Distributed electrochemical energy storage fire alarm system based on intelligent algorithm

Abstract

The invention discloses a distributed electrochemical energy storage fire alarm system based on an intelligent algorithm, and relates to the technical field of energy storage battery safety monitoring. Technical processes of multi-source data acquisition and preprocessing, key health factor extraction, multi-modal feature fusion, depth model detection, cross-system intervention and adaptive strategy optimization are provided; a high-quality data stream is generated through real-time synchronous sampling and noise filtering, weak fault symptoms such as internal short circuit or material deterioration are highlighted by applying a multi-dimensional feature structure, a deep learning model is utilized to accurately recognize and output early warning, and then load reduction, isolation and temperature control measures are carried out in cooperation with a BMS, an EMS and a fire fighting system. An evolutionary algorithm or a reinforcement learning continuous iteration intervention strategy is introduced in a simulation environment, high-sensitivity capture and rapid suppression of latent risks under multi-scene complex working conditions are achieved, the rate of missing report and false report can be remarkably reduced, and the overall safety of the system is remarkably enhanced.

Description

Technical Field

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

[0002] With the increasing improvement of the coordination degree between new energy power generation and power grid dispatching, electrochemical energy storage power stations play a key role in peak shaving, valley filling, emergency power supply, etc. Such energy storage facilities usually include a large-scale battery module array, covering a variety of chemical systems (such as lithium-ion, sodium-sulfur, all-vanadium redox flow, etc.), and are equipped with complex energy management systems to meet the requirements of high-concurrency charging and discharging. In the long-term operating environment, it is not only necessary to ensure the power balance and dispatching efficiency at all levels, but also to cope with the interference of multiple external factors such as temperature, humidity, and sudden load changes. Since energy storage power stations are mostly in a state of non-stop operation throughout the year, if there are slight manufacturing defects in individual batteries or local material aging occurs during operation, it may induce unpredictable micro-shorts or degradation reactions, and gradually evolve into thermal runaway under the effects of heat accumulation and superposition. If potential hazards accumulate and cannot be detected in time, even if the daily monitoring remains within the nominal range, it may suddenly break out at a certain moment, causing fires, explosions, or even the paralysis of the entire station, resulting in huge losses to power grid safety and property safety. Therefore, how to capture the subtle abnormal signals inside the battery at an early stage and make a forward-looking intervention in a timely manner for possible fault situations has become the core requirement in the field of energy storage system operation and maintenance.

[0003] Existing energy storage power stations mostly rely on basic battery management systems (BMS) for threshold-based alarms, and can only give prompts when parameters such as voltage and temperature deviate significantly, making it difficult to capture more concealed latent faults in a timely manner. When the internal short circuit or material degradation of a single module is still in the early stage, the parameter changes often show a small and insignificant distribution, and traditional methods are extremely likely to ignore such "small fluctuation signals"; at the same time, the acquisition cost of fault data under actual working conditions is very high, resulting in limited model training samples and restricted extraction and recognition accuracy of early fault signs. In addition, although a few R & D teams have tried to integrate multi-source sensor information (such as voltage, temperature, gas composition, etc.), there is a lack of unified multi-modal fusion and in-depth analysis means, making it difficult to form a complete description of the fault mechanism; the cooperation mechanism of cross-system linkage (such as fire-fighting equipment, temperature control units, energy management systems) has not been systematized, and often only passive after-the-fact remedies are carried out after an accident occurs.

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

[0005] 1. Technical issues to be resolved In view of the shortcomings of the prior art, 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, and uses multi-dimensional feature construction to highlight weak fault signs such as internal short circuits or material degradation. It uses deep learning models to accurately identify and output early warnings, and then coordinates with BMS, EMS and fire protection systems to reduce load, isolate and control temperature measures. It introduces evolutionary algorithms or reinforcement learning continuous iterative intervention strategies in a simulation environment, so as to achieve highly sensitive capture and rapid suppression of potential risks under 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.

[0006] (II) Technical solution 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 missing compensation of multi-source sensor outputs, eliminating random noise and serious error points, and generating high-quality data streams that characterize the battery operating status; Perform multi-scale wavelet energy analysis and vector fusion on multi-source signal dimensions to highlight their hidden features and output them in the form of multi-modal feature vectors; When the multimodal feature vector enters the thermal runaway anomaly detection model training and anomaly detection, the limited fault samples and normal samples are used to adjust the bidirectional LSTM and attention weights, and the anomaly score is calculated at each moment in the online stage. If it exceeds the dynamic threshold, the risk warning signal is output immediately; If the 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 the pre-set command interface, and intervention data is sent back to update the threshold baseline of the thermal runaway anomaly detection model; After completing the predefined intervention and starting 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, and a better intervention strategy is generated after continuous iteration.

[0007] Furthermore, various 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; after real-time filtering and interpolation of the multi-source signal set, all filtered and compensated signals are reconstructed in the same period to obtain a preprocessed data stream.

[0008] Furthermore, after obtaining the preprocessed signal output by each sensor, a feature construction method based on the combination of wavelet energy distribution and local trend analysis is adopted to define the health factor for the preprocessed signal of each sensor at a discrete moment.

[0009] Furthermore, 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 multi-dimensional health factor set is obtained, and a multi-modal feature vector is obtained at each time through a fusion operation function using the health factor sequence.

[0010] Furthermore, using the multi-modal feature vectors and their corresponding time tags, they are organized into a training set during the training phase; Introduce a hybrid framework of a bidirectional recurrent network and a self-attention mechanism. First, use a bidirectional long short-term memory network to perform context encoding on the input sequence. After obtaining the hidden representation, use the attention module to calculate the attention weights to highlight the critical moments in the sequence, and output the temporal feature vector; Iteratively update the parameter set through stochastic gradient descent or an adaptive optimization algorithm until convergence on the validation set or reaching the specified metrics, obtaining a trained thermal runaway anomaly detection model that can be used for online detection.

[0011] Furthermore, after receiving the multi-modal feature vectors in real-time during the online phase, input the multi-modal feature vectors into the trained thermal runaway anomaly detection model, calculate the aggregated vector through the same network structure as in offline training, and output the anomaly score; 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, corresponding intervention measures are executed, and the determination results generated by the online detection are fed back to the thermal runaway anomaly detection model in real-time.

[0012] Furthermore, receive the anomaly score and the dynamic threshold, and determine the degree of risk occurrence according to their relative relationship. Among them, according to the return value of the risk level function, establish linkage interfaces with the BMS, EMS, and fire protection system: in the case of moderate risk, submit a status monitoring enhancement instruction to the BMS; in the case of high-risk, send emergency intervention instructions to the BMS, EMS, and fire protection system simultaneously.

[0013] Furthermore, after obtaining the risk level determination result, combine the current operating status information of the BMS and EMS; Based on the risk level, give a hierarchical execution plan, and each subsystem needs to report the execution status to the cross-system linkage control unit in real-time.

[0014] Furthermore, in the case of moderate risk, instruct the BMS to perform load reduction or isolation operations on the suspicious battery module, notify the EMS to complete load transfer or peak shaving at the scheduling level, and put the fire protection system on standby; In the case of high-risk, completely disconnect the corresponding suspicious battery module from the main circuit, or use the extreme load reduction method to link the temperature control unit and perform rapid cooling operations, and start the fire protection system to pre-pressurize the fire extinguishing medium or explosion suppression device.

[0015] Furthermore, collect the process data and execution status data after each system executes, the newly generated sensing data and working condition information after intervention, and conduct comprehensive analysis with the abnormal score records to form a new training or verification data set; during the idle stage, perform small-scale incremental updates on the parameters or dynamic thresholds of the thermal runaway abnormal detection model.

[0016] Furthermore, continue to use the safety intervention actions, and after collecting the abnormal score or failure probability information, construct a simulation environment in the digital twin platform to simulate the evolution process of internal battery failures and the state changes brought about by performing different intervention actions; By running the intervention plan in the simulation environment, record the corresponding failure evolution curve and operation cost, and obtain a diverse test data set for subsequent optimization algorithm training.

[0017] Furthermore, abstract the intervention actions into multi-dimensional policy vectors in the simulation environment, search for the optimal intervention plan through multiple rounds of trial and error, and maximize the objective function to balance safety and cost control. Among them, The genetic algorithm can be used to regard the intervention strategy vector as an individual chromosome, and cycle through the selection, crossover, and mutation operations to evolve a better intervention plan under the guidance of the objective function.

[0018] Furthermore, when the genetic algorithm reaches convergence or finds the optimal or sub-optimal intervention plan in the simulation environment, deploy the output optimal policy vector to the real energy storage system. During real operation, if the risk level is high, preferentially call the adaptive policy, dynamically generate or correct the intervention actions, and execute them through the BMS, EMS, and fire protection systems; After deployment, real-time monitor whether the safety index and operation cost index during the intervention process are consistent with those during simulation, and return the difference data to the algorithm fine-tuning link again.

[0019] (III) Beneficial Effects The present invention provides a distributed electrochemical energy storage fire alarm system based on intelligent algorithms, having the following beneficial effects: A complete technical solution from multi-source data collection to active safety intervention is constructed in a large-scale energy storage power station, which can accurately identify potential thermal runaway hazards at an early stage and quickly take effective measures. The specific beneficial effects are as follows: With the high-speed synchronous sampling and online filtering mechanism, comprehensively obtain and preprocess the signals from multiple sensors such as voltage, temperature, and gas, and form a high-quality data stream through missing compensation and abnormal removal operations, laying a stable and reliable foundation for subsequent analysis; Relying on the multi-dimensional feature construction algorithm, convert the preprocessed data into key health factors that can efficiently characterize the signs of internal short circuit or material deterioration, and through a unified multi-modal feature vector Corroborating information from different sensors improves sensitivity to early and subtle fault signs; The deep model combines bidirectional LSTM and attention mechanism to fully explore the contextual dependencies of sequence data, focus on key moments, and use anomaly scoring to identify key moments. Or the probability of failure quantifies the level of hidden danger; when the 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. With the help of cross-system linkage, the BMS (battery management system) and EMS (energy management system) work together to implement active intervention strategies for load reduction, isolation, temperature control and fire protection systems, so as to eliminate risks in the early stage in the shortest way. At the same time, the intervention process data is fed back to the deep model for cyclic correction to ensure the continuous optimization of detection accuracy and response mechanism. Through evolutionary algorithms or reinforcement learning frameworks, multiple rounds of trial and error iterations in a simulation environment or safety test bench are carried out to automatically explore the optimal intervention action, generate a policy function with adaptive capabilities, and seamlessly apply it to the linkage module in the fourth step after verification of convergence, greatly improving the flexibility and adaptability in multiple scenarios and multiple constraints.

[0020] In general, the multimodal feature vector , Anomaly score With dynamic threshold It is closely integrated with other technical features, and has a rigorous logical connection from early high-quality data collection to later self-learning intervention, which can significantly reduce missed alarms and false alarms, protect the safety of energy storage equipment to the greatest extent, reduce downtime and fire risks, and form a full-cycle safety management closed loop from early detection to rapid intervention to continuous evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a schematic diagram of the distributed electrochemical energy storage fire alarm process based on the intelligent algorithm of the present invention. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0023] See also Figure 1 The present invention provides a distributed electrochemical energy storage fire alarm system based on intelligent algorithm, including: Step 1: The data acquisition module performs online filtering, missing compensation, and anomaly rejection operations on the outputs of multi-source sensors such as voltage, temperature, and gas, and combines sampling synchronization calibration means to form a preprocessed data stream with a unified timestamp, effectively suppressed noise interference, and significantly improved data integrity. The first step includes the following: Step 101: Real-time acquisition of multi-source data Deploy various types of sensors inside the energy storage power station to record the following signals respectively: The output of the voltage sensor is denoted as , where represents the th voltage sensor , represents time; the output of the temperature sensor is denoted as , where represents the th temperature sensor ; the output of the gas concentration sensor is denoted as , where represents the th gas sensor ; Through a unified data acquisition module, read the multi-source sensor signals synchronously at a fixed time interval to generate a discretized sequence , where ; perform unified timestamp calibration on various sensors to ensure that the observed values from different sensors correspond to the same physical time point to form a multi-source signal set ; through the distributed layout of multi-type sensors, various trace abnormal signals that may appear inside the battery can be captured at an early stage. Through the multi-source signal set a same-screen observation platform is constructed, laying a foundation for subsequent multi-modal fusion.

[0024] Step 102: Dynamic preprocessing Perform real-time filtering on the multi-source signal set to suppress high-frequency or low-frequency interference, and use the following non-linear smoothing function:

[0025] where: represents the output of any sensor in the multi-source signal set, such as , or ; and are both positive real numbers, and the smoothing coefficient , the non-linear exponent ; For non-linearly reducing abnormal peaks or impulse noises while retaining the minute fluctuations that may carry fault signs; If a certain sensor has data missing due to instantaneous jitter at a certain moment, interpolation is performed using the following time-series missing compensation function:

[0026] where represents the number of steps before or after the missing point, which can enable the same sensor to remain continuous in the case of accidental missing; perform co-periodic reconstruction on all filtered and compensated signals to obtain the preprocessed data stream where the signal value at each moment

[0027] When in use, while suppressing random impulse noises, maximize the retention of the fluctuation characteristics required for early abnormal signals, enhancing the subsequent recognition ability of subtle fault signs. The time-series missing compensation function improves the signal integrity and reduces the impact that data interruption may bring to subsequent analysis; Step 101 generates an accurate multi-source signal set based on high-speed synchronous acquisition of multi-source sensors; Step 102 then performs dynamic preprocessing on these signals using non-linear smoothing and time-series missing compensation, and finally outputs the preprocessed data stream after reconstruction .

[0028] In Step Two, when the preprocessed data stream is fed into the feature extraction stage, the multi-modal fusion module performs wavelet energy decomposition and trend enhancement operations on multi-dimensional signals such as voltage, temperature, and gas, and combines the scattered features into a multi-modal feature vector , highlighting early fault signs such as internal short circuits or material degradation and significantly enhancing the subsequent recognition sensitivity; The said Step Two includes the following contents: Step 201, extraction of key health factors Obtain the preprocessed data stream , which contains the preprocessed signals from multi-source sensors such as voltage, temperature, and gas. The outputs of each sensor are collectively referred to as preprocessed signals , where is the sensor index, ; To more accurately capture early fault modes such as internal short circuits and material degradation, a feature construction method combining wavelet energy distribution and local trend analysis is adopted. For each sensor at discrete moments the preprocessed signal define the health factor:

[0029] In the formula: represents the coefficient obtained by performing discrete wavelet transform on the preprocessed signal at the wavelet scale , and the subscript is used to distinguish different frequency bands or scale intervals. ; 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 change rate of the wavelet coefficient with time evolution; is the amplitude exponent. , by taking the power of the result of , the small but high-frequency subtle anomalies can be amplified in the overall integration; is the attenuation factor. ; represents the time range for local integration before and after , ; represents taking the vector norm (which can be the 2-norm or other appropriate norm) of the wavelet coefficient after time differentiation to obtain the amplitude magnitude; The output of the above formula constitutes one of the key health factors, and the above factors will be used as the output health factor sequence of step 201. By aggregating the results of the same sensor at different times , a one-dimensional health factor sequence can be obtained:

[0030] When traverses all sensors, a multi-dimensional health factor set is obtained.

[0031] In use, with the help of wavelet multi-scale decomposition, signal changes in different frequency bands can be captured simultaneously, enhancing the sensitivity to instantaneous anomalies. The introduction of the non-linear power exponent increases the proportion of pulses or fluctuations with small amplitudes but containing fault information in the overall energy, making early faults more prominent; it can balance the amplification of weak features and the adaptive filtering 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 in the same processing framework, enhancing the comparability between feature dimensions and the convenience of subsequent fusion.

[0032] Step 202, Multi-modal fusion Directly use the health factor sequence output by step 201 , where ; In order to make full use of the fault information contained in the health factors of each sensor, a fusion operation function is defined:

[0033] In the formula: , which is used to perform the same-power transformation on each health factor; The output of is a -dimensional vector, and each component corresponds to a weighted health factor; At each moment , a unique multi-modal feature vector is obtained through the fusion operation function :

[0034] Multi-modal feature vector Is exactly the final output of the second step and is used as the core input in the third step (thermal runaway anomaly detection model training and online anomaly detection); When in use, combining health factors of different types into a unified multi-modal feature representation can simplify the structural design and training difficulty of the thermal runaway anomaly detection model. The abnormal patterns of different sensors can be mutually verified in the fusion result, further improving the overall insight ability into multi-source fault signs; completing the conversion from preprocessed data to the high-level feature vector . First, the construction method of health factors based on multi-scale wavelet energy effectively highlights the weak signals of early faults; second, combining the power fusion strategy unifies the health factors of different sensors into the same feature vector space, thus forming a close technical synergy: 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 the training and online inference of the thermal runaway anomaly detection model.

[0035] Step three, when the output multi-modal feature vector enters the thermal runaway anomaly detection model training and online detection process, the training module first uses the limited fault data and a large number of normal samples to perform iterative learning on the fusion network of bidirectional LSTM and attention mechanism to obtain a deep representation of complex temporal dependencies, and then calculates the anomaly score at each moment in the online stage and compares it with the dynamic threshold . If it exceeds the limit, an early warning signal is immediately output; The third step includes the following content: Step 301, Thermal runaway anomaly detection model training Use the multi-modal feature vector and its corresponding time label , which is uniformly organized as a training set during the training phase , where represents known labels (such as normal, mildly abnormal, severely abnormal, etc.). When there is a lack of fault samples, controlled experiments or numerical simulation data can be used for supplementation to balance the proportion of positive and negative samples.

[0036] 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. First, the bidirectional long short-term memory network (Bi-LSTM) is used to perform context encoding on the input sequence to obtain a hidden representation , and then on the hidden representation An attention module is used to calculate attention weights to highlight the critical moments in the sequence, and an output temporal feature vector is obtained , which is formally expressed as follows:[[]]

[0037] In the formula:[[]] is a set of time window index sets adjacent to ; is the attention weight, and its magnitude determines the contribution degree of the hidden representation in the merged output ; is the final aggregated temporal feature vector; The attention weight can be obtained through a learnable affine mapping function and softmax operation, for example:[[]]

[0038] where , , are all trainable parameters used to learn how to allocate attention to hidden states at different moments within the same time window; first use and to project the hidden state and pass it through nonlinear mapping, and then use the vector to take the dot product with the projection result to obtain the unnormalized attention energy , and finally perform softmax on all within the same time window to obtain the normalized attention weight for weighted aggregation of hidden states; The final model output is used for fault classification or anomaly scoring through a fully connected layer or other discriminative heads. To enhance the ability to identify rare faults, a composite objective function is introduced:[[]]

[0039] Wherein: represents the overall mapping process from the input multi-modal feature vector to the output prediction value (such as fault category or anomaly probability), and the parameter set is denoted as ; is the conventional classification or regression loss for approximating the known sample labels; is the additional regularization term or reconstruction loss based on the attention aggregation vector which can encourage the model to maintain a robust mapping of the sequence structure; is used to control the balance between the main loss and the regularization term; The parameter set is iteratively updated through stochastic gradient descent or adaptive optimization algorithms (such as Adam) until convergence on the validation set or reaching the specified metrics. After training is completed, the final model parameter set ; In use, the bidirectional cyclic structure combined with the attention module can extract the context dependencies of key anomaly moments from the multi-modal feature vector sequence to improve the sensitivity to early fault signals. The composite objective function can, in the scenario of limited fault samples, improve the ability to capture unknown patterns through additional regularization or reconstruction terms, and can reduce the missed detection rate; The offline training method enables the model to fully learn in an environment with rich data or simulated data assistance, and then seamlessly connect to the online detection stage; The fusion design of attention aggregation and bidirectional recurrent network takes into account both deep temporal modeling and local key feature extraction, and has stronger generalization ability than traditional simple LSTM or convolutional networks. Combining fault classification or anomaly scoring with the reconstruction or regularization term of the sequence structure in a loss function takes into account both discriminant accuracy and adaptability to unknown fault types.

[0040] Step 302, Online anomaly detection In the online stage, the multi-modal feature vector is received in real time, and at each moment the newly arrived data; The multi-modal feature vector is input into the thermal runaway anomaly detection model trained in step 301 (where represents the optimal parameters obtained from training), and through the same network structure as in offline training, the aggregation vector is calculated and the anomaly score or fault probability is output:

[0041] This result is the core indicator of this step and will be directly compared with the dynamic threshold in the follow-up to determine whether a fault warning is generated; 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 score statistics within a recent window or based on the fine-tuning results of the thermal runaway anomaly detection model during idle periods; If , it is determined that there is a potential risk of thermal runaway, and a warning signal is immediately output; If , it enters the normal observation state and continues to wait for data at the next moment; Among them: the trained thermal runaway anomaly detection model refers to the deep neural network used for anomaly detection in the third step, and its structure can be summarized as follows: the bidirectional long short-term memory network (Bi-LSTM) layer captures the forward and backward time dependencies of multimodal feature vectors and outputs the hidden representation ; The attention aggregation layer calculates the normalized attention weights at each moment with the parameter set and performs a weighted sum on the hidden state to obtain the aggregated feature ; ; The fully connected discriminative head uses the weight and the bias to map ) to classification logits, and then outputs the anomaly score through Softmax or regression . After training, all the trainable parameters of this network, that is , are fixed and used for anomaly inference in the online stage.

[0042] If the warning signal is triggered, corresponding intervention measures will be executed. The determination results generated by the 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.

[0043] 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 according to changes in the environment and working conditions, avoiding false negatives or false positives caused by a single fixed threshold; the linkage call mechanism with the next step (the fourth step) can quickly switch the system strategy when the fault risk emerges, minimizing the impact of accidents to the greatest extent; the concept of dynamic threshold is introduced in the online detection, which not only considers historical scores but also combines external conditions (such as environmental temperature, load level, etc.), giving this solution stronger adaptability.

[0044] Step 4: When the warning signal is triggered, a load reduction or isolation instruction is immediately sent to the BMS (battery management system), EMS (energy management system) and fire protection system, and the temperature control unit is linked to execute rapid cooling or pre-pressurization fire extinguishing measures according to the actual thermal runaway risk level, so that high-risk modules are quickly suppressed in the budding stage. The execution status and operating condition data generated during the intervention process are sent back to the thermal runaway anomaly detection model in real time for calibration and update; The step 4 includes the following contents: Step 401: Risk level determination and instruction distribution Receive anomaly score With dynamic threshold , and determine the degree of risk based on the relative relationship between the two; in order to make the intervention measures more targeted, the degree of failure can be graded in the over-limit situation, and the risk level function can be defined:

[0045] Where: is the risk level function; the value 0 indicates normal, 1 indicates moderate risk, and 2 indicates high risk; Based on the current threshold Dynamically adjusted high-risk cutoff function, which can be defined ,in is an adjustable multiplication factor used to distinguish the specific boundaries between moderate and high risk; According to the risk level function The return value establishes a linkage interface with BMS, EMS, and fire protection systems: When the risk level = 1 (moderate risk), submit a status monitoring enhancement instruction to the BMS, and simultaneously inform the EMS to prepare for capacity or power dispatch; When the risk level = 2 (high risk), emergency intervention instructions need to be sent to the BMS, EMS and fire protection systems at the same time so that stronger safety measures can be implemented in the next step; When used, by scoring the anomaly Making more precise secondary judgments can avoid possible hysteresis or false alarms of a single warning threshold under complex working conditions; clear risk levels facilitate unified scheduling of multiple systems when executing intervention strategies of different intensities, improve coordination efficiency, and introduce dynamically scalable high-risk demarcation functions , allowing adaptive adjustment of high-risk thresholds based on operating conditions or historical warning accuracy, which helps to flexibly configure in actual deployment.

[0046] Step 402: Active security intervention strategy execution Obtaining risk level determination results Then, combined with the current operating status information of BMS and EMS, a graded execution plan is given based on the risk level: When the risk is moderate, instruct the BMS to reduce the load or isolate the suspicious battery module to reduce its working current, notify the EMS to complete load transfer or peak shaving at the dispatching level to reduce the overall operating pressure of the system, and put the fire protection system in standby mode, but do not force the activation of the fire extinguishing medium or explosion suppression device; When the risk is high, the corresponding suspicious battery module is completely disconnected from the main circuit, or the temperature control unit is linked to the extreme load reduction method and a rapid cooling operation is performed. The custom control function can be used:

[0047] 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 the cooling power with the temperature difference; 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.

[0048] 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.

[0049] Among them, the suspicious battery module is determined by the online detection module in the third step after giving a local abnormality score to each battery module. Specifically: In step 2, for each module The multi-source sensor data independently constructs health factors and further fuses them into multi-modal feature vectors ; In the online phase of step 3, the multimodal feature vector Input into the same trained thermal runaway anomaly detection model , and the module is calculated at time Anomaly score ; Maintain a local threshold for each module , can be adjusted adaptively based on historical scoring statistics or online feedback. When the module is detected, it is determined that there are abnormal signs and marked as a suspicious battery module.

[0050] During use, intervention measures are taken differentially according to the risk level to avoid unnecessary resource consumption caused by excessive intervention in low-risk situations, and to give priority to ensuring the safety of personnel and equipment in high-risk situations; through the use of a rapid cooling control function, an 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 in the form of a power function introduced into the intervention process, which is different from traditional linear or constant power refrigeration methods and can achieve cooling in a shorter time; relying on the centralized scheduling and instruction issuance of the cross-system linkage control unit, refined intervention based on the actual health status of the modules is achieved, which is beneficial to the management and expansion of large-scale energy storage scenarios.

[0051] Step 403, Feedback correction and data closed-loop Collect the process data and execution status after the execution of each system in step 402, including: The degree of load reduction and the number of isolated battery modules returned by the BMS; the load transfer record and scheduling optimization result returned by the EMS; the activation status of the explosion suppression device and the remaining amount of the fire extinguishing medium returned by the fire protection system, etc.; Collect the newly generated sensing data and working condition information after the above intervention, and conduct a comprehensive analysis with the abnormal score record, such as fault annotation and effect evaluation, feature re-extraction, etc., to form a new training or verification data set ; In the idle stage, the parameters of the thermal runaway anomaly detection model or the dynamic threshold are updated incrementally on a small scale, such as using the following formula:

[0052] In the formula: is the model parameter currently in use online; is the objective function for the feedback data, such as the small-batch cross-entropy or the reconstruction error of the updated new fault label; is the update step size, which can be determined in combination with the real-time requirements of the system; During the incremental update, according to the new fault distribution and system operation status, the dynamic threshold or the high-risk demarcation function is fine-tuned by using the method of exponential smoothing + quantile update to ensure that the threshold setting continues to adapt to the new environment; after the update is completed, return to the online anomaly detection stage in the third step, and continue to execute the next round of detection and intervention with the thermal runaway anomaly detection model and dynamic threshold after feedback correction.

[0053] During use, by incorporating the intervention process data into subsequent training or verification processes, the thermal runaway anomaly detection model can continuously adapt to real operating conditions, improving the detection rate and accuracy. The threshold self-adaptive adjustment can cope with the diverse operating environments of energy storage systems and seasonal load fluctuations, reducing false alarms and missed detections. The dual fine-tuning of the dynamic threshold and the high-risk demarcation function enables the system to maintain high adaptability to sudden events or major environmental changes. The risk grading mechanism makes the intervention methods more precise, the proactive safety strategy enables potential thermal runaway to be quickly controlled, and the feedback correction precipitates the intervention experience back into the thermal runaway anomaly detection model and threshold setting, increasing the ability of self-learning and self-correction, providing safety guarantees for large-scale energy storage power stations to cope with the risk of internal battery failures.

[0054] Based on multi-modal features and the detection results of the thermal runaway anomaly detection model, it is already possible to execute predefined hierarchical intervention strategies for potential thermal runaway risks. However, fixed strategies cannot achieve optimality between all fault scenarios and operating objectives (such as minimizing downtime costs, maximizing safety margins, or balancing the two). To further enhance the adaptability to complex fault situations, in a simulation environment or a safety test bench, an evolutionary algorithm or a reinforcement learning framework can be used to automatically explore better intervention methods and continuously optimize the strategy through multiple rounds of iteration, thereby achieving self-learning proactive safety intervention on the premise of meeting the overall operating conditions.

[0055] Step Five: After completing the predefined intervention strategy and the system enters a controllable simulation environment, the adaptive module conducts multiple rounds of interactive trial and error on intervention action parameters such as the load reduction ratio, temperature control power, and fire extinguishing medium dosage based on the genetic algorithm, and evaluates their comprehensive impact on the safety margin and downtime cost. Then, an optimal intervention strategy is iteratively generated. After the optimal intervention strategy converges and is verified in the virtual environment, it can be seamlessly connected to the cross-system linkage control unit; The content of the above Step Five includes the following: Step 501: Simulation environment construction and test data preparation Adopt the safety intervention actions in the fourth step (such as isolation ratio, load reduction amplitude, temperature control power, fire extinguishing medium dosage, etc.), and collect the anomaly scores or failure probability information as indicators to evaluate the intervention effect; In the digital twin platform, construct a simulation environment based on the electrochemical, heat dissipation, and fire response models of actual energy storage battery modules , within the simulation environment , the evolution process of internal battery faults and the state changes (such as temperature rise amplitude, module capacity attenuation) brought about by executing different intervention actions can be simulated; By in the simulation environment Run a certain number of random or preset intervention scenarios, record the corresponding fault evolution curves and operation costs (such as downtime losses, energy consumption), and obtain a diverse experimental dataset for subsequent optimization algorithm training. ; When in use, with the help of the simulation environment , various extreme scenarios and intervention scenarios can be quickly tested without affecting the safety of the real system, providing basic data for the exploration of adaptive strategies. The diverse initial fault conditions and intervention combinations ensure the richness and representativeness of the experimental dataset. Combining the actual energy storage system parameters and electrochemical mechanisms, a simulation environment with high fidelity is built which can significantly reduce the deviation or risk when the strategy is implemented in the real scenario.

[0056] Step 502: Optimization of intervention strategies based on evolutionary algorithms or reinforcement learning In the simulation environment , abstract the intervention actions (such as isolation ratio, temperature control power, fire extinguishing medium dosage, etc.) into a multi-dimensional strategy vector , where is the set of actionable actions, and define the system state , representing information such as the current fault level, temperature distribution, and load conditions; To automatically search for the optimal intervention scenario in multiple rounds of trial and error, define the objective function :

[0057] Where: is used to quantify the degree of improvement in safety margin (such as reducing the probability of thermal runaway, reducing the risk of smoking or fire), and the larger the value, the higher the safety; measures the losses incurred by implementing the intervention (such as downtime cost, energy consumption cost, etc.), and the larger the value, the higher the cost; The ultimate goal is to maximize the objective function to balance safety and cost control; The genetic algorithm can be used to regard the intervention strategy vector as an individual chromosome, and repeatedly perform selection, crossover, and mutation operations to evolve a better intervention scenario under the guidance of the objective function ; When in use, through repeated simulation and trial and error, hidden efficient intervention action combinations can be automatically discovered, and the safety requirements and operation costs are taken into account during the evolution or learning process, avoiding resource waste or overprotection that may be caused by fixed strategies; parameterize the intervention actions in the fourth step and incorporate them into the optimizable strategy vector, changing from predefined to self-learning through evolution or reinforcement learning, significantly improving the scalability and automation level of the solution in complex actual scenarios.

[0058] Step 503, Policy Deployment and Field Verification When the genetic algorithm reaches convergence or finds the optimal (or sub - optimal) intervention plan in the simulation environment After that, the output optimal policy vector Is deployed to the real energy storage system; During actual operation, if the risk level output in the third step is relatively high (triggered by linkage in the fourth step), this adaptive policy can be preferentially called to dynamically generate or correct intervention actions, which are executed through the BMS, EMS, and fire protection systems to achieve more flexible safety prevention and control. If the risk level is relatively low, the predefined policy is continued to ensure overall stability.

[0059] After deployment, continuously monitor whether the safety index and operation cost index during the intervention process are consistent with those during simulation, and send the difference data back to the algorithm fine - tuning link to continuously iterate and upgrade the adaptive policy; During use, smoothly implement the optimal policy in the simulation stage into the real scenario, and link it with the hierarchical intervention in the fourth step to achieve a practical adaptive safety policy. Field verification can promptly capture the gap between the simulation environment and the actual working conditions, enhance the online correction ability, and increase the emergency handling ability. Based on the hierarchical intervention strategy in the fourth step, repeatedly explore and optimize the intervention means in the simulation platform and the safety test environment, and dock the finally evolved or learned advanced policy with the cross - system linkage control unit to form a closed - loop mechanism of self - learning - self - adaptation - re - verification. This can not only greatly expand the applicable scope of the original hierarchical intervention, but also more flexibly balance safety and downtime cost in complex and variable energy storage system scenarios, realizing the continuous optimization of proactive safety management.

[0060] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0061] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0062] In several embodiments provided by the present 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 illustrative. For example, the division of the units is only for some logical function divisions. In actual implementation, there may be other division methods. For example, 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 displayed or discussed coupling or direct coupling or communication connection to each other can be through some interfaces. The indirect coupling or communication connection of the devices or units can be in electrical, mechanical, or other forms.

[0063] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0064] As described above, the above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A distributed electrochemical energy storage fire alarm system based on intelligent algorithms, characterized in that: including, online filtering, differential interpolation, and missing compensation are performed on the outputs of multi-source sensors to eliminate random noise and serious error points, generating a high-quality data stream that depicts the battery operating state; multi-scale wavelet energy analysis and vectorization fusion are performed on the dimensions of multi-source signals to highlight their respective hidden features and output them in the form of multi-modal feature vectors; When the multi-modal feature vectors enter the thermal runaway anomaly detection model training and anomaly detection stage, finite fault samples and normal samples are used to adjust the bidirectional LSTM and attention weights. Anomaly scores are calculated for each moment in the online stage. If the score exceeds the dynamic threshold, a risk warning signal is immediately output; If the 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-charging 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 starting strategy optimization in the simulation environment, the genetic algorithm is used to interactively trial and error the intervention action parameters and evaluate their comprehensive impact on safety margin and shutdown cost. A better intervention strategy is generated after continuous iteration.

2. The distributed electrochemical energy storage fire alarm system according to claim 1, characterized in that: Multiple types of sensors are deployed inside the energy storage power station, and multi-source sensor signals are synchronously read at fixed time intervals to form a multi-source signal set; After real-time filtering and interpolation of the multi-source signal set, all the filtered and compensated signals are reconstructed in the same period to obtain a preprocessed data stream.

3. The distributed electrochemical energy storage fire alarm system according to claim 2, characterized in that: After obtaining the preprocessed signals output by each sensor, based on the feature construction method combining wavelet energy distribution and local trend analysis, a health factor is defined for the preprocessed signal of each sensor at discrete moments.

4. The distributed electrochemical energy storage fire alarm system according to claim 3, characterized in that: The health factors of the same sensor at different moments are aggregated to obtain a one-dimensional health factor sequence. After traversing all sensors, a multi-dimensional health factor set is obtained. The multi-modal feature vector is obtained at each moment using the health factor sequence through a fusion operation function.

5. The distributed electrochemical energy storage fire alarm system according to claim 4, characterized in that: Using the multi-modal feature vector and its corresponding time tag, it is organized as a training set in the training stage; Introduce a hybrid framework of a bidirectional recurrent network and a self-attention mechanism. First, use the bidirectional long short-term memory network to encode the input sequence contextually. After obtaining the hidden representation, use the attention module to calculate the attention weights to highlight the critical moments in the sequence, and output the temporal feature vector; The parameter set is iteratively updated through an adaptive optimization algorithm until convergence on the validation set or reaching the specified metrics, obtaining a trained thermal runaway anomaly detection model for online detection.

6. The distributed electrochemical energy storage fire alarm system according to claim 5, characterized in that: After receiving the multi-modal feature vectors in real time during the online phase, input the multi-modal feature vectors into the trained thermal runaway anomaly detection model, calculate the aggregated vector through the same network structure as in the offline training, and output the anomaly score; 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; when the warning signal is triggered, corresponding intervention measures are executed, and the determination results generated by the online detection are fed back to the thermal runaway anomaly detection model in real time.

7. The distributed electrochemical energy storage fire alarm system according to claim 6, characterized in that: Receive the anomaly score and the dynamic threshold, and determine the degree of risk occurrence according to the relative relationship between the two, where According to the return value of the risk level function, establish linkage interfaces with the BMS, EMS, and fire protection system: in the case of medium risk, submit a status monitoring enhancement instruction to the BMS, and in the case of high risk, send emergency intervention instructions to the BMS, EMS, and fire protection system simultaneously.

8. The distributed electrochemical energy storage fire alarm system according to claim 7, characterized in that: After obtaining the risk level determination result, combine the current operating status information of the BMS and EMS; give a hierarchical execution plan based on the risk level, and each subsystem reports the execution status to the cross-system linkage control unit in real time.

9. The distributed electrochemical energy storage fire alarm system according to claim 8, characterized in that: In the case of medium risk, instruct the BMS to perform load reduction or isolation operations on the suspicious battery modules, notify the EMS to complete load transfer or peak shaving at the dispatching level, and place the fire protection system on standby; In the case of high risk, completely disconnect the corresponding suspicious battery modules from the main circuit, or use the extreme load reduction method to link the temperature control unit and perform rapid cooling operations, and start the fire protection system to pre-pressurize the fire extinguishing medium or explosion suppression device.

10. The distributed electrochemical energy storage fire alarm system according to claim 9, characterized in that: Collect the process data and execution status data after each system executes, the new sensing data and working condition information generated after the intervention, and conduct comprehensive analysis with the anomaly score records to form a new training or verification data set; During the idle phase, perform small-scale incremental updates on the parameters or dynamic thresholds of the thermal runaway anomaly detection model.

11. The distributed electrochemical energy storage fire alarm system according to claim 10, characterized in that: After collecting the anomaly score or failure probability information, build a simulation environment in the digital twin platform to simulate the evolution process of the internal battery fault and the state changes brought about by performing different intervention actions; By running the intervention plan in the simulation environment, record the corresponding fault evolution curve and operation cost, and obtain a diverse test data set for subsequent optimization algorithm training.

12. The distributed electrochemical energy storage fire alarm system according to claim 11, characterized in that: Abstract the intervention actions as multi-dimensional strategy vectors in the simulation environment, search for the optimal intervention plan through multiple rounds of trial and error, and maximize the objective function to balance safety and cost control, where The genetic algorithm can be used to regard the intervention strategy vector as an individual chromosome, and the selection crossover and mutation operations are cyclically executed to evolve a better intervention plan under the guidance of the objective function.

13. The distributed electrochemical energy storage fire alarm system according to claim 12, wherein: When the genetic algorithm converges in the simulation environment or finds the optimal or sub-optimal intervention plan, the output optimal strategy vector is deployed to the real energy storage system; during real operation, if the risk level is high, the adaptive strategy is preferentially called to dynamically generate or correct intervention actions, and they are executed through the BMS, EMS and fire protection systems; After deployment, it is monitored in real time whether the safety index and operation cost index during the intervention process are consistent with those during simulation, and the difference data is fed back to the algorithm fine-tuning link again.

Images