constraint operators include energy conservation operators, capacity decay operators, and temperature rise threshold operators.

[0033] During the training process of the symbolic regression network, candidate expression nodes are combined and optimized using a neural network parameterization approach to form an interpretable symbolic structure.

[0034] The constraint-generated expressions can simultaneously adapt to the characteristics of operating data and the physical mechanisms of energy storage systems.

[0035] Furthermore, the symbolic expression search includes:

[0036] Candidate symbolic expressions are generated based on a preset set of operators, and the generation process adopts an expression evolution mechanism based on a genetic algorithm.

[0037] During the evolution process, the fitting error of the candidate expression is used as the fitness index, and a penalty term is set in combination with energy conservation constraints, capacity decay constraints and temperature rise boundary constraints.

[0038] The weight of the penalty term is dynamically adjusted with the number of iterations, so that expressions that do not meet the physical constraints are gradually eliminated.

[0039] After multiple rounds of evolution, a target expression that simultaneously satisfies the data fitting accuracy and physical constraints is obtained, and this target expression is used as the core result for modeling the operating state of the energy storage system.

[0040] Furthermore, the prediction of the future operating state of the energy storage system based on the target expression includes:

[0041] The objective expression is applied to the time series data for extrapolation to obtain the battery cluster temperature change curve, capacity decay trend and power fluctuation sequence at each predicted time in the future;

[0042] In the prediction results, when the rising rate of the temperature change curve exceeds a preset safety threshold, or the falling rate of the capacity decay trend exceeds a set threshold, or the amplitude of the power fluctuation sequence exceeds an allowable threshold, corresponding fault warning information is generated.

[0043] The fault warning information is output in the form of timestamp, triggering conditions and corresponding risk level, so that the energy storage system management platform can perform real-time alarm and subsequent scheduling.

[0044] Furthermore, the contributions of each variable and operator in the analytical objective expression include:

[0045] Perform a sensitivity analysis on the target expression to calculate the partial derivative of each variable with respect to the change in the predicted output, in order to determine the relative impact of different operating parameters on the prediction results;

[0046] The structural contribution evaluation is performed on the operator to identify the strength of the role of substructures, including energy conservation operators, capacity decay operators, and temperature rise threshold operators, in the overall expression.

[0047] Based on the sensitivity analysis and structural contribution assessment results, the key parameters leading to fault evolution are identified, and the fault category is determined according to the type of key parameters and the corresponding operator structure, including overheating fault, capacity decay fault, power converter abnormality, and insufficient heat dissipation fault.

[0048] Furthermore, the detailed rules for determining the fault category include:

[0049] The determination of the fault category does not use threshold triggering, but is based on symbol substructure pattern matching for identification;

[0050] The target expression is decomposed into substructures to extract symbolic substructures containing energy conservation operators, capacity decay operators, and temperature rise threshold operators;

[0051] The symbol substructure is matched with a pre-built fault mode library. When a substructure combination that matches the mode library appears, it is determined to be the corresponding fault category.

[0052] The overheating fault corresponds to a substructure that couples a temperature rise threshold operator with cooling-related variables;

[0053] The capacity decay fault corresponds to a substructure that couples the capacity decay operator with the loop count variable;

[0054] Power anomalies correspond to substructures where the energy conservation operator deviates.

[0055] The insufficient heat dissipation fault corresponds to the temperature rise threshold operator and the substructure of the imbalance of heat dissipation parameters.

[0056] Furthermore, the dynamic weight penalty mechanism includes:

[0057] During the symbolic expression search process, the residual distribution between the predicted values ​​of candidate expressions and the actual running data is calculated;

[0058] When the residual distribution shows a systematic shift on the boundary variables related to energy conservation, capacity decay, or temperature rise, the penalty weight is adaptively adjusted according to the characteristics of the residual distribution.

[0059] The competition among candidate expressions is modeled as a game process. Search resources are dynamically allocated through evolutionary game strategies. Expressions that conform to physical constraints gain higher fitness, while expressions that violate physical constraints are gradually eliminated.

[0060] The beneficial effects of this invention are:

[0061] The mechanism enables symbolic expression search to not only focus on the accuracy of data fitting, but also to gradually eliminate candidate expressions that violate the boundary conditions of energy conservation, capacity decay and temperature rise, ensuring that the target expression conforms to the physical laws of the energy storage system, thereby improving the reliability of the prediction results.

[0062] This invention decomposes the target expression into symbolic substructures and replaces the traditional threshold triggering mechanism with pattern matching. It can directly identify types of faults such as overheating faults, capacity decay faults, power anomalies, and insufficient heat dissipation faults based on the expression structure, achieving a unified approach from data prediction to mechanism diagnosis, and significantly improving the interpretability and precision of fault diagnosis.

[0063] This invention combines feature representation transformation of multi-source operational data with structural optimization of symbolic regression networks, enabling early identification of potential hidden anomalies in energy storage systems, reducing the risk of false alarms and missed alarms, and providing a reliable basis for safe operation and maintenance decisions of energy storage systems. It has broad engineering application value. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a flowchart of an artificial intelligence-based method for fault prediction and diagnosis of energy storage systems proposed in this invention;

[0066] Figure 2 This is a flowchart of the fault determination pattern matching process for an artificial intelligence-based energy storage system fault prediction and diagnosis method proposed in this invention.

[0067] Figure 3 This is a flowchart of the dynamic weighted penalty mechanism of an artificial intelligence-based energy storage system fault prediction and diagnosis method proposed in this invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figure 1 - Figure 3 A method for fault prediction and diagnosis of energy storage systems based on artificial intelligence includes the following steps:

[0070] Collect operational data from energy storage systems;

[0071] The running data is preprocessed to generate a preprocessed multidimensional feature sequence;

[0072] The multidimensional feature sequence is subjected to feature representation transformation to extract low-dimensional latent feature representations;

[0073] Based on the low-dimensional latent feature representation, a symbolic regression network is constructed, wherein the operator set of the symbolic regression network includes conventional mathematical operators and energy storage mechanism constraint operators;

[0074] Symbolic expression search is performed based on the symbolic regression network. During the search process, dynamic weight penalties are applied to candidate expressions that do not meet the energy conservation constraint, capacity decay constraint, or temperature rise boundary constraint, and the target expression is output.

[0075] The future operating status of the energy storage system is predicted based on the target expression, and a fault warning message is generated when the prediction result exceeds a preset threshold.

[0076] The contribution of each variable and operator in the target expression is analyzed to determine the key parameters that cause the fault and to identify the corresponding fault category.

[0077] In this embodiment, the operational data includes: at the battery cluster level, the terminal voltage, surface temperature, state of charge, and overall charging and discharging current of individual cells; on the electrical side, the input power, output power, and power factor parameters of the energy storage system; on the thermal management side, the fan speed, coolant temperature, and radiator temperature distribution; and on the environmental side, ambient temperature, humidity, and external power fluctuations. The acquisition of the operational data adopts a time synchronization mechanism, using a global clock to uniformly timestamp the data collected by each sensor.

[0078] At the battery cluster level, the terminal voltage of a single cell is acquired by a voltage sampling chip mounted on the battery terminal, with a sampling accuracy of ±1mV and a sampling frequency of 1Hz; the surface temperature of a single cell is acquired by a thermocouple sensor attached to the battery casing, with a sampling accuracy of ±0.1℃ and a sampling frequency of 1Hz; the state of charge of a single cell is calculated using the coulomb integration method, with an integration step of 1 second, and is corrected based on the open-circuit voltage after each charge and discharge cycle; the overall charge and discharge current of the battery cluster is acquired by a Hall current sensor, with a sampling frequency of 10Hz and a measurement range of 0–500A.

[0079] On the electrical side, the input power and output power are calculated by multiplying the sampled voltage and sampled current, with a calculation period of 1 second; the power factor parameter is calculated by measuring the phase difference between the voltage and current using a phase detection circuit, with a measurement accuracy of 0.01 and a calculation period of 1 second.

[0080] On the thermal management side, the cooling fan speed is collected by a hot-film wind speed sensor installed at the air outlet, with a sampling frequency of 1Hz and a measurement range of 0–20m / s; the coolant temperature is collected by a digital temperature sensor embedded in the cooling pipe, with a sampling frequency of 1Hz and a measurement range of -40–125℃; the heat sink temperature distribution is collected by an NTC thermistor array arranged at different locations on the heat sink, with a total of 8 sampling points and a sampling frequency of 1Hz.

[0081] On the environmental side, ambient temperature and humidity are collected by digital temperature and humidity sensors installed outside the energy storage system, with a sampling frequency of 1Hz and measurement ranges of -40–85℃ and 0–100%RH, respectively; external power fluctuations are collected by the power quality monitoring module, and the monitoring parameters include the effective value of voltage, voltage fluctuation amplitude and frequency offset, with a sampling frequency of 1Hz.

[0082] All acquired data is transmitted to the edge computing node via a CAN bus with a speed of 500kbps. The global clock is provided by the GPS timing module with a timing accuracy of ±1μs. All data acquisition controllers immediately add a unified timestamp to the sampled data after receiving the timing signal, with a timestamp accuracy of milliseconds.

[0083] After completing the data acquisition, the data is preprocessed to generate a preprocessed multidimensional feature sequence. The steps are as follows:

[0084] The unified sampling period is set to 1 second. After receiving the global clock timing signal, the sampling controller writes a millisecond-level timestamp for each record. The data management program sorts the data by timestamp and resamples the data from different sources to whole-second time points. For records that fall outside the whole second, linear interpolation is performed according to the time ratio of the adjacent valid records, and the result is written to the corresponding whole-second time point to ensure that each parameter corresponds one-to-one at the same time point.

[0085] Sliding windows are set for three key parameters: voltage, current, and temperature, with a window length of 60 seconds. Within each window, the system calculates the mean and standard deviation of the parameter. When a sampled value deviates from the window mean by more than three times the standard deviation, the sampled value is marked as an anomaly. For marked anomaly samples, the mean of their respective windows is used as a replacement, and the replacement mark is recorded for traceability.

[0086] When a single time point is missing, linear interpolation based on the time ratio of adjacent valid samples is used for imputation, and the resulting data is labeled "Interpolation Imputation". When the duration of consecutive missing data exceeds 5 seconds, a linear regression model pre-trained on historical data with frozen parameters is used for imputation: for voltage parameters, current and temperature from the same cluster are used as independent variables to output voltage imputation values; for current parameters, voltage and temperature are used as independent variables to output current imputation values; and for temperature parameters, current and cooling fan speed are used as independent variables to output temperature imputation values. The training data for the above models covers the most recent 30 days, and the coefficients are not adjusted during runtime after training. All imputation records are labeled "Model Imputation".

[0087] Parameters such as voltage, current, power, and temperature are linearly mapped to the 0-1 range. The mapping endpoints are fixed at the 1st and 99th percentiles of the parameter's most recent 30-day samples; values ​​outside these endpoints are truncated before mapping to prevent extreme values ​​from affecting the scale. For the power factor parameter, a dimensionless transformation is performed centered on the mean of the most recent 30-day samples and scaled by the sample standard deviation, outputting a relative value in units of "standard deviation". All transformed results maintain the field correspondence with the original quantities and are written to the "preprocessed feature sequence" with time as the primary key for direct reading in subsequent feature representation transformation steps.

[0088] After preprocessing, the multidimensional feature sequence is transformed into a feature representation, as follows:

[0089] Using a fixed 60-second window, the preprocessed multidimensional feature sequence is divided into several non-overlapping subsequences in chronological order. Each subsequence contains data from 60 time points, each corresponding to all parameters including voltage, current, temperature, power, power factor, cooling fan speed, coolant temperature, heat sink temperature, ambient temperature, ambient humidity, and power fluctuation.

[0090] For each subsequence, a feature matrix is ​​established, and principal component analysis is performed to reduce dimensionality along the dimensional direction. For continuous parameters such as voltage, current, temperature, and power, the top principal components with a cumulative variance contribution rate of 95% are extracted, and the results are written into a "linear component vector".

[0091] The subsequence is input into a pre-trained autoencoder network. The number of nodes in the input layer equals the total number of all parameter dimensions of the subsequence, the number of nodes in the hidden layer is fixed at one-quarter of the input layer, and the number of nodes in the output layer is the same as the input layer. The network adopts a fully connected structure, the activation function is ReLU, and the goal during training is to minimize the reconstruction error. The hidden layer vector output by the encoder is denoted as the "nonlinear latent feature".

[0092] The "linear component vector" and "nonlinear latent feature" corresponding to each subsequence are concatenated to generate a low-dimensional latent feature representation of fixed length. The concatenation order is: linear component first, nonlinear latent feature second. The generated low-dimensional latent feature representation is uniformly stored in matrix form, with row indices representing time window numbers and column indices representing feature dimension numbers.

[0093] Perform integrity checks on the generated low-dimensional latent feature representations. Each window must contain two types of fields: linear components and nonlinear latent features. Window with missing fields is marked as invalid and removed from subsequent modeling.

[0094] After the above steps, a low-dimensional latent feature representation is obtained. This representation contains both the temporal evolution information of the running data and the nonlinear interaction relationship between different physical quantities, ensuring that it can be used as input to the symbolic regression network.

[0095] After obtaining the low-dimensional latent feature representation, the symbolic regression network is constructed, including:

[0096] Construct a symbolic expression search tree in the input space. The root node of the search tree is the set of input variables. Each node is expanded to generate child nodes through operators. The maximum depth is fixed at six levels. The leaf nodes are numerical constants or input variables, and the intermediate nodes are operators.

[0097] A set of operators is defined, divided into conventional mathematical operators and energy storage mechanism constraint operators. Conventional mathematical operators include addition, subtraction, multiplication, division, exponentiation, and logarithms, used to describe general functional relationships. Energy storage mechanism constraint operators include energy conservation operators, capacity decay operators, and temperature rise threshold operators. The energy conservation operator receives input power, output power, and power loss, internally performs two difference operations, and outputs a power balance deviation value, used to limit candidate expressions from violating the law of energy conservation. The capacity decay operator receives the number of cycles, current capacity, and initial capacity, internally looks up the decay rate from a table and performs a subtraction operation, outputting a capacity residual, used to ensure that the prediction result conforms to the mechanism of gradual capacity decay with the number of cycles. The temperature rise threshold operator receives the current temperature, the previous temperature, the cooling fan speed, and the coolant temperature, internally first calculates the temperature rise rate, then retrieves the allowable temperature rise upper limit from the cooling capacity table and performs a difference operation, outputting a temperature rise deviation, used to limit the temperature rise rate from exceeding the cooling boundary. All the above operators are stored in the operator library in symbolic form. Each operator node contains a fixed number of input ports and one output port. The internal operation order is determined, and the output data is a floating-point number, which can be directly used as the intermediate or final result of the expression.

[0098] In the candidate expression generation process, a neural network is used to parameterize and control the operator nodes. Each operator node has a weight vector, which is output by the neural network and normalized to determine the selection probability of the child node. This ensures that the selection of operators is regulated by the network weights rather than being completely random during expression search.

[0099] Constant nodes are introduced into the symbolic regression network. The initial values ​​of the constant nodes are taken from a fixed range of values ​​and stored as floating-point numbers. During training, the values ​​are adjusted through gradient update methods so that the candidate expressions can better fit the running data.

[0100] As the search and training proceed, the candidate expressions gradually evolve, eventually forming a symbolic structure composed of input variables, constant nodes, conventional mathematical operators, and energy storage mechanism constraint operators. This symbolic structure is stored in the form of explicit expressions, recording operator paths and parameter values, providing input for subsequent symbolic expression search, prediction, and diagnosis.

[0101] After the symbolic regression network is constructed, symbolic expression search is performed based on the network.

[0102] Candidate expressions are generated under the constraints of the operator set and input variables. Each candidate expression is composed of input variables, constant nodes, conventional mathematical operators, and energy storage mechanism constraint operators, combined according to search tree rules. The initial population size is fixed at 500 candidate expressions, and the maximum depth of each candidate expression does not exceed six levels.

[0103] The fitting error of the candidate expression is calculated. The fitting error is based on the difference between the predicted output of the target variable and the candidate expression, and is calculated at each time point using a point-by-point comparison method. The final result is the average error of the expression over a time window. This error value serves as the basic indicator of the fitness of the candidate expression.

[0104] A dynamic penalty mechanism based on physical constraints is introduced into the fitness evaluation. For candidate expressions that do not satisfy the energy conservation constraint, the deviation between the difference between input power and output power and the power loss is calculated and the degree of deviation is recorded. For candidate expressions that do not satisfy the capacity decay constraint, the deviation between the predicted capacity and the decay model capacity is calculated and the degree of deviation is recorded. For candidate expressions that do not satisfy the temperature rise boundary constraint, the magnitude of the temperature rise rate exceeding the limit is calculated and the degree of deviation is recorded. The above deviations are weighted by residual distribution analysis and distributed to each candidate expression in a game-theoretic adaptive manner during the iteration process. Specifically, a competitive relationship is established among candidate expressions: the fitness value of candidate expressions that violate the constraints decreases, while the fitness value of candidate expressions that meet the constraints increases, thus forming a dynamic survival of the fittest during the evolution process.

[0105] A genetic algorithm is used to evolve candidate expressions. Genetic operations include crossover, mutation, and selection. Crossover involves exchanging operator nodes and subtree structures to ensure the generated expressions maintain structural rationality. Mutation randomly replaces operators or adjusts constant values ​​at expression nodes. Selection is based on fitness ranking, retaining the top 20% of expressions as the elite group, and selecting the remaining expressions from the remaining candidates using a roulette wheel selection process.

[0106] After 100 rounds of iteration, a target expression that simultaneously satisfies the data fitting accuracy and energy storage physical constraints is selected from the candidate expression set. This target expression is stored in symbolic form, containing a complete node structure and parameter configuration, and serves as the core basis for subsequent prediction and fault diagnosis.

[0107] After obtaining the target expression, the future operating state of the energy storage system is predicted based on this expression, and fault warning information is generated. The specific steps are as follows:

[0108] The objective expression is applied to the preprocessed feature sequence, and extrapolation calculations are performed in chronological order. The extrapolation time step is fixed at 1 second, and the prediction duration is set to the next 600 seconds. During the prediction process, the system reads the low-dimensional latent feature representation point by point as input, and generates battery cluster temperature sequence, capacity change sequence, and power fluctuation sequence through the objective expression. All prediction results are stored in a structured table with time index as the primary key and second by second.

[0109] Trend analysis is performed on the prediction results. The temperature series obtains the temperature rise rate by calculating the difference between two consecutive points; the capacity series obtains the capacity decay rate by calculating the capacity difference per minute; and the power series obtains the power fluctuation amplitude by calculating the peak difference within 5 seconds. These three indicators are compared with preset safety boundary parameters, which are fixedly stored in the system configuration file. The upper limit for the temperature rise rate is 0.5°C per second, the upper limit for the capacity decay rate is 0.05% per minute, and the upper limit for the power fluctuation amplitude is 10% of the rated power.

[0110] When any predicted indicator exceeds its corresponding boundary, the fault warning generation module is immediately triggered. The warning information includes three parts: timestamp, triggering condition, and risk level. The risk level is divided into three levels: if the predicted indicator exceeds the boundary value by less than 20%, it is classified as Level 1 risk; if it exceeds 20% but is less than 50%, it is classified as Level 2 risk; and if it exceeds 50%, it is classified as Level 3 risk.

[0111] Fault warning information is sent to the energy storage system management platform via a communication interface. Information transmission uses the MQTT protocol, with a message subject of "Fault / Warning" and a message payload including a JSON-formatted timestamp, indicator type, predicted value, boundary value, and risk level. Upon receiving the information, the management platform records it in its database and triggers the corresponding scheduling strategy.

[0112] After predicting future operating conditions, the target expression is parsed to identify key parameters and determine the fault category. The specific steps are as follows:

[0113] Perform sensitivity analysis on the target expression. The parsing program iterates through the input variables in the expression one by one, applying a ±1% perturbation to the target variable while keeping other variables constant, and recording the magnitude of the output change. The sensitivity value of each variable is obtained by calculating the output difference caused by the perturbation. The higher the sensitivity value, the stronger the influence of that variable on the prediction result. All sensitivity values ​​are sorted in descending order and stored in the diagnostic results table.

[0114] The structural contribution of operators in the target expression is evaluated. The parser starts from the root node of the expression and recursively identifies all substructures containing energy conservation operators, capacity decay operators, and temperature rise threshold operators. For each substructure, the difference between its independent output and the overall output is calculated, and this difference is used as the substructure's contribution. The contribution value is recorded in a substructure index table for comparison with preset fault modes.

[0115] Perform symbolic substructure pattern matching. The system compares the operator substructures extracted from the target expression with a pre-built fault mode library. The pattern library contains four types of patterns:

[0116] Overheating fault mode: includes a coupled substructure of temperature rise threshold operator and cooling fan speed variable;

[0117] Capacity decay failure mode: includes a coupled substructure of capacity decay operator and loop count variable;

[0118] Power anomaly mode: Substructures that contain energy conservation operators and whose outputs exhibit continuous deviations;

[0119] Insufficient heat dissipation mode: This includes a substructure where the temperature rise threshold operator and the heat sink temperature parameters are unbalanced. When one of the above modes is matched, the system determines it to be the corresponding fault category.

[0120] The diagnostic results are generated into standardized output, including fault category, key parameters, trigger substructures, and sensitivity ranking. The diagnostic results are written to a log file in JSON format and sent to the energy storage system management platform via the MQTT protocol for reference by operations and maintenance personnel and for subsequent control strategy invocation.

[0121] Example 1:

[0122] To verify the feasibility of this invention in practice, it was applied to a battery energy storage system with a rated power of 500kW and a capacity of 1MWh. The system consists of 1000 lithium iron phosphate cells, each with an initial capacity of 100Ah and a rated voltage of 3.2V. It adopts a hybrid thermal management method of air cooling and liquid cooling, and the ambient temperature range of the operating location is -10℃ to 40℃.

[0123] During the experimental period, the system ran for 30 days with a data sampling period of 1 second. The collected operational data, after preprocessing and feature representation transformation, was input into a symbolic regression network to obtain the target expression. The table below compares key data from a specific charge-discharge cycle with the predicted results:

[0124] Table 1 Comparison of Energy Storage System Operation Data and Prediction Results

[0125] Time (s) Input power (kW) Output power (kW) Average temperature of battery cluster (°C) Measured capacity (Ah) Predicted capacity (Ah) Power deviation (%) Temperature rise rate (°C / s) 0 250 245 28.5 99.8 99.7 2.0 0.00 300 260 253 29.8 98.9 98.8 2.7 0.004 600 270 262 31.5 97.9 97.7 3.0 0.006 900 280 270 34.2 96.8 96.5 3.6 0.009 1200 290 278 37.0 95.6 95.2 4.1 0.011 1500 300 286 40.1 94.3 93.8 4.7 0.014 1800 310 295 43.8 93.1 92.5 4.8 0.017

[0126] As can be seen from Table 1, during the entire charging and discharging process:

[0127] The difference between input power and output power is always kept within 5%, and the target expression predicted by the symbolic regression network includes an energy conservation operator to constrain the power balance, so that the predicted power deviation is kept below 5%.

[0128] As charging and discharging proceeded, the measured capacity gradually decreased from 99.8 Ah to 93.1 Ah, while the predicted capacity decreased from 99.7 Ah to 92.5 Ah. The deviation between the predicted and measured values ​​was less than 0.6 Ah, proving that the capacity decay operator in the target expression can accurately reflect the law of capacity change with the number of cycles.

[0129] Within 0–1800 seconds, the average temperature of the battery cluster rose from 28.5℃ to 43.8℃, with the temperature rise rate increasing from 0.004℃ / s to 0.017℃ / s. When the temperature rise rate exceeded the preset safety threshold of 0.01℃ / s, the system triggered a warning message and identified it as an "overheating risk." By analyzing the substructure in the objective expression that couples the temperature rise threshold operator with the cooling fan speed variable, it was determined that the key parameter causing overheating was insufficient fan speed.

[0130] After 1500 seconds, the contribution of the temperature rise threshold operator in the objective expression increased significantly, and the symbolic substructure pattern matched the "insufficient heat dissipation" pattern library entry. The system ultimately determined the fault category for this cycle to be "insufficient heat dissipation," and the diagnosis result was consistent with the subsequent performance degradation of the cooling fan observed by maintenance personnel.

[0131] This embodiment verifies that the symbolic regression network method proposed in this invention can not only make high-precision predictions of the operating status of energy storage systems, but also make interpretable diagnoses of the causes of faults by parsing the target expression structure, thus having practical engineering application value.

[0132] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for fault prediction and diagnosis of energy storage systems based on artificial intelligence, characterized in that, Includes the following steps: Collect operational data from energy storage systems; The running data is preprocessed to generate a preprocessed multidimensional feature sequence; The multidimensional feature sequence is subjected to feature representation transformation to extract low-dimensional latent feature representations; Based on the low-dimensional latent feature representation, a symbolic regression network is constructed, wherein the operator set of the symbolic regression network includes conventional mathematical operators and energy storage mechanism constraint operators; Symbolic expression search is performed based on the symbolic regression network. During the search process, dynamic weight penalties are applied to candidate expressions that do not meet the energy conservation constraint, capacity decay constraint, or temperature rise boundary constraint, and the target expression is output. The future operating status of the energy storage system is predicted based on the target expression, and a fault warning message is generated when the prediction result exceeds a preset threshold. The contribution of each variable and operator in the target expression is analyzed to determine the key parameters that cause the fault and to identify the corresponding fault category. The construction of the symbolic regression network includes: Define a symbolic regression expression search tree on the input space of low-dimensional latent feature representations; The search tree includes a set of predefined operators, which includes conventional mathematical operators and energy storage mechanism constraint operators. The conventional mathematical operators include addition, subtraction, multiplication, division, exponential, and logarithmic operators. The energy storage mechanism constraint operators include energy conservation operators, capacity decay operators, and temperature rise threshold operators. During the training process of the symbolic regression network, candidate expression nodes are combined and optimized using a neural network parameterization approach to form an interpretable symbolic structure, constraining the generated expressions to simultaneously adapt to the characteristics of operating data and the physical mechanism of the energy storage system.

2. The method for fault prediction and diagnosis of energy storage systems based on artificial intelligence according to claim 1, characterized in that, The operational data includes: At the battery cluster level, the terminal voltage, surface temperature, state of charge of individual cells, and overall charging and discharging current of the battery cluster are collected. On the electrical side, the input power, output power, and power factor parameters of the energy storage system are collected. On the thermal management side, the airflow speed of the cooling fan, the temperature of the coolant, and the temperature distribution of the heat sink are collected. On the environmental side, data on ambient temperature, humidity, and external power fluctuations are collected. The data collection adopts a time synchronization mechanism, which uses a global clock to uniformly timestamp the data collected by each sensor.

3. The method for fault prediction and diagnosis of energy storage systems based on artificial intelligence according to claim 1, characterized in that, The preprocessing includes: All types of operational data are time-aligned according to a unified sampling period and timestamp; Outlier removal is performed on data with noise interference, and the outlier removal adopts a combination of statistical threshold detection and sliding window mean correction. Imputation processing is performed on missing data, which includes two methods: temporal neighborhood interpolation and regression prediction based on relevant parameters. The parameters of different dimensions are normalized by a combination of interval normalization and Z-score standardization.

4. The method for fault prediction and diagnosis of energy storage systems based on artificial intelligence according to claim 1, characterized in that, The feature representation transformation includes: The preprocessed multidimensional feature sequence is divided into several sub-sequences according to the time window; Dimensionality reduction is performed on each subsequence, including extracting the main linear components using principal component analysis and extracting nonlinear latent features using an autoencoder. Linear components are combined with nonlinear latent features to form a low-dimensional latent feature representation; The low-dimensional latent features represent both the preservation of the temporal evolution characteristics of the operational data and the retention of the spatial correlation between parameters of the battery cluster, electrical side, thermal management side, and environmental side.

5. The method for fault prediction and diagnosis of energy storage systems based on artificial intelligence according to claim 1, characterized in that, The symbolic expression search includes: Candidate symbolic expressions are generated based on a preset set of operators, and the generation process adopts an expression evolution mechanism based on a genetic algorithm. During the evolution process, the fitting error of the candidate expression is used as the fitness index, and a penalty term is set in combination with energy conservation constraints, capacity decay constraints and temperature rise boundary constraints. The weight of the penalty term is dynamically adjusted with the number of iterations, so that expressions that do not meet the physical constraints are gradually eliminated. After multiple rounds of evolution, a target expression that simultaneously satisfies the data fitting accuracy and physical constraints is obtained, and this target expression is used as the core result of energy storage system operation state modeling.

6. The method for fault prediction and diagnosis of energy storage systems based on artificial intelligence according to claim 1, characterized in that, The prediction of the future operating state of the energy storage system based on the target expression includes: The objective expression is applied to the time series data for extrapolation to obtain the battery cluster temperature change curve, capacity decay trend and power fluctuation sequence at each predicted time in the future; In the prediction results, when the rising rate of the temperature change curve exceeds a preset safety threshold, or the falling rate of the capacity decay trend exceeds a set threshold, or the amplitude of the power fluctuation sequence exceeds an allowable threshold, corresponding fault warning information is generated. The fault warning information is output in the form of timestamp, triggering conditions and corresponding risk level, so that the energy storage system management platform can perform real-time alarm and subsequent scheduling.

7. The method for fault prediction and diagnosis of energy storage systems based on artificial intelligence according to claim 1, characterized in that, The contributions of each variable and operator in the analytical objective expression include: Perform a sensitivity analysis on the target expression to calculate the partial derivative of each variable with respect to the change in the predicted output, in order to determine the relative impact of different operating parameters on the prediction results; The structural contribution evaluation is performed on the operator to identify the strength of the role of substructures, including energy conservation operators, capacity decay operators, and temperature rise threshold operators, in the overall expression. Based on the sensitivity analysis and structural contribution assessment results, the key parameters leading to fault evolution are identified, and the fault category is determined according to the type of key parameters and the corresponding operator structure, including overheating fault, capacity decay fault, power converter abnormality, and insufficient heat dissipation fault.

8. The method for fault prediction and diagnosis of energy storage systems based on artificial intelligence according to claim 1, characterized in that, The detailed rules for determining the fault category include: The determination of the fault category does not use threshold triggering, but is based on symbol substructure pattern matching for identification; The target expression is decomposed into substructures to extract symbolic substructures containing energy conservation operators, capacity decay operators, and temperature rise threshold operators; The symbol substructure is matched with a pre-built fault mode library. When a substructure combination that matches the mode library appears, it is determined to be the corresponding fault category. The overheating fault corresponds to a substructure that couples a temperature rise threshold operator with cooling-related variables; The capacity decay fault corresponds to a substructure that couples the capacity decay operator with the loop count variable; Power anomalies correspond to substructures where the energy conservation operator deviates. The insufficient heat dissipation fault corresponds to the temperature rise threshold operator and the substructure of the imbalance of heat dissipation parameters.

9. The method for fault prediction and diagnosis of energy storage systems based on artificial intelligence according to claim 1, characterized in that, The dynamic weight penalty mechanism includes: During the symbolic expression search process, the residual distribution between the predicted values ​​of candidate expressions and the actual running data is calculated; When the residual distribution shows a systematic shift on the boundary variables related to energy conservation, capacity decay, or temperature rise, the penalty weight is adaptively adjusted according to the characteristics of the residual distribution. The competition among candidate expressions is modeled as a game process. Search resources are dynamically allocated through evolutionary game strategies. Expressions that conform to physical constraints gain higher fitness, while expressions that violate physical constraints are gradually eliminated.

Images