Distributed electrochemical energy storage fire alarm system based on intelligent algorithm

Abstract

The application discloses a distributed electrochemical energy storage fire alarm system based on an intelligent algorithm, relates to the technical field of energy storage battery safety monitoring, and aims at the early thermal runaway risk of batteries in a large energy storage power station, and proposes a technical process of multi-source data acquisition and preprocessing, key health factor extraction, multi-modal feature fusion, deep model detection, cross-system intervention and adaptive strategy optimization; high-quality data streams are generated through real-time synchronous sampling and noise filtering, multi-dimensional features are applied to highlight weak fault signs such as internal short circuit or material degradation, a deep learning model is used for accurate identification and output of early warning, and the BMS, EMS and fire extinguishing system are cooperated to implement load reduction, isolation and temperature control measures; an evolutionary algorithm or reinforcement learning is introduced into a simulation environment to continuously iterate the intervention strategy, so that the latent risks under multiple scene complex working conditions can be captured with high sensitivity and inhibited quickly, the false alarm rate and the false alarm rate can be reduced significantly, and the overall safety of the system can be improved significantly.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage battery safety monitoring, and in particular to a distributed electrochemical energy storage fire alarm system based on an intelligent algorithm. Background Art

[0002] With the increasing coordination between renewable energy generation and grid dispatch, electrochemical energy storage power stations are playing a key role in peak shaving and valley filling, as well as emergency power supply. Such energy storage facilities typically consist of large-scale battery module arrays covering a variety of chemistries (such as lithium-ion, sodium-sulfur, and all-vanadium liquid flow), and are equipped with complex energy management systems to cope with high concurrent charging and discharging requirements. In long-term operation, they must not only ensure power balance and dispatch efficiency at all levels, but also cope with multiple external factors such as temperature, humidity, and sudden load fluctuations. Because energy storage power stations often operate continuously year-round, even minor manufacturing defects in individual cells or localized material degradation during operation can trigger unpredictable micro-shorts or degradation reactions, which, through heat accumulation and cumulative effects, can gradually develop into thermal runaway. If these hidden dangers accumulate and go undetected, even if routine monitoring remains within nominal ranges, they could suddenly erupt, causing fires, explosions, or even complete station failure, resulting in significant losses to grid and property safety. Therefore, detecting subtle abnormal signals within batteries at an early stage and proactively addressing potential failures has become a core challenge in energy storage system operation and maintenance.

[0003] Existing energy storage power plants often rely on basic battery management systems (BMS) for threshold-based alarms, which can only provide alerts when parameters such as voltage and temperature deviate significantly, making it difficult to detect more subtle latent faults. When internal short circuits or material degradation in individual modules are still in the early stages, parameter changes often exhibit a small and insignificant distribution, making these "small fluctuations" easily overlooked by traditional methods. Furthermore, the high cost of collecting fault data under actual operating conditions results in limited model training samples, limiting the accuracy of extracting and identifying early fault signs. Furthermore, while a few R&D teams have attempted to integrate multi-source sensor information (such as voltage, temperature, and gas composition), the lack of unified multimodal fusion and in-depth analysis methods makes it difficult to fully characterize the fault mechanism. Furthermore, cross-system coordination mechanisms (such as firefighting equipment, temperature control units, and energy management systems) are not systematic, often resulting in only reactive post-event remediation after an incident has occurred.

[0004] In summary, how to achieve early warning and proactive prevention of potential thermal runaway through more sensitive anomaly recognition algorithms and more efficient linkage intervention modes in large-scale energy storage environments with huge capacity and complex operating conditions constitutes a technical challenge that the industry urgently needs to solve. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In response to the shortcomings of the existing technology, the present invention provides a distributed electrochemical energy storage fire alarm system based on intelligent algorithms. Aiming at the risk of early thermal runaway of batteries in large-scale energy storage power stations, the system generates high-quality data streams through real-time synchronous sampling and noise filtering, applies multi-dimensional feature construction to highlight weak fault signs such as internal short circuits or material degradation, and uses deep learning models to accurately identify and output early warnings. It then coordinates with BMS, EMS and fire protection systems to implement load reduction, isolation and temperature control measures, and introduces evolutionary algorithms or reinforcement learning continuous iterative intervention strategies in a simulation environment to achieve highly sensitive capture and rapid suppression of potential risks in complex working conditions in multiple scenarios, which can significantly reduce the rates of missed reports and false alarms, and solve the technical problems recorded in the background technology.

[0007] (2) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a distributed electrochemical energy storage fire alarm system based on an intelligent algorithm, including online filtering, differential interpolation and loss compensation of multi-source sensor outputs to eliminate random noise and severe error points, and generate a high-quality data stream that depicts the battery operating status;

[0009] Perform multi-scale wavelet energy analysis and vector fusion on multi-source signal dimensions to highlight their respective hidden features and output them in the form of multimodal feature vectors;

[0010] When the multimodal feature vector enters the thermal runaway anomaly detection model training and anomaly detection, a limited number of fault samples and normal samples are used to adjust the bidirectional LSTM and attention weights. In the online phase, an anomaly score is calculated at each moment. If it exceeds the dynamic threshold, a risk warning signal is immediately output.

[0011] If a risk warning signal is triggered, load reduction, isolation, and temperature control commands are issued to the BMS, EMS, and fire protection systems. Explosion suppression or pre-pressurization measures are linked through a pre-set command interface, and intervention data is transmitted back to update the threshold baseline of the thermal runaway anomaly detection model.

[0012] After completing the predefined intervention and initiating multiple rounds of strategy optimization in the simulation environment, a genetic algorithm is used to interactively trial and error the intervention action parameters and evaluate their combined impact on the safety margin and downtime cost. After continuous iteration, a better intervention strategy is generated.

[0013] Furthermore, various types of sensors are deployed in the energy storage power station, and the multi-source sensor signals are synchronously read at fixed time intervals to form a multi-source signal set; after the multi-source signal set is filtered and interpolated in real time, all filtered and compensated signals are reconstructed in the same period to obtain the preprocessed data stream.

[0014] Further, after obtaining the preprocessed signals of each sensor output, a feature construction method based on wavelet energy distribution combined with local trend analysis is adopted to define a health factor for the preprocessed signal of each sensor at discrete time.

[0015] Further, the health factors of the same sensor at different times are collected to obtain a one-dimensional health factor sequence, and a multi-dimensional health factor set is obtained after traversing all sensors. The health factor sequence is used to obtain a multi-modal feature vector at each time through a fusion operation function.

[0016] Further, the multi-modal feature vector and its corresponding time label are used to organize a training set in the training stage.

[0017] A hybrid framework of bidirectional recurrent network and self-attention mechanism is introduced. First, a bidirectional long short-term memory network is used to encode the input sequence in context. After obtaining the hidden representation, an attention module is used to calculate the attention weight to highlight the key moments in the sequence, and output the time series feature vector.

[0018] The parameter set is iteratively updated by stochastic gradient descent or adaptive optimization algorithm until convergence on the validation set or specified indicators are reached, obtaining a trained thermal runaway anomaly detection model that can be used for online detection.

[0019] Further, after receiving the multi-modal feature vector in real time in the online stage, the multi-modal feature vector is input into the trained thermal runaway anomaly detection model, and the aggregated vector is calculated through the same network structure as the offline training, and the anomaly score is output.

[0020] If the anomaly score is not less than the corresponding dynamic threshold, it is determined that there is a potential thermal runaway risk, and a warning signal is immediately output. If the warning signal is triggered, the corresponding intervention measures are executed, and the determination result generated by online detection is fed back to the thermal runaway anomaly detection model in real time.

[0021] Further, the anomaly score and the dynamic threshold are received, and the risk occurrence degree is determined according to the relative relationship between the two. According to the return value of the risk level function, a linkage interface is established with BMS, EMS, and fire fighting system: when the risk is moderate, a state monitoring reinforcement instruction is submitted to BMS, and when the risk is high, an emergency intervention instruction is sent to BMS, EMS, and fire fighting system.

[0022] Further, after obtaining the risk level determination result, the current running state information of BMS and EMS is combined.

[0023] Based on the risk level, a hierarchical execution scheme is given, and each subsystem needs to report the execution state to the cross-system linkage control unit in real time.

[0024] Further, at the moderate risk, instruct the BMS to perform the load reduction or isolation operation on the suspicious battery module, notify the EMS to complete the load transfer or peak shaving at the scheduling level, and put the fire extinguishing system in standby mode;

[0025] At the high-risk, the corresponding suspicious battery module is completely cut off from the main circuit, or the limit load reduction mode is adopted to link the temperature control unit and perform the rapid cooling operation, and the pre-pressurized fire extinguishing medium or explosion suppression device of the fire extinguishing system is started.

[0026] Further, the process data and execution state data after the execution of each system, the newly generated sensing data and working condition information after the intervention, and the abnormal score record are comprehensively analyzed to form a new training or verification data set; in the idle stage, the parameters or dynamic threshold of the thermal runaway abnormality detection model are incrementally updated in a small scale.

[0027] Further, the safety intervention action is continued, and after the abnormal score or fault probability information is collected, a simulation environment is constructed in the digital twin platform to simulate the evolution process of the battery internal fault and the state change caused by the execution of different intervention actions.

[0028] By running the intervention scheme in the simulation environment, the corresponding fault evolution curve and operation cost are recorded, and a diversified test data set for subsequent optimization algorithm training is obtained.

[0029] Further, the intervention action is abstracted as a multi-dimensional strategy vector in the simulation environment, and the optimal intervention scheme is searched in multiple rounds of trial and error, and the target function is maximized to balance safety and cost control, wherein,

[0030] The intervention strategy vector can be regarded as an individual chromosome by using a genetic algorithm, and selection, crossover and mutation operations are cyclically executed to evolve a better intervention scheme under the guidance of the target function.

[0031] Further, when the genetic algorithm reaches convergence or finds the optimal or suboptimal intervention scheme in the simulation environment, the output optimal strategy vector is deployed to the real energy storage system, and in real operation, if the risk level is high, the adaptive strategy is called first, the intervention action is dynamically generated or corrected, and the BMS, EMS and fire extinguishing system are executed.

[0032] After deployment, the safety indicators and operation cost indicators in the intervention process are monitored in real time to determine whether they are consistent with those in the simulation, and the difference data is again fed back to the algorithm fine-tuning link.

[0033] (Three) beneficial effects

[0034] The present application provides a distributed electrochemical energy storage fire alarm system based on intelligent algorithm, which has the following beneficial effects:

[0035] A complete technical solution, from multi-source data collection to active safety intervention, has been built for large-scale energy storage power plants. This allows for accurate early identification of potential thermal runaway hazards and prompt implementation of effective measures. Specific benefits are as follows:

[0036] High-speed synchronous sampling and online filtering mechanisms are used to comprehensively acquire and pre-process signals from multiple sensors such as voltage, temperature, and gas. Loss compensation and anomaly elimination operations are used to generate high-quality data streams, laying a stable and reliable foundation for subsequent analysis.

[0037] Relying on a multi-dimensional feature construction algorithm, the pre-processed data is converted into key health factors that can effectively characterize signs of internal short circuits or material degradation, and a unified multi-modal feature vector is used to Corroborating information from different sensors improves sensitivity to early and subtle fault signs;

[0038] The deep model combines bidirectional LSTM and attention mechanism, which can not only fully explore the contextual dependencies of sequence data, but also focus on key moments and use anomaly scoring to identify key moments. Or failure probability quantifies the hidden danger level; when abnormal score Exceeding dynamic threshold When a thermal runaway occurs, the early warning signal is output immediately, which can greatly reduce the probability of thermal runaway being ignored in its infancy.

[0039] Leveraging cross-system linkage, the BMS (Battery Management System) and EMS (Energy Management System) collaborate to implement load shedding, isolation, and proactive intervention strategies for temperature control and fire protection systems, eliminating risks in the shortest possible time. Intervention process data is fed back to the deep learning model for cyclical correction, ensuring continuous optimization of detection accuracy and response mechanisms.

[0040] Through evolutionary algorithms or reinforcement learning frameworks, multiple rounds of trial and error iterations in simulation environments or safety test benches are carried out to automatically explore the optimal intervention actions, generate adaptive policy functions, and seamlessly apply them to the linkage module in the fourth step after verification of convergence, greatly improving the flexibility and adaptability in multiple scenarios and multiple constraints.

[0041] In general, the multimodal feature vector , Abnormality Score With dynamic threshold The close integration of technical features such as and so on, has a rigorous logical connection from high-quality collection in the early stage to self-learning intervention in the later stage, which can significantly reduce missed alarms and false alarms to protect the safety of energy storage equipment to the greatest extent, reduce the risk of downtime and fire, and form a full-cycle safety management closed loop from early detection to rapid intervention to continuous evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1This is a schematic diagram of the distributed electrochemical energy storage fire alarm process based on the intelligent algorithm of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] See also Figure 1 The present invention provides a distributed electrochemical energy storage fire alarm system based on an intelligent algorithm, comprising:

[0045] Step 1: The data acquisition module performs online filtering, missing compensation, and anomaly rejection on the outputs of multiple source sensors, such as voltage, temperature, and gas. Combined with sampling synchronization and calibration, this module generates a pre-processed data stream with a unified timestamp, effectively suppressed noise interference, and significantly improved data integrity.

[0046] The step 1 includes the following:

[0047] Step 101: Real-time collection of multi-source data

[0048] Various types of sensors are deployed in the energy storage power station to record the following signals:

[0049] The voltage sensor output is recorded as ,in Indicates the voltage sensors , represents time; the temperature sensor output is recorded as ,in Indicates the Temperature sensors ; The output of the gas concentration sensor is recorded as ,in Indicates the Gas sensors ;

[0050] Through a unified data collection module, at fixed time intervals Synchronously read multi-source sensor signals to generate discrete sequences ,in ; Perform unified timestamp calibration on various sensors to ensure that the observations from different sensors correspond to the same physical time point , forming a multi-source signal set Through the distributed deployment of multiple types of sensors, various trace abnormal signals that may appear inside the battery can be captured at an early stage. A same-screen observation platform was built, laying the foundation for subsequent multimodal fusion.

[0051] Step 102: Dynamic preprocessing

[0052] For multi-source signal sets Perform real-time filtering to suppress high-frequency or low-frequency interference, using the following nonlinear smoothing function:

[0053]

[0054] in: Represents any sensor output in a multi-source signal set, such as 、 or ; and are all positive real numbers, and the smoothing coefficient , nonlinear index ; Used to nonlinearly reduce abnormal peaks or impulse noise while retaining slight fluctuations that may carry fault signs;

[0055] If a sensor at a certain moment has missing data due to instantaneous jitter, the following timing missing compensation function is used for interpolation:

[0056]

[0057] in, represents the number of steps before or after the missing point, It can make the same sensor maintain continuity in the case of occasional loss; reconstruct all filtered and compensated signals in the same period to obtain the pre-processed data stream , where every moment The signal values ​​above are complete and pre-processed result sets.

[0058] When used, it suppresses random pulse noise while maximizing the retention of the fluctuation characteristics required for early abnormal signals, improving the ability to identify subtle fault signs later, and the timing loss compensation function Improve the signal integrity and reduce the impact of data interruption on subsequent analysis; Step 101 generates an accurate multi-source signal set based on high-speed synchronous acquisition of multi-source sensors; Step 102 dynamically pre-processes these signals using nonlinear smoothing and timing loss compensation, and finally outputs a reconstructed pre-processed data stream .

[0059] Step 2: When the pre-processed data stream is sent to the feature extraction stage, the multimodal fusion module performs wavelet energy decomposition and trend enhancement operations on multidimensional signals such as voltage, temperature and gas, and merges the scattered features into a multimodal feature vector , highlighting early fault signs such as internal short circuits or material degradation and significantly improving subsequent identification sensitivity;

[0060] The second step includes the following:

[0061] Step 201: Extraction of key health factors

[0062] Get preprocessed data stream , which contains pre-processed signals from multiple source sensors such as voltage, temperature, and gas. The output of each sensor is collectively referred to as the pre-processed signal ,in is the sensor index, ;

[0063] In order to more accurately capture early failure modes such as internal short circuit and material degradation, a feature construction method based on wavelet energy distribution and local trend analysis is used to generate a characteristic for each sensor. At discrete moments Preprocessed signal on Define the health factor:

[0064]

[0065] Where: Represented on the wavelet scale Preprocessed signal The coefficients obtained by discrete wavelet transform are Used to distinguish different frequency bands or scale intervals, ;

[0066] Represents the first-order difference of the wavelet coefficient in time (or the first-order partial derivative in the continuous case), which is used to measure the instantaneous rate of change of the wavelet coefficient over time;

[0067] is the amplitude index, , through The result of Power can magnify small but high-frequency subtle anomalies in the overall integral; is the attenuation factor, ; Indicates The time range for local integration before and after, ; Indicates the vector norm of the wavelet coefficients after time differentiation (which can be 2-norm or other suitable norm) to obtain the magnitude;

[0068] The output of the above formula is This constitutes one of the key health factors, which will be used as the output health factor sequence of step 201 Elements, for the same sensor At different times By aggregating the above results, we can get a one-dimensional health factor sequence:

[0069]

[0070] when After traversing all sensors, we get a multidimensional health factor set .

[0071] When used, with the help of wavelet multi-scale decomposition, signal changes in different frequency bands can be captured simultaneously, and the sensitivity to transient anomalies can be enhanced. The introduction of wavelet energy distribution increases the proportion of pulses or fluctuations with smaller amplitude but containing fault information in the overall energy, making early faults more easily highlighted. It can take into account both the amplification of weak features and the adaptive screening of noise, providing more significant specificity for subsequent fault identification. The original outputs of different sensor types (voltage, temperature, gas) are uniformly measured using wavelet energy distribution under the same processing framework, which enhances the comparability between feature dimensions and the convenience of subsequent fusion.

[0072] Step 202: Multimodal Fusion

[0073] Directly use the health factor sequence output in step 201 ,in In order to make full use of the fault information contained in the health factor of each sensor, the fusion operation function is defined:

[0074]

[0075] Where: , used to perform the same power transformation on each health factor; The output is a dimensional vector, each component corresponds to a weighted health factor;

[0076] At every moment , the unique multimodal feature vector is obtained by fusion operation function :

[0077]

[0078] Multimodal feature vector It is the final output of the second step that is used as the core input in the third step (thermal runaway anomaly detection model training and online anomaly detection);

[0079] When used, different types of health factors are combined into a unified multimodal feature representation, which can simplify the structural design and training difficulty of the thermal runaway anomaly detection model. Different sensor anomaly modes can be mutually verified in the fusion results, further improving the overall insight into multi-source fault signs; completing the pre-processing of data Towards high-level feature vector First, a health factor construction method based on multi-scale wavelet energy effectively highlights weak signals of early faults. Second, a power fusion strategy is combined to unify the health factors of different sensors into the same feature vector space, thus forming a close technical collaboration: the former provides a factor sequence with high fault sensitivity, and the latter fuses these factor sequences into a high-dimensional vector that can be directly applied to thermal runaway anomaly detection model training and online inference.

[0080] Step 3: When the output multimodal feature vector When entering the thermal runaway anomaly detection model training and online detection process, the training module first uses limited fault data and a large number of normal samples to iteratively learn the fusion network of bidirectional LSTM and attention mechanism to obtain a deep representation of complex time series dependencies, and then calculates the anomaly score at each moment in the online stage. and with dynamic threshold Compare and immediately output warning signal if the limit is exceeded;

[0081] The step three includes the following:

[0082] Step 301: Thermal runaway anomaly detection model training

[0083] Using multimodal feature vectors and its corresponding time tag , organized as a training set during the training phase ,in Indicates known labels (such as normal, slightly abnormal, severely abnormal, etc.). When fault samples are lacking, they can be supplemented by controlled experiments or numerical simulation data to balance the ratio of positive and negative samples.

[0084] In order to take into account the coupling of time dependence and multi-dimensional features, a hybrid framework of bidirectional recurrent network and self-attention mechanism is introduced. The bidirectional long short-term memory network (Bi-LSTM) is first used to contextually encode the input sequence to obtain the hidden representation. , and then the hidden representation Use the attention module to calculate the attention weight to highlight the key moments in the sequence and output the time series feature vector , which is formally expressed as follows:

[0085]

[0086] Where: For A collection of adjacent time window indices;

[0087] is the attention weight, whose size determines the hidden representation In the merged output Contribution in

[0088] is the final time series feature vector after aggregation;

[0089] Attention weight It can be obtained through a learnable affine mapping function and a softmax operation, for example:

[0090]

[0091] in 、 、 are all trainable parameters, used to learn how to allocate attention to hidden states at different times within the same time window; first use and Hide state Projection and passing Nonlinear mapping, then use vector Do a dot product with the projection result to get the unnormalized attention energy , and finally for the same time window All within Perform softmax to obtain normalized attention weights , used for weighted aggregation of hidden states;

[0092] The final model output is passed through a fully connected layer or other discriminant head for fault classification or anomaly scoring. To enhance the ability to identify rare faults, a composite objective function is introduced:

[0093]

[0094] in: Represents the multimodal feature vector from the input The overall mapping process to the output prediction value (such as fault category or abnormal probability), the parameter set is recorded as ; It is a conventional classification or regression loss, which is used to approximate the known sample labels;

[0095] is the attention-based aggregation vector An additional regularization term or reconstruction loss can encourage the model to maintain a robust mapping of the sequence structure; Used to control the balance between the main loss and the regularization term;

[0096] Iteratively update the parameter set using stochastic gradient descent or an adaptive optimization algorithm (such as Adam) , until it converges on the validation set or reaches the specified indicator. After the training is completed, the final model parameter set that can be used for online detection is obtained ;

[0097] When used, the bidirectional recurrent structure combined with the attention module can extract the contextual dependencies of key abnormal moments from the multimodal feature vector sequence to enhance the sensitivity to early fault signals. The composite objective function can improve the ability to capture unknown patterns through additional regularization or reconstruction terms in a limited fault sample scenario, thereby reducing the missed detection rate. The offline training method enables the model to fully learn in a data-rich or simulated data-assisted environment, and then seamlessly connect to the online detection stage. The fusion design of attention aggregation and bidirectional recurrent network takes into account both deep time series modeling and local key feature extraction, and has stronger generalization capabilities than traditional simple LSTM or convolutional networks. It combines fault classification or anomaly scoring with reconstruction or regularization terms of sequence structure in one loss function, taking into account both discrimination accuracy and adaptability to unknown fault types.

[0098] Step 302: Online anomaly detection

[0099] Receive multimodal feature vectors in real time during the online phase , at every moment Freshly arrived data;

[0100] The multimodal feature vector Input to the thermal runaway anomaly detection model trained in step 301 (in Represents the best parameters obtained through training), the aggregation vector is calculated through the same network structure as offline training , and output anomaly score or failure probability:

[0101]

[0102] This result is the core indicator of this step and is subsequently directly compared with the dynamic threshold to determine whether a fault warning is generated;

[0103] In order to cope with the time-varying nature of fault signals in complex operating environments, a dynamic threshold can be maintained online. , for example, based on historical rating statistics within a recent window or fine-tuning the thermal runaway anomaly detection model based on idle periods;

[0104] like , it is determined that there is a potential risk of thermal runaway and an early warning signal is output immediately;

[0105] like , then enter the normal observation state and continue to wait for the data at the next moment;

[0106] Among them: trained thermal runaway anomaly detection model Refers to the deep neural network used for anomaly detection in the third step. Its structure can be summarized as follows: the bidirectional long short-term memory network (Bi-LSTM) layer captures the multimodal feature vector The forward and backward time dependencies of ;

[0107] The attention aggregation layer is parameterized Calculate the normalized attention weight at each moment , and the hidden state Weighted summation to obtain aggregate features ;

[0108] Fully connected discriminant head uses weights and bias Will ) is mapped to classification logits, and then outputs anomaly scores through Softmax or regression After the training is completed, all the trainable parameters of the network, namely , is fixed and used for anomaly inference in the online stage.

[0109] If the warning signal is triggered, corresponding intervention measures will be implemented. The judgment results generated by online detection will also be fed back to the thermal runaway anomaly detection model in real time for fine-tuning in subsequent batch updates or incremental learning, forming a closed-loop process of continuous improvement.

[0110] When in use, the same thermal runaway anomaly detection model is used for forward inference in the online stage to ensure the consistency of training and detection logic, significantly reducing redundancy and errors during model inference; the dynamic threshold mechanism can be adaptively adjusted as the environment and working conditions change, avoiding missed reports or false alarms caused by a single fixed threshold; the linkage call mechanism with the next step (step 4) can quickly switch system strategies when the risk of failure increases, minimizing the impact of accidents; dynamic thresholds are introduced in online detection The concept of not only based on historical scores but also combined with external conditions (such as ambient temperature, load level, etc.) gives the solution stronger adaptability.

[0111] Step four, when the early warning signal is triggered, immediately send the BMS (battery management system), EMS (energy management system) and fire fighting system to send the load or isolation instructions, and according to the actual thermal runaway risk level linkage temperature control unit to execute rapid cooling or pre-charging pressure fire extinguishing measures, the high risk module in the bud stage is quickly inhibited, the execution state and working condition data generated in the intervention process are real-time returned to the thermal runaway abnormal detection model for calibration update;

[0112] The step four includes the following contents:

[0113] Step 401, risk level determination and instruction distribution

[0114] Receive abnormal score And the dynamic threshold , and determine the risk occurrence degree according to the relative relationship; In order to make the intervention measures more targeted, the fault degree can be graded under the over-limit condition, and the risk level function is defined:

[0115]

[0116] In the formula: The risk level function; The value 0 represents normal, 1 represents moderate risk, and 2 represents high risk;

[0117] The high-risk boundary function based on the current threshold Dynamic adjustment, which can be defined as , wherein The adjustable multiplication factor is used to distinguish the specific boundary between moderate and high risk;

[0118] According to the return value of the risk level function , a linkage interface is established with the BMS, EMS and fire fighting system:

[0119] When the risk level = 1 (moderate risk), submit the state monitoring intensive instruction to the BMS, and notify the EMS to prepare for capacity or power scheduling at the same time;

[0120] When the risk level = 2 (high risk), emergency intervention instructions need to be sent to the BMS, EMS and fire fighting system at the same time, so as to execute stronger safety measures in the next step;

[0121] When used, through more detailed secondary determination of abnormal score , the hysteresis or false alarm that may exist under complex working conditions can be avoided; The clear risk level facilitates the unified scheduling of multiple systems when executing intervention strategies of different intensities, improves the cooperation efficiency, and introduces the high-risk boundary function which can be dynamically expanded , allowing adaptive adjustment of high-risk thresholds based on operating conditions or historical warning accuracy, which facilitates flexible configuration in actual deployment.

[0122] Step 402: Active security intervention strategy execution

[0123] Obtaining risk level determination results Then, based on the current operating status information of BMS and EMS, a graded execution plan is given based on the risk level:

[0124] When the risk is moderate, the BMS is instructed to reduce the load or isolate the suspicious battery module to reduce its operating current. The EMS is notified to complete load transfer or peak shaving at the dispatch level to reduce the overall operating pressure of the system. The fire protection system is placed in standby mode, but the fire extinguishing medium or explosion suppression device is not forcibly activated for the time being.

[0125] When the risk is high, the corresponding suspicious battery module will be completely cut off from the main circuit, or the temperature control unit will be linked to the extreme load reduction method and a rapid cooling operation will be performed. This can be done through a custom control function:

[0126]

[0127] Where: Indicates the current module temperature. Indicates the target safety temperature, is the temperature difference response index, which is used to quickly increase the response intensity of cooling power with temperature difference;

[0128] Start the fire protection system to pre-pressurize the fire extinguishing medium or explosion suppression device to ensure effective protection before thermal runaway actually occurs.

[0129] Each subsystem needs to report the execution status (such as the load reduction degree of the BMS, the load transfer ratio of the EMS, the fire extinguishing medium reserve of the fire protection system, etc.) to the cross-system linkage control unit in real time for evaluation and feedback in the subsequent step 403.

[0130] Among them, suspicious battery modules are determined by the online detection module in the third step after giving local abnormality scores to each battery module. Specifically:

[0131] In step 2, for each module The multi-source sensor data independently constructs health factors and further fuses them into multimodal feature vectors ; In the online phase of step 3, the multimodal feature vector Input into the same trained thermal runaway anomaly detection model , it is calculated that the module is at time Anomaly score ;

[0132] Maintain a local threshold for each module , can be adjusted adaptively based on historical rating statistics or online feedback. When the module is detected, it is determined that there are abnormal signs and marked as a suspicious battery module.

[0133] When in use, differentiated intervention measures are taken according to risk classification to avoid unnecessary resource consumption caused by excessive intervention in low-risk situations, and to prioritize the safety of personnel and equipment in high-risk situations; through the use of rapid cooling control functions, adaptive linkage with the temperature control unit is achieved, significantly reducing the probability of further spread of thermal runaway. Each subsystem shares the execution status in real time to ensure the traceability and controllability of intervention actions, reducing omissions or delays caused by information asymmetry; the temperature control strategy is expressed in the form of a power function The introduction of an intervention process, different from traditional linear or constant power cooling methods, can achieve cooling in a shorter time; relying on centralized scheduling and command issuance of cross-system linkage control units, it can achieve refined intervention based on the actual module health status, which is conducive to the management and expansion of large-scale energy storage scenarios.

[0134] Step 403: Feedback correction and data closed loop

[0135] Collect process data and execution status of each system after execution in step 402, including:

[0136] The load reduction degree and number of isolated battery modules returned by the BMS; the load transfer records and scheduling optimization results returned by the EMS; the activation status of the explosion suppression device and the remaining amount of fire extinguishing medium returned by the fire protection system;

[0137] Collect the newly generated sensor data and working condition information after the above intervention and compare them with the abnormality score Record and conduct comprehensive analysis, such as fault labeling and effect evaluation, feature re-extraction, etc., to form new training or verification data sets ; In the idle phase, the parameters of the thermal runaway anomaly detection model or dynamic threshold Perform small incremental updates, such as using the following formula:

[0138]

[0139] Where: are the model parameters currently used online; is the objective function for the feedback data, such as mini-batch cross entropy or the reconstruction error of the new fault label update; To update the step size, it can be determined in combination with the real-time requirements of the system;

[0140] While incrementally updating, dynamic thresholds can also be adjusted based on new fault distribution and system operation status. or high-risk cutoff function Exponential smoothing + quantile update is used for fine-tuning to ensure that the threshold setting continues to adapt to the new environment. After the update is completed, return to the third step of online anomaly detection, using the thermal runaway anomaly detection model and dynamic threshold after feedback correction. Proceed to the next round of detection and intervention.

[0141] During use, by incorporating intervention process data into subsequent training or verification links, the thermal runaway anomaly detection model can continuously adapt to changes in real operating conditions, improving the detection rate and accuracy; threshold adaptive adjustment can cope with the diverse operating environments and seasonal load fluctuations of the energy storage system, reducing false alarms and missed alarms; the dual fine-tuning of dynamic thresholds and high-risk boundary functions enables the system to maintain high adaptability to sudden events or major environmental changes; the risk grading mechanism makes the intervention method more precise, the active safety strategy enables potential thermal runaway to be quickly controlled, and the feedback correction deposits the intervention experience back into the thermal runaway anomaly detection model and threshold setting, increasing the ability of self-learning and self-correction, and providing safety protection for large-scale energy storage power stations to deal with the risk of internal battery failure.

[0142] Based on multimodal features and the detection results of thermal runaway anomaly detection models, it is now possible to implement predefined hierarchical intervention strategies for potential thermal runaway risks. However, fixed strategies cannot achieve the optimal balance between all failure scenarios and operational objectives (such as minimizing downtime costs, maximizing safety margins, or a balance between the two). To further improve adaptability to complex failure scenarios, evolutionary algorithms or reinforcement learning frameworks can be used in simulation environments or safety testbeds to automatically explore more optimal intervention methods and continuously optimize strategies through multiple rounds of iterations, thereby achieving self-learning active safety intervention while meeting the overall operating conditions.

[0143] Step 5: After the predefined intervention strategy is completed and the system enters a controllable simulation environment, the adaptive module uses a genetic algorithm to conduct multiple rounds of interactive trial and error on intervention action parameters such as load reduction ratio, temperature control power, and fire protection medium release amount, and evaluates their comprehensive impact on safety margin and downtime cost. It then iteratively generates the optimal intervention strategy. Once the optimal intervention strategy reaches convergence and is verified in the virtual environment, it can be seamlessly integrated into the cross-system linkage control unit.

[0144] The step five includes the following:

[0145] Step 501: Simulation environment construction and test data preparation

[0146] Continue with the safety intervention actions in step 4 (such as isolation ratio, load reduction range, temperature control power, fire protection medium dosage, etc.) and collect abnormality scores or failure probability information as an indicator for evaluating intervention effectiveness;

[0147] In the digital twin platform, a simulation environment is built based on the electrochemical, heat dissipation, and fire response models of actual energy storage battery modules. , in the simulation environment It can simulate the evolution of internal battery faults and the state changes caused by performing different intervention actions (such as temperature rise and module capacity attenuation).

[0148] Through the simulation environment Run a certain number of random or preset intervention plans, record the corresponding fault evolution curves and operation costs (such as downtime loss, energy consumption), and obtain a diverse test data set for subsequent optimization algorithm training. ;

[0149] When using, with the help of simulation environment , can quickly test various extreme scenarios and intervention plans without affecting the safety of the real system, provide basic data for the exploration of adaptive strategies, and the diverse initial fault conditions and intervention combinations ensure the test data set The richness and representativeness of the model, combined with the actual energy storage system parameters and electrochemical mechanisms, build a simulation environment with high fidelity It can significantly reduce the deviation or risk when the strategy is implemented in real scenarios.

[0150] Step 502: Optimization of intervention strategy based on evolutionary algorithm or reinforcement learning

[0151] In the simulation environment In the process, intervention actions (such as isolation ratio, temperature control power, fire protection medium delivery, etc.) are abstracted into multi-dimensional strategy vectors. ,in A set of possible actions that defines the system state , indicating the current fault level, temperature distribution load status and other information;

[0152] To automatically search for the optimal intervention plan in multiple rounds of trial and error, define the objective function :

[0153]

[0154] in: Used to quantify the degree of improvement in safety margin (such as reducing the probability of thermal runaway, reducing the risk of smoke or fire). A larger value indicates higher safety. It measures the loss caused by the execution intervention (such as downtime cost, energy consumption cost, etc.). The larger the value, the higher the cost.

[0155] The ultimate goal is to maximize the objective function To balance safety and cost control;

[0156] Genetic algorithm can be used to treat the intervention strategy vector as an individual chromosome, and cyclically perform selection, crossover and mutation operations. Evolve better intervention plans under the guidance of

[0157] During use, through repeated simulation and trial and error, hidden combinations of efficient intervention actions can be automatically discovered, and security requirements and operating costs can be taken into account during the evolution or learning process, avoiding the waste of resources or excessive protection that may be caused by fixed strategies; the intervention actions in the fourth step are parameterized and integrated into optimizable strategy vectors, and transformed from predefined to self-learning through evolutionary or reinforcement learning, which significantly improves the scalability and automation level of the solution in complex practical scenarios.

[0158] Step 503: Strategy deployment and field verification

[0159] When the genetic algorithm is used in the simulation environment After reaching convergence or finding the optimal (or suboptimal) intervention plan, the optimal strategy vector will be output Deployed to real energy storage systems;

[0160] In actual operation, if the risk level output in the third step is high (corresponding to the linkage trigger in the fourth step), the adaptive strategy can be called first to dynamically generate or modify intervention actions and execute them through the BMS, EMS and fire protection systems to achieve more flexible safety control. If the risk level is low, the predefined strategy will continue to be used to ensure overall stability.

[0161] After deployment, the safety indicators and operating cost indicators during the intervention process are monitored in real time to ensure they are consistent with those during simulation. The difference data is then fed back into the algorithm fine-tuning phase to continuously iterate and upgrade the adaptive strategy.

[0162] During use, the optimal strategy of the simulation phase is smoothly implemented in the real scenario, and linked with the hierarchical intervention in the fourth step to realize a practical adaptive safety strategy. Field verification can timely capture the gap between the simulation environment and the actual working conditions, enhance the online correction capability and increase the emergency handling capability. On the basis of the hierarchical intervention strategy in the fourth step, the intervention means are repeatedly explored and optimized in the simulation platform and safety test environment, and the high-level strategy finally evolved or learned is connected with the cross-system linkage control unit to form a closed-loop mechanism of self-learning-adaptation-reverification. This will not only greatly expand the scope of application of the original hierarchical intervention, but also more flexibly balance safety and downtime costs in complex and changeable energy storage system scenarios, and realize the continuous optimization of active safety management.

[0163] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0164] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0165] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is only for some logical functions. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0166] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0167] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. Distributed electrochemical energy storage fire alarm system based on intelligent algorithm, characterized by: include, The multi-source sensor outputs are filtered online, and differential interpolation is used to compensate for missing values, eliminating random noise and severe error points to generate a high-quality data stream that depicts the battery's operating status. After real-time filtering and interpolation of the multi-source signal set, all filtered and compensated signals are reconstructed in the same cycle to obtain a pre-processed data stream. Perform multi-scale wavelet energy analysis and vector fusion on multi-source signal dimensions to highlight their respective hidden features and output them in the form of multimodal feature vectors; When the multimodal feature vector enters the thermal runaway anomaly detection model training and anomaly detection phase, a limited number of fault samples and normal samples are used to adjust the bidirectional LSTM and attention weights. Anomaly scores are calculated at each moment in the online phase. If the score exceeds the dynamic threshold, a risk warning signal is immediately output. If a risk warning signal is triggered, load reduction, isolation, and temperature control commands are issued to the BMS, EMS, and fire protection systems. Explosion suppression or pre-pressurization measures are linked through a pre-set command interface, and intervention data is transmitted back to update the threshold baseline of the thermal runaway anomaly detection model. After completing the preset intervention and initiating strategy optimization in the simulation environment, a genetic algorithm is used to interactively trial and error the intervention action parameters and evaluate their combined impact on safety margins and downtime costs. A more optimal intervention strategy is generated through continuous iteration. When the genetic algorithm reaches convergence or finds the optimal or suboptimal intervention plan in the simulation environment, the output optimal strategy vector is deployed to the actual energy storage system. In real-world operation, if the risk level is high, the adaptive strategy is prioritized to dynamically generate or modify intervention actions, which are then executed through the BMS, EMS, and fire protection systems. After deployment, the safety and cost indicators during the intervention process are monitored in real time to ensure they are consistent with those during simulation, and any discrepancies are fed back into the algorithm fine-tuning phase. After obtaining the pre-processed signals output by each sensor, a feature construction method combining wavelet energy distribution and local trend analysis is used to define a health factor for each sensor's pre-processed signal at discrete moments. The health factors of the same sensor at different times are aggregated to obtain a one-dimensional health factor sequence. After traversing all sensors, a multidimensional health factor set is obtained. The health factor sequence is used to obtain a multimodal feature vector through a fusion operation function at each time.

2. The distributed electrochemical energy storage fire alarm system according to claim 1, characterized in that: Multiple types of sensors are deployed in the energy storage power station, and multi-source sensor signals are synchronously read at fixed time intervals to form a multi-source signal set.

3. The distributed electrochemical energy storage fire alarm system according to claim 2, characterized in that: Utilize multimodal feature vectors and their corresponding time labels to organize them into training sets during the training phase; A hybrid framework of bidirectional recurrent networks and self-attention mechanisms is introduced. A bidirectional long short-term memory network is first used to contextually encode the input sequence. After obtaining the hidden representation, an attention module is used to calculate attention weights to highlight key moments in the sequence and output a time series feature vector. The parameter set is iteratively updated through the adaptive optimization algorithm until convergence on the validation set or the specified indicator is reached, thereby obtaining a trained thermal runaway anomaly detection model for online detection.

4. The distributed electrochemical energy storage fire alarm system according to claim 3, characterized in that: After receiving the multimodal feature vector in real time during the online phase, the multimodal feature vector is input into the trained thermal runaway anomaly detection model. The aggregate vector is calculated using the same network structure as offline training and the anomaly score is output. If the anomaly score is not less than the corresponding dynamic threshold, it is determined that there is a potential thermal runaway risk and an early warning signal is immediately output; when the early warning signal is triggered, the corresponding intervention measures are executed, and the judgment results generated by the online detection are fed back to the thermal runaway anomaly detection model in real time.

5. The distributed electrochemical energy storage fire alarm system according to claim 4, characterized in that: Receive the anomaly score and dynamic threshold, and determine the degree of risk based on the relative relationship between the two, where: Based on the return value of the risk level function, a linkage interface is established with the BMS, EMS, and fire protection system: when the risk is moderate, a status monitoring enhancement instruction is submitted to the BMS; when the risk is high, an emergency intervention instruction is sent to the BMS, EMS, and fire protection system at the same time.

6. The distributed electrochemical energy storage fire alarm system according to claim 5, characterized in that: After obtaining the risk level determination result, the current operating status information of BMS and EMS is combined; a graded execution plan is given based on the risk level, and each subsystem reports the execution status to the cross-system linkage control unit in real time.

7. The distributed electrochemical energy storage fire alarm system according to claim 6, characterized in that: At moderate risk, the BMS is instructed to reduce the load or isolate the suspicious battery modules, the EMS is notified to complete load transfer or peak shaving at the dispatch level, and the fire protection system is placed in standby mode. In case of high risk, the corresponding suspicious battery module will be completely cut off from the main circuit, or the temperature control unit will be linked with extreme load reduction and rapid cooling operation will be performed, and the fire protection system will be started to pre-pressurize the fire extinguishing medium or explosion suppression device.

8. The distributed electrochemical energy storage fire alarm system according to claim 7, characterized in that: Collect process data and execution status data of each system after execution, newly generated sensor data and working condition information after intervention, and conduct comprehensive analysis with abnormal score records to form a new training or verification data set; During the idle phase, small-scale incremental updates are performed on the parameters or dynamic thresholds of the thermal runaway anomaly detection model.

9. The distributed electrochemical energy storage fire alarm system according to claim 8, characterized in that: After collecting anomaly scores or failure probability information, a simulation environment is built on the digital twin platform to simulate the evolution of internal battery failures and the state changes caused by executing different intervention actions; By running the intervention plan in a simulation environment and recording the corresponding fault evolution curve and operation cost, a diverse test data set is obtained for subsequent optimization algorithm training.

10. The distributed electrochemical energy storage fire alarm system according to claim 9, characterized in that: In the simulation environment, the intervention action is abstracted into a multi-dimensional strategy vector, the optimal intervention plan is searched in multiple rounds of trial and error, and the objective function is maximized to balance safety and cost control. A genetic algorithm is used to regard the multidimensional strategy vector as an individual chromosome, and cyclically perform selection, crossover and mutation operations to evolve a better intervention plan under the guidance of the objective function.

Images