A method of thermal management for an energy storage system

Abstract

This application discloses a thermal management method for an energy storage system, comprising: acquiring an infrared image sequence collected by an infrared thermal imaging array in the energy storage system; constructing a hybrid model of physical information neural network and Transformer, using the partial differential equation of heat conduction as a soft-constraint embedding loss function, and extracting thermal feature tensors that conform to the laws of thermodynamics from the infrared image sequence; mapping multiple modal features, including at least infrared modes, to a common embedding space through four-modal contrastive learning pre-training, performing zero-sample anomaly detection without fault samples, and using model-independent element learning to adapt to the current energy storage system type; using a temporal graph attention network to dynamically learn a directed heat propagation path based on the thermal feature tensor, and outputting heat source location results and heat spread prediction results; and outputting thermal management collaborative control commands based on the heat source location results and the heat spread prediction results.

Description

Technical Field

[0001] This application relates to the field of energy storage safety monitoring technology, and in particular to a thermal management method for energy storage systems. Background Technology

[0002] With the rapid growth of installed capacity of new energy power generation, power energy storage systems such as lithium-ion battery energy storage, flow battery energy storage, flywheel energy storage, and compressed air energy storage have become key infrastructure for ensuring the stable operation of the power grid. Abnormal temperatures are a core cause of failures and even thermal runaway accidents in various energy storage systems. Infrared thermal imaging technology has been gradually applied to energy storage temperature monitoring, for example, by installing infrared imaging temperature sensors inside energy storage containers, arranging infrared probes around flow battery stacks, and setting infrared temperature measurement points at the ends of flywheel housings.

[0003] However, existing technologies still have the following prominent shortcomings:

[0004] First, pure data-driven deep learning models lack physical constraints. When sensor coverage is sparse or training data is insufficient, the predicted temperature field may produce abnormal results that violate the laws of heat conduction (e.g., the direction of the temperature gradient contradicts the location of the heat source), leading to false alarms or missed alarms.

[0005] Second, anomaly detection heavily relies on fault samples. Novel thermal runaway modes in real-world energy storage systems (such as slow thermal accumulation in specific material systems) often lack historical fault data, and existing supervised or semi-supervised methods cannot achieve zero-sample detection, rendering them ineffective for unknown fault types.

[0006] Third, the heat propagation path is invisible and cannot be actively blocked. Current technology can only detect existing hot spots, cannot predict the direction and speed of heat spread, and cannot cut off the heat propagation chain by actively adjusting cooling strategies.

[0007] Fourth, independent control of multiple cooling branches can easily lead to local competition. In traditional thermal management, each PID controller adjusts independently, which may result in a non-coordinated state where area A is overcooled and wastes energy while area B remains overheated, leading to poor overall energy efficiency and temperature uniformity.

[0008] Fifth, the fault diagnosis output is limited in form and lacks interpretability and interactive decision-making capabilities. Existing alarm systems only provide numerical or audible and visual signals, cannot generate professional-grade diagnostic reports, and do not support maintenance personnel in using natural language to perform counterfactual reasoning based on "what if...".

[0009] Sixth, there are data silos among multiple energy storage sites, joint modeling faces the risk of privacy leakage, and the AI ​​model itself is vulnerable to adversarial example attacks (such as maliciously constructed infrared images leading to missed detections). Summary of the Invention

[0010] This application provides a thermal management method for energy storage systems, aiming to solve the problems in the prior art where pure data-driven models lack physical constraints, cannot detect unknown faults with zero samples, cannot predict heat propagation paths to achieve active blocking, lack coordination in cooling control, have poor diagnostic interactivity, and are insufficient in data security and model robustness.

[0011] Firstly, a thermal management method for an energy storage system is provided, including:

[0012] Acquire the infrared image sequence acquired by the infrared thermal imaging array in the energy storage system;

[0013] A hybrid model combining physical information neural network and Transformer is constructed, and the partial differential equation of heat conduction is used as a soft-constraint embedding loss function to extract thermal feature tensors that conform to the laws of thermodynamics from the infrared image sequence.

[0014] By using four-modal contrastive learning pre-training, multiple modal features, including at least infrared modes, are mapped to a common embedding space to perform zero-sample anomaly detection without fault samples, and model-independent meta-learning is used to adapt to the current energy storage system type.

[0015] A temporal graph attention network is used to dynamically learn a directed heat propagation path based on the thermal feature tensor, and outputs heat source localization results and heat spread prediction results.

[0016] Based on the heat source location results and the heat spread prediction results, output thermal management collaborative control commands.

[0017] Optionally, in the above scheme, the step of using the heat conduction partial differential equation as a soft constraint embedding loss function specifically involves:

[0018] A residual term of the heat conduction equation is added to the overall loss function of the hybrid model. This residual term is used to constrain the temperature field prediction results output by the network to satisfy the partial differential equation of heat conduction.

[0019] Optionally, in the above scheme, the step of mapping multiple modal features, including at least the infrared modality, to a common embedding space through four-modal contrastive learning pre-training specifically involves:

[0020] Data from four sensors—infrared mode, visible light mode, gas mode, and vibration mode—are acquired simultaneously, and feature vectors are extracted from each of the four channels.

[0021] Using samples from the same modality at different times in the four feature vectors as negative sample pairs and samples from different modalities at the same time as positive sample pairs, the feature distance between positive sample pairs is brought closer and the feature distance between negative sample pairs is pushed apart by the comparison loss function, and the four feature vectors are mapped to the common embedding space.

[0022] After mapping the feature vector of the sample to be detected to the common embedding space, the distance between it and the cluster center of the normal sample features is calculated. When the distance exceeds a preset threshold, it is determined to be abnormal.

[0023] Optionally, in the above scheme, the step of using model-independent element learning to adapt to the current energy storage system type specifically includes:

[0024] Acquire historical data from at least two of the following systems: lithium-ion battery energy storage system, flow battery energy storage system, flywheel energy storage system, and compressed air energy storage system; and learn general initialization parameters during the meta-training phase.

[0025] During the meta-testing phase, the general initialization parameters are updated one or more times using a small number of labeled samples of the current energy storage system type to obtain model parameters that are adapted to the current energy storage system type.

[0026] Optionally, in the above scheme, the use of a temporal graph attention network to dynamically learn a directed heat propagation path based on the thermal feature tensor specifically involves:

[0027] A dynamic graph structure is constructed using each monitoring node in the energy storage system as a graph node and the thermal influence relationship between each monitoring node as an edge.

[0028] For each time step, the temporal attention coefficient between the current monitoring node and its neighboring monitoring nodes is calculated, and the temporal attention coefficient is used as the weight of the directed edge to represent the direction of heat flow.

[0029] Backpropagation is performed on the temperature change sequence of each monitoring node along the time dimension to trace the earliest starting monitoring node where the temperature anomaly occurred as the heat source location result.

[0030] Based on the current temperature and its changing trend of each monitoring node, and combined with the weight of the directed edge, the arrival time and temperature peak of heat spread to adjacent monitoring nodes are predicted.

[0031] Optionally, in the above scheme, the output thermal management collaborative control command specifically includes:

[0032] Construct a multi-agent deep reinforcement learning controller, treating each cooling branch in the energy storage system as an agent, with each agent sharing global state information;

[0033] With the optimization objectives of minimizing total global energy consumption and maximizing overall temperature field uniformity, the control actions of each agent are output through centralized training and distributed execution strategies.

[0034] The thermal management coordinated control commands include cooling flow commands for each cooling branch, fan speed commands, and pre-cooling start commands.

[0035] In the above scheme, optionally, the reward function of the multi-agent deep reinforcement learning controller is:

[0036] ;

[0037] in, The total energy consumption of all cooling branches. Let be the root mean square deviation of the temperature at each monitoring point within the energy storage system. For the action of the i-th agent, The average value of all agent actions. , , The weighting coefficients are used; the reward function is used to jointly optimize the minimization of total global energy consumption, the maximization of overall temperature field uniformity, and the minimization of the differences in actions of each agent.

[0038] Optionally, in the above scheme, after the output heat source location result and the heat spread prediction result, the following steps are also included:

[0039] Acquire sparse multi-angle infrared images simultaneously captured by infrared thermal imagers from at least two different perspectives deployed in the energy storage system, as well as the pose parameters and internal parameters of each infrared thermal imager.

[0040] The sparse multi-angle infrared image is input into the neural radiation field network. The pose parameters and internal parameters are used as light sampling parameters. Sampling is performed along each light ray in three-dimensional space. The position coordinates and direction vector of the sampling point are input into the neural radiation field network. The temperature and voxel density of each sampling point are output.

[0041] A dense three-dimensional temperature voxel field is generated through voxel rendering integration. The spatial resolution of the three-dimensional temperature voxel field is higher than that of a single frame image in the infrared image sequence.

[0042] Optionally, in the above scheme, after the output thermal management collaborative control command, the following further includes:

[0043] The large language model, which has been fine-tuned by instructions from the energy storage field, is accessed. The heat source location result, the heat spread prediction result, and the current operating status parameters of the energy storage system are concatenated into prompt words and input into the large language model.

[0044] By using a retrieval-enhanced generation mechanism, historical fault cases matching the current anomaly type are retrieved from the knowledge graph. The large language model then combines these historical fault cases to generate and output a natural language fault diagnosis report.

[0045] In response to a counterfactual hypothesis query input by the user, the system invokes a physical simulator to simulate the temperature evolution process of the energy storage system under the conditions specified in the counterfactual hypothesis query, or the counterfactual hypothesis query is directly input into the large language model for it to infer and generate the expected result, and the expected result is fed back to the user.

[0046] Optionally, the above scheme may also include:

[0047] A federated learning architecture is adopted. The local model training gradients of each energy storage site are encrypted and then uploaded to the central server. The central server aggregates the encrypted gradients of each site and distributes the aggregated global model parameters to each site for local updates.

[0048] An adversarial attack detection module is used to monitor the input data of each modality in real time, calculate the local intrinsic dimension of each input sample in the feature space, and when the local intrinsic dimension exceeds a preset security threshold, the current input sample is determined to be an adversarial sample and is blocked.

[0049] Compared with the prior art, this application has at least the following beneficial effects:

[0050] Based on further analysis and research of existing technical problems, this application recognizes that existing technologies suffer from issues such as a lack of physical constraints in purely data-driven models, inability to detect unknown faults with zero samples, inability to predict heat propagation paths for active blocking, lack of coordination in cooling control, poor diagnostic interactivity, and insufficient data security and model robustness. This application addresses these issues by constructing a hybrid model of physical information neural networks and Transformers, using the partial differential equation of heat conduction as a soft-constraint embedding loss function. This ensures that the extraction process of thermal feature tensors is constrained by thermodynamic laws, eliminating abnormal predictions that violate physical laws. Furthermore, through four-modal contrastive learning pre-training, multimodal features are aligned to a common embedding space. Combined with model-independent element learning, this allows for rapid adaptation to different energy storage systems, achieving fault-free... The system can identify early signs of unknown thermal runaway by sampling; it dynamically learns the directed heat propagation path between monitoring nodes through a time-series graph attention network, traces back the earliest abnormal node along the time dimension to locate the heat source and predict the heat spread trend, providing a basis for proactive intervention; by outputting collaborative control commands based on the heat source location and heat spread prediction results, it enables cooling resources to be pre-deployed according to the heat propagation direction and each branch to work together, achieving high-confidence temperature perception that integrates physical laws and data-driven approaches, cross-system zero-sample anomaly early warning without fault samples, visualized tracing and proactive blocking of heat propagation paths, and energy efficiency optimization effect of on-demand collaboration of multiple cooling branches, fundamentally breaking through the limitations of traditional threshold alarms that are passively responded to and cannot predict the direction of heat spread. Attached Figure Description

[0051] Figure 1A schematic flowchart illustrating a thermal management method for an energy storage system provided in one embodiment of this application;

[0052] Figure 2 A schematic flowchart of a thermal management method for an energy storage system provided in another embodiment of this application;

[0053] Figure 3 This is a schematic diagram of a physical information neural network-spatiotemporal Transformer hybrid model structure provided in one embodiment of this application;

[0054] Figure 4 This is a schematic diagram of a four-modal contrastive learning pre-training framework provided in one embodiment of this application;

[0055] Figure 5 This is a schematic diagram of heat propagation tracing in a timing graph attention network provided in one embodiment of this application;

[0056] Figure 6 A schematic diagram of super-resolution reconstruction of the three-dimensional temperature field of the neural radiation field provided in one embodiment of this application;

[0057] Figure 7 This is a flowchart of generative adversarial network abnormal scenario inference and adversarial training provided in one embodiment of this application;

[0058] Figure 8 A diagram of a multi-agent deep reinforcement learning collaborative thermal management architecture provided in one embodiment of this application;

[0059] Figure 9 A flowchart of interactive diagnosis and counterfactual reasoning using a large language model and knowledge graph provided in one embodiment of this application;

[0060] Figure 10 A schematic diagram of a federated learning and adversarial defense mechanism provided in one embodiment of this application;

[0061] Figure 11 A schematic diagram illustrating the deployment of an infrared thermal imaging array in a lithium-ion battery energy storage container according to an embodiment of this application;

[0062] Figure 12 This is a schematic diagram illustrating the deployment of an infrared thermal imaging array in an underground compressed air energy storage tunnel, according to one embodiment of this application. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "including," "comprising," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0065] To facilitate understanding, this specification uses different numbering systems to describe the same technical solution from multiple perspectives. Steps A1 to A5 present the complete process of the solution in a relatively general manner, while steps S1 to S7 elaborate on the specific implementation details of each stage of the process. The technical solutions described by the two numbering systems are essentially the same, both being a complete thermal management process from infrared image acquisition to coordinated control command output. There are no substantial differences between them, and both can be read together for a more comprehensive understanding of the embodiments of the present invention.

[0066] In one embodiment, a thermal management method for an energy storage system is provided, such as Figure 1 As shown, the method includes steps A1-A5:

[0067] Step A1: Obtain the infrared image sequence acquired by the infrared thermal imaging array in the energy storage system.

[0068] Specifically, an infrared thermal imaging array, consisting of multiple infrared thermal imagers, is deployed in the energy storage system. For lithium-ion battery energy storage containers, each infrared thermal imager is fixedly mounted on a bracket on the top of the container, with partial overlap in the field of view of adjacent infrared thermal imagers to cover the entire monitoring area of ​​the energy storage system. Each thermal imager covers a predetermined number of battery clusters (e.g., 2 to 4 battery clusters). For underground tunnels storing compressed air energy, infrared thermal imagers are arranged at intervals along the inner wall of the tunnel.

[0069] Infrared thermal imagers continuously acquire infrared images at a preset sampling frequency (e.g., one frame every 5 seconds). Each infrared thermal imager is connected to the data acquisition unit of the energy storage management platform via Ethernet or an industrial bus. The data acquisition unit receives the infrared images acquired by each infrared thermal imager, forms an infrared image sequence, and stores it in a cache or database after aligning it with timestamps for subsequent steps.

[0070] In one embodiment, the acquisition frequency of the infrared thermal imager is not a fixed value, but is dynamically adjusted according to the rate of temperature change: the acquisition frequency is increased when the rate of temperature change increases, and the acquisition frequency is decreased when the rate of temperature change decreases, so as to reduce data storage and transmission overhead while ensuring monitoring accuracy.

[0071] In one embodiment, the infrared thermal imaging array can be deployed not only in lithium-ion battery energy storage containers or compressed air energy storage chambers, but also around flow battery stacks or at the ends of flywheel energy storage housings to meet the monitoring needs of different energy storage system types.

[0072] Step A2: Construct a hybrid model of physical information neural network and Transformer, using the partial differential equation of heat conduction as a soft constraint embedding loss function to extract thermal feature tensors that conform to the laws of thermodynamics from the infrared image sequence.

[0073] Specifically, the hybrid model of Physics-Informed Neural Network (PINN) and Spatiotemporal Transformer includes a spatiotemporal Transformer encoder and a physical information constraint layer.

[0074] The spatiotemporal Transformer encoder is used to encode spatiotemporal features of infrared image sequences. For the input infrared image sequence, the spatial feature maps of each frame are first extracted through convolutional layers. Then, the spatial feature maps at different time steps are concatenated along the time dimension. The multi-head self-attention mechanism of the Transformer is used to capture the long-range spatiotemporal dependencies between the monitoring areas, and the preliminary thermal feature tensor is output.

[0075] The physical information constraint layer is used to embed the heat conduction partial differential equation into the model training process. The heat conduction partial differential equation describes the variation of the temperature field with time and space. In this embodiment, this partial differential equation is used as a soft constraint, and a residual term of the heat conduction equation is added to the overall loss function of the model.

[0076] Specifically, the overall loss function is expressed as:

[0077] ;

[0078] in, This refers to the data fitting loss (e.g., the mean square error between the predicted and actual temperature values). For physical constraint loss, These are the weighting coefficients.

[0079] The physical constraint loss By randomly sampling several collocation points within the spatiotemporal region of the network's output temperature field, the residuals of the partial differential equation for heat conduction at each collocation point are calculated, and the residuals are summed by mean square. This residual term is used to constrain the network's output temperature field predictions to satisfy the partial differential equation for heat conduction, meaning that the spatial and temporal variation of the temperature field must conform to the second law of thermodynamics. By minimizing the overall loss function that includes this residual term, the model learns not only to fit the observed data during training but also to satisfy physical laws, thereby eliminating anomalous predictions that violate thermodynamic laws, such as non-physical results where the temperature gradient direction contradicts the location of the heat source, that might arise from a purely data-driven model.

[0080] In one embodiment, the aforementioned physical information constraints are not limited to being implemented as residual terms in the loss function. Alternatively, the partial differential equation of heat conduction can be directly constructed as an independent physical information layer of the network. This layer receives the temperature field output from the previous network layer, calculates its residual satisfying the heat conduction equation, and feeds it back to the network as a regularization term. Alternatively, a physics-guided data augmentation approach can be used, where training samples are transformed according to the heat conduction equation (e.g., a small perturbation to the temperature field satisfying the heat conduction equation) to generate additional training samples, thereby implicitly enabling the network to learn physical laws.

[0081] Different governing equations can be used as soft constraints for different types of energy storage systems. For example, for flow battery energy storage systems, due to the significant convective heat transfer effect of the electrolyte, a convection-diffusion equation can be added to the heat conduction equation as a supplementary soft constraint to more accurately describe its heat transfer characteristics.

[0082] Step A3 involves mapping multiple modal features, including at least the infrared modality, to a common embedding space through four-modal contrastive learning pre-training, performing zero-sample anomaly detection without fault samples, and utilizing model-independent meta-learning to adapt to the current energy storage system type.

[0083] In one embodiment, this step further includes a four-modal contrastive learning pre-training phase, a zero-shot anomaly detection phase, and a model-independent meta-learning adaptation phase.

[0084] (a) Pre-training stage of four-modal contrastive learning.

[0085] Multimodal sensors are deployed simultaneously in the energy storage system, including: an infrared thermal imaging array (infrared mode), a visible light camera (visible mode), a gas sensor (gas mode), and a vibration sensor (vibration mode). Each sensor synchronously acquires multimodal data of the energy storage system at the same time.

[0086] Sensor data from the four modes at the same time are input into the corresponding feature extraction networks to extract four feature vectors. The feature extraction of the infrared mode is completed by the hybrid model described in step A2, and the output thermal feature tensor is used as the infrared mode feature vector after global pooling. The visible light mode, gas mode, and vibration mode are respectively extracted by their respective encoder networks (e.g., convolutional neural networks or deep belief networks) to extract the corresponding feature vectors.

[0087] A contrastive learning framework is employed to map four feature vectors to a common embedding space. Specifically, feature vectors from different modalities at the same time step are used as positive sample pairs (e.g., infrared and visible light features at the same time step constitute a positive sample pair), while feature vectors from the same modality at different time steps are used as negative sample pairs (e.g., infrared features at time t1 and infrared features at time t2 constitute a negative sample pair). A contrastive loss function (e.g., the InfoNCE loss function) is used for training to narrow the feature distance between positive sample pairs in the common embedding space while widening the feature distance between negative sample pairs. Through this cross-modal contrastive learning, the model learns the semantic representations shared between different modalities, meaning that under normal conditions, the features of each modality form a compact cluster distribution in the common embedding space.

[0088] In one embodiment, the multimodal sensor is not limited to infrared, visible light, gas, and vibration sensors; it can also include acoustic emission sensors or ultrasonic detection sensors to provide more dimensions of state information. In another alternative embodiment, the contrastive learning pre-training is not limited to four-modal joint training; it can also employ any combination of two or three modalities for contrastive learning. In yet another alternative embodiment, the contrastive learning loss function is not limited to InfoNCE; it can also use the SupCon loss function or the SimCLR loss function.

[0089] (ii) Zero-sample anomaly detection phase.

[0090] After the contrastive learning pre-training is completed, only normal samples are used for training in the common embedding space. Therefore, the feature vectors of normal samples form a compact cluster distribution in the embedding space, while the feature vectors of abnormal samples (including known fault types and unknown fault types) will deviate from this cluster distribution.

[0091] The zero-sample anomaly detection is achieved by mapping the modal feature vectors of the sample to be detected to the common embedding space, calculating the distance (e.g., Euclidean distance or cosine distance) between the feature vector of the sample to be detected and the cluster center of the normal sample features, and determining that the current energy storage system is abnormal when the distance exceeds a preset detection threshold. The detection threshold is determined statistically based on the distribution range of normal samples in the embedding space.

[0092] Since this anomaly detection process does not rely on any labeled data of fault samples, it only needs normal samples to construct the detection boundary. Therefore, it is also capable of identifying newly emerging and unrecorded thermal runaway precursors or fault types, breaking through the dependence of traditional supervised learning methods on fault samples.

[0093] (III) Model-independent meta-learning and adaptation stage.

[0094] Considering that different types of energy storage systems (lithium-ion batteries, flow batteries, flywheel energy storage, and compressed air energy storage) have different thermophysical characteristics and monitoring layouts, this embodiment uses model-agnostic meta-learning (MAML) to enable a single model to quickly adapt to multiple energy storage systems.

[0095] During the meta-training phase, historical data from at least two types of energy storage systems are acquired, and the data for each system are constructed into different meta-tasks. Each meta-task contains a support set and a query set. During meta-training, the model performs one or more gradient updates on the support set, calculates the loss on the query set, and optimizes the model's initial parameters through backpropagation, ultimately learning a set of universal initialization parameters. These universal initialization parameters exhibit good cross-system generalization ability.

[0096] During the meta-testing phase (i.e. the actual deployment phase), a small number of labeled samples of the current energy storage system type (such as a small amount of temperature data under normal conditions) are obtained. Based on the general initialization parameters, one or more gradient updates are performed to quickly adapt to the specific thermal characteristics of the current energy storage system and obtain the adapted model parameters.

[0097] Using the MAML framework described above, the same pre-trained model can be applied sequentially or simultaneously to four systems: lithium-ion battery energy storage, flow battery energy storage, flywheel energy storage, and compressed air energy storage, without the need to train a model from scratch for each system.

[0098] Step A4: A temporal graph attention network is used to dynamically learn the directed heat propagation path based on the thermal feature tensor, and output the heat source localization result and the heat spread prediction result.

[0099] Specifically, the Temporal Graph Attention Network (TGAT) is used to model the dynamic heat propagation relationships between different monitoring areas of an energy storage system.

[0100] First, a dynamic graph structure is constructed using each monitoring node in the energy storage system as a graph node and the thermal influence relationships between monitoring nodes as edges. The granularity of the monitoring nodes can be any of the following: individual battery cells, battery modules, or battery clusters, determined according to the actual monitoring accuracy requirements. For compressed air energy storage chambers, the monitoring nodes can be divided into spatial grids.

[0101] The feature vector of each graph node at each time step is composed of the temperature value, temperature gradient value, and temperature change rate value corresponding to the node in the thermal feature tensor output in step A2.

[0102] Then, for each time step, the temporal attention coefficient between each graph node and its neighboring nodes is calculated. Specifically, for the current graph node i and its neighboring node j, the feature sequence of node i within the historical time window is concatenated with the current features of node j and input into the attention function to obtain the attention weight. This attention weight serves as the weight of the directed edge from node j to node i, characterizing whether the heat from node j propagates to node i and the intensity of the propagation, thus modeling the direction of heat flow.

[0103] Next, the temperature change sequence of each monitoring node is backpropagated along the time dimension. Specifically, when the temperature or its rate of change at any monitoring node exceeds the warning threshold, the heat propagation tracing process is triggered. In the tracing process, the timestamps of the abnormal temperature rise at each monitoring node are checked sequentially in reverse chronological order, and the monitoring node with the earliest abnormal occurrence time is identified as the heat source location result.

[0104] Finally, based on the current temperature values ​​and trends of each monitoring node, and combined with the weights of the aforementioned directed edges, the heat transfer path is extrapolated forward along the direction of the directed edges to predict the estimated arrival time and peak temperature of heat spreading to adjacent monitoring nodes. This prediction allows maintenance personnel to know the direction and speed of heat spread before thermal runaway actually occurs.

[0105] In one embodiment, the heat propagation modeling is not limited to a temporal graph attention network, but can also be implemented using a graph convolutional recurrent neural network or a spatiotemporal graph convolutional network. In another alternative embodiment, heat source localization is not limited to the aforementioned backpropagation method along the time dimension, but can also be achieved by solving the inverse problem of the heat conduction equation (i.e., calculating the initial heat source location from the current temperature distribution).

[0106] Step A5: Output thermal management collaborative control commands based on the heat source location results and the heat spread prediction results.

[0107] In one embodiment, thermal management collaborative control commands are generated based on the heat source location results and heat spread prediction results output in step A4. The collaborative control commands include at least cooling flow commands for each cooling branch, fan speed commands, and pre-cooling commands.

[0108] The collaborative control is implemented by constructing a multi-agent deep reinforcement learning controller, treating each cooling branch in the energy storage system (e.g., the air-cooled or liquid-cooled branch corresponding to each battery cluster) as an agent. Each agent shares global state information (including the temperature of each monitoring node, the location of heat sources, and the predicted path of heat spread, etc.), and outputs the control actions of each agent through a strategy of centralized training and distributed execution.

[0109] In one embodiment, the reward function of the multi-agent deep reinforcement learning controller aims to minimize global total energy consumption and maximize overall temperature field uniformity. It also includes a penalty term for the differences in actions among the agents to suppress action conflicts arising from independent decision-making between cooling branches. Through this reward function design, each agent gradually forms a cooperative strategy during the learning process: increasing cooling intensity in advance in the heat source location area, pre-activating cooling along the predicted heat spread path, and maintaining low-power operation in unaffected areas, thereby achieving globally optimized allocation of cooling resources.

[0110] In one embodiment, the framework of the multi-agent deep reinforcement learning controller is not limited to QMIX or MAPPO, but can also employ algorithms such as MA-DDPG or COMA. In another alternative embodiment, the action difference penalty term in the reward function can be replaced with a mutual information maximization method or a mean-field game method. In yet another alternative embodiment, the control action is not limited to continuous control values ​​(such as precise flow rates), but can be discretized into multi-level control commands (e.g., dividing the cooling flow rate into low, medium, and high levels) to simplify the training and execution of the control policy.

[0111] Through the above steps A1 to A5, this embodiment realizes a complete closed-loop thermal management process from temperature sensing, anomaly detection, heat propagation tracing to collaborative control.

[0112] In one embodiment, a method for reconstructing a three-dimensional temperature field of a neural radiation field is also provided.

[0113] After outputting the heat source location results and heat spread prediction results in step A4, the following steps are also included:

[0114] Acquire sparse multi-angle infrared images simultaneously captured by infrared thermal imagers from at least two different perspectives deployed in the energy storage system, as well as the pose parameters and internal parameters (including focal length, principal point coordinates, etc.) of each infrared thermal imager.

[0115] The sparse multi-angle infrared image is input into a Neural Radiance Field (NeRF) network. The NeRF network takes the location coordinates of the sampling point in three-dimensional space and the observation direction vector as input, and outputs the temperature and voxel density of that sampling point. Specifically, for each ray of light emitted from the camera's optical center through an image pixel, dense sampling is performed along the ray in three-dimensional space. The location coordinates and direction vector of each sampling point are input into the NeRF network, and the network outputs the predicted temperature and voxel density of each sampling point.

[0116] Then, using voxel rendering integration technology, the temperature and voxel density of all sampling points on each ray are weighted and integrated to obtain the rendering temperature value at the corresponding pixel location of that ray. By optimizing the network parameters to minimize the error between the rendering temperature value and the actual acquired infrared image temperature value, a dense three-dimensional temperature voxel field is finally obtained.

[0117] The spatial resolution of the three-dimensional temperature voxel field is significantly higher than that of a single frame in the infrared image sequence, achieving super-resolution three-dimensional temperature field reconstruction. This technology is particularly suitable for solving the problem of blind zone monitoring in areas such as flywheel rotors and compressed air chambers where it is difficult to deploy dense sensors.

[0118] In one embodiment, the neural radiation field can be replaced by a Gaussian process regression method or a sparse voxel hashing method, which can also achieve super-resolution 3D temperature field reconstruction. For scenarios that do not require super-resolution reconstruction, a multi-view stereo matching algorithm can also be used directly for 3D temperature field reconstruction.

[0119] In one embodiment, a method for actively extrapolating extreme thermal runaway trajectories using generative adversarial networks is further provided.

[0120] After outputting the heat source location results and heat spread prediction results in step A4, the following steps are also included:

[0121] The extreme thermal runaway trajectory is actively inferred using a Generative Adversarial Network (GAN). Specifically, the current operating state parameters of the energy storage system (including current temperature distribution, state of charge, charging and discharging power, etc.) and the thermal feature tensor extracted in step A2 are obtained as the conditional inputs to the generator.

[0122] Under the control of the input conditions, the generator generates multiple sequences of thermal runaway evolution trajectories with different initial conditions and different evolution paths. These thermal runaway trajectory sequences include temperature-time curves and thermal propagation range-time curves for each monitoring node. A discriminator is used to distinguish between real thermal runaway trajectories (derived from actual historical fault records) and thermal runaway trajectories synthesized by the generator.

[0123] Through adversarial game training between the generator and the discriminator, the generator learns to produce synthetic trajectories that are statistically indistinguishable from real thermal runaway trajectories. These synthetic trajectories represent thermal runaway scenarios that have not yet actually occurred but are physically plausible, including two modes: rapid thermal runaway and slow thermal accumulation under extreme conditions.

[0124] After training, the confidence level output by the discriminator (i.e. the probability that the discriminator considers the current input to be a real fault trajectory) is used as a risk assessment indicator and fed back to the risk assessment model of the energy storage system. The risk assessment model is then subjected to adversarial training so that it can identify and respond to extreme fault scenarios that exceed the coverage of the training data, thereby improving the robustness of the model to rare but highly dangerous "black swan" events.

[0125] In one embodiment, the generative adversarial network can be replaced by a diffusion model or a variational autoencoder (VAE), which can also generate diverse thermal runaway trajectories. In another alternative embodiment, adversarial training can also employ virtual adversarial training instead of explicit generator-discriminator adversarial training.

[0126] In one embodiment, a method for intelligent diagnosis and counterfactual reasoning interaction based on large language models is further provided.

[0127] Following the output of the thermal management collaborative control command in step A5, the following steps are also included:

[0128] Access the Large Language Model (LLM) that has been fine-tuned by instructions from the energy storage field. Concatenate the heat source location results, heat spread prediction results, and current operating status parameters of the energy storage system output in step A4 into a prompt word, and input it into the Large Language Model.

[0129] Simultaneously, through a Retrieval-Augmented Generation (RAG) mechanism, historical failure cases and handling solutions matching the current anomaly type are retrieved from the knowledge graph. Specifically, the knowledge graph stores historical failure cases in a graph structure, including entities such as failure type, failure characteristic patterns, handling methods, and handling effects, as well as their relationships. Based on the characteristic patterns of the current anomaly (e.g., the rate and spatial distribution pattern of abnormal temperature rise), similarity matching is performed in the knowledge graph to retrieve the most similar historical cases.

[0130] The large language model, combined with the historical fault cases, automatically generates a fault diagnosis report in natural language format. The fault diagnosis report includes at least a description of the anomaly location, an anomaly type prediction, suggested handling measures, and an explanation of the expected effects.

[0131] In one embodiment, counterfactual reasoning interaction is also supported. Specifically, in response to user-input counterfactual hypothetical questions (e.g., "If the cooling flow rate of branch A is increased by 10%, how will the temperature in area B change?" or "If the current load is reduced to 80% of rated power, will heat spread be suppressed?"), the energy storage management platform performs one of the following operations:

[0132] One approach is to invoke a physical simulator, recalculate the temperature field evolution process of the energy storage system under the conditions specified in the counterfactual hypothesis query, and feed back the simulation results to the user in the form of natural language descriptions and / or temperature change trend graphs, which are then displayed through the user interface.

[0133] Another approach is to directly input the counterfactual hypothesis query into the large language model, which then generates and returns the expected result based on its pre-trained knowledge and the current context. This method offers faster response times but relatively lower accuracy, making it suitable for rapid decision-making scenarios.

[0134] In one embodiment, the large language model is not limited to LLaMA, but can also use GPT-4, ChatGLM, or Wenxin Yiyan, etc. In another alternative embodiment, the knowledge graph can be replaced by a vector database, and historical cases can be obtained through similarity retrieval instead of graph reasoning. In yet another embodiment, counterfactual reasoning can be implemented using only one of the two methods described above.

[0135] In one embodiment, a federated learning and adversarial defense approach is further provided.

[0136] This embodiment also includes a step of cross-site joint training using a federated learning architecture. Specifically, for multiple energy storage sites (e.g., multiple energy storage power stations located in different geographical locations), each site deploys a local model (including the model described in steps A2 to A4). Each site trains its local model using locally collected data and calculates the model gradient. Each site encrypts the gradient (e.g., using homomorphic encryption or a secure aggregation protocol) before uploading it to the central server, instead of uploading the original data.

[0137] The central server aggregates the encrypted gradients uploaded by each site (e.g., using the Federated Avg algorithm or the Federated Stochastic Gradient Descent algorithm), and distributes the aggregated global model parameters to each site. Each site then updates its local model parameters. Through this method, the original data of each site never leaves its local machine, achieving cross-site joint modeling while protecting data privacy and improving the model's generalization ability.

[0138] This embodiment also includes a security protection step using an adversarial attack detection module. Specifically, the input data of each modality is monitored in real time, and the local intrinsic dimension (LID) of each input sample in the feature space is calculated. Normal samples are usually distributed on a low-dimensional manifold in the feature space, with a low local intrinsic dimension; adversarial samples, due to malicious perturbation, have a significantly higher local intrinsic dimension than normal samples.

[0139] Therefore, when the calculated local intrinsic dimension exceeds the preset security threshold, the current input sample is determined to be an adversarial sample and is intercepted, and does not participate in the subsequent processing steps, thereby resisting malicious attacks against infrared images or control commands.

[0140] In one embodiment, adversarial attack detection is not limited to local intrinsic dimension methods, but can also be achieved using kernel density estimation or Mahalanobis distance methods.

[0141] In one embodiment, combined Figure 2 This provides the implementation steps for another thermal management method for energy storage systems:

[0142] S1: Deploy infrared thermal imaging arrays and multimodal sensor networks in the energy storage system. Different deployment configurations are used depending on the energy storage technology type: multiple infrared imaging temperature sensors are installed along sensor tracks inside the lithium-ion battery container; infrared probes are arranged around the outer periphery of the flow battery stack; infrared thermal imaging inspection devices are positioned on the outside of the flywheel housing, aligned with the motor rotor end; and infrared thermal imaging inspection devices are deployed circumferentially along the inner wall of the compressed air underground chamber. The multimodal sensor network includes visible light cameras and gas sensors (detecting CO, etc.). (VOCs, etc.) and vibration sensors.

[0143] S2: Construct a hybrid model of physical information neural network and spatiotemporal Transformer. (Reference) Figure 3 The model employs a 3D convolutional neural network at the front end to extract short-term local spatiotemporal features, and a spatiotemporal Transformer module at the back end. Its innovation lies in applying the two-dimensional unsteady-state heat conduction equation... Residual term embedding loss function: ,in To predict the mean square error of temperature and infrared observations, This represents the mean square error of the residuals in the heat conduction equation. The network output is constrained by the physical equations in the unobserved region, eliminating predictions that violate the laws of thermodynamics.

[0144] S3: Construct an adaptive identification model for cross-energy storage systems based on four-modal contrastive learning pre-training and meta-learning. (Reference) Figure 4 Utilizing a contrastive learning framework, infrared thermal images, visible light edge images, gas concentration spectrum encoding vectors, and vibration spectrum encoding vectors from the same time and spatial location are used as positive sample pairs, while encoding vectors from different locations or times are used as negative sample pairs. An InfoNCE loss model is used to train a multimodal encoder, enabling normal operating condition features to cluster tightly in a common embedding space. During deployment, zero-sample anomaly detection (requiring no fault samples) is achieved by calculating the Mahalanobis distance between real-time multimodal features and normal cluster centers. Simultaneously, a Model-Independent Meta-Learning (MAML) framework is employed for meta-training on a dataset containing four energy storage systems. This allows the model to quickly adapt to any energy storage system with a small number of samples, outputting corresponding anomaly features (lithium battery: hot spot location, temperature uniformity index; flow battery: electrolyte leakage location, stack temperature distribution; flywheel: rotor temperature, bearing anomalies; compressed air: seal leakage location).

[0145] S4: Construct a 3D temperature field super-resolution reconstruction and heat propagation source tracing model that integrates a Temporal Graph Attention Network (TGAT) and a Neural Radiation Field (NeRF). Reference Figure 5 and Figure 6 TGAT treats each monitoring unit (single battery, module, stack) as a graph node, with node characteristics being temperature and temperature change rate. It dynamically calculates the directed attention coefficient through temporal and spatial attention mechanisms. , representing the heat flow intensity from node j to node i. By tracing the heat flow path matrix back, the earliest heat source node exhibiting anomalies is automatically identified, and potentially overheated node regions within the next Δt time interval are predicted. Simultaneously, NeRF learns a continuous three-dimensional temperature field function using sparse infrared images from multiple angles and their corresponding camera poses as input. High-resolution three-dimensional temperature voxel field can be reconstructed through volume rendering technology, with spatial resolution reaching 5 to 10 times that of the sensor's physical resolution.

[0146] S5: Construct an anomaly scenario simulation and adversarial risk assessment model based on Generative Adversarial Networks (GANs). (Reference) Figure 7The generator G, based on a conditional variational autoencoder architecture and using historical normal temperature field sequences as conditions, samples latent variables to generate multiple possible future thermal runaway evolution trajectories. The discriminator D distinguishes between real fault trajectories and generated trajectories. After training, the generator is used to: ① generate a large number of physically plausible but non-occurring extreme fault scenarios to expand the training set; ② conduct adversarial attack tests on the risk assessment model to screen for and strengthen vulnerabilities; ③ offline simulation of the effectiveness of different thermal management strategies under extreme scenarios. The risk assessment engine integrates TGAT output, NeRF temperature field, gas concentration, vibration signals, and adversarial samples provided by the generator to output the comprehensive safety risk level (Level I, Level II, Level III) and anomaly type based on adversarial robustness.

[0147] S6: Constructing a multi-agent deep reinforcement learning collaborative thermal management controller. (Reference) Figure 8 Each independent cooling branch (such as a liquid cooling circuit or fan assembly) is defined as an agent, and a centralized training distributed execution framework (such as QMIX or MAPPO) is used. The state space consists of the dimensionality-reduced temperature field characteristics of the area covered by each agent and the risk level of global broadcast; the action space consists of continuous quantities: coolant flow rate (0 to rated maximum), fan speed (0 to maximum speed), and pre-cooling switch. The reward function is designed as follows: The third item is a penalty for differences in actions between adjacent agents, used to encourage spatial cooperation. After training, each agent runs in real time on its local PLC to achieve global optimized control.

[0148] S7: An interactive fault diagnosis system enhanced with Large Language Model (LLM) and Knowledge Graph (KG) in the energy storage field. (Reference) Figure 9 The knowledge graph uses a basic triplet structure of "fault phenomenon - fault cause - handling measures - spare parts" and contains over 50,000 expert knowledge entries. When an anomaly is detected, the system first outputs a heat propagation map and a three-dimensional temperature field via TGAT and NeRF. Then, it retrieves relevant subgraphs from the knowledge graph (KG), constructs structured prompts by combining the subgraphs with real-time monitoring data, and inputs them into the LLaMA-3-8B model, which has been fine-tuned using energy storage field instructions, to generate a diagnostic report containing a chain of evidence. The system supports counterfactual reasoning interaction: maintenance personnel can propose hypothetical operations in natural language (e.g., "If the cooling flow rate in area B is increased by 10%, how will the temperature change?"). The system then calls a simplified heat-flow simulation model at the edge to quickly perform the simulation, and the LLM returns the simulation results in natural language.

[0149] Furthermore, this invention also includes an edge-cloud collaborative federated learning and adversarial defense mechanism. (Reference) Figure 10Each energy storage site's edge AI computing unit updates model parameters locally, uploading only homomorphically encrypted or securely aggregated gradients to the cloud server; the cloud maintains the global model and distributes updates. An adversarial example detection module based on Local Intrinsic Dimension (LID) is deployed at the edge to identify and block malicious attacks targeting infrared images or sensor data.

[0150] In one embodiment, a system for implementing the above method is also provided, comprising:

[0151] Infrared thermal imaging array and multimodal sensor network;

[0152] Edge AI computing unit: integrates a physical information neural network Transformer inference engine, a temporal graph attention network inference engine, a neural radiation field reconstruction engine, and a lightweight zero-shot anomaly detection module; built-in Trusted Execution Environment (TEE) protects data privacy;

[0153] Generative Adversarial Simulation Server: Used for offline training of abnormal scenario generators;

[0154] Multi-agent thermal management controller: Deployed in a distributed PLC cluster, each agent exchanges collaborative information through Time-Sensitive Network (TSN) or 5G communication;

[0155] Large Language Model and Knowledge Graph Interaction Platform: Deployed on a cloud server, it includes a retrieval enhancement generation module and a physical simulator call interface;

[0156] Federated Learning and Security Defense Server: Used to aggregate gradients from various sites, distribute global models, and distribute adversarial defense rules.

[0157] Compared with the prior art, the present invention has the following beneficial effects:

[0158] 1. Significantly improved physical consistency: The physical information neural network uses the heat conduction equation as a soft constraint, reducing the prediction error by 58% in sparse sensor regions, completely eliminating predictions that violate the laws of thermodynamics, and enhancing the reliability of the model.

[0159] 2. Zero-sample anomaly detection capability: Four-modal contrastive learning enables the system to detect unknown anomaly types without any fault samples, and its generalization ability for novel thermal runaway modes far exceeds that of existing supervised methods.

[0160] 3. Visualization and active blocking of heat propagation: The time-series graph attention network outputs a directed heat propagation map, which can accurately locate heat sources (with single-unit accuracy) and predict heat spread paths, providing spatial targets for active cooling and increasing the early warning lead time to 5-10 minutes.

[0161] 4. Significant energy savings through collaborative thermal management: Multi-agent deep reinforcement learning enables global collaboration of cooling branches. Compared with independent PID or single-agent DDPG, total energy consumption is reduced by 20-25%, and temperature uniformity is improved by 35%.

[0162] 5. Super-resolution three-dimensional temperature field reconstruction: Neural radiation field technology can reconstruct a high-resolution three-dimensional temperature voxel field using a small number of infrared images, improving spatial resolution by 5 to 10 times. It is especially suitable for monitoring the temperature inside flywheel rotors and deep layers of compressed air chambers.

[0163] 6. High coverage of extreme scenarios: Generative adversarial networks actively simulate non-occurring thermal runaway trajectories, increasing the coverage of risk assessment models for "black swan" events from less than 30% to over 95%.

[0164] 7. Natural Language Interaction and Counterfactual Reasoning: The large language model combined with knowledge graphs not only outputs expert-level diagnostic reports, but also supports operations and maintenance personnel to make decision-making deductions based on "what if..." through natural language, significantly improving the efficiency of human-machine collaboration.

[0165] 8. Data privacy and security: Federated learning and homomorphic encryption enable cross-site joint modeling without leaking original data; local intrinsic dimension adversarial detection can identify more than 99% of white-box / black-box attacks, ensuring system robustness.

[0166] In one embodiment, a complete application in a lithium-ion battery energy storage container is provided.

[0167] A 5MWh lithium-ion battery energy storage container contains 16 battery clusters, each cluster consisting of 12 battery modules connected in series. Figure 11 Deployment: A mobile sensor track is installed on the top of the container, with six infrared imaging temperature sensors (wavelength 8–14 μm, temperature measurement accuracy) mounted on the track. (30Hz frame rate) to achieve multi-angle scanning; simultaneously deploy 4 gas sensors (detecting CO, ... It has 8 vibration sensors (triaxial MEMS, sampling rate 10kHz) and 2 visible light cameras.

[0168] Step S2 implementation: The edge AI computing unit (NVIDIA Jetson Orin, 32GB memory) receives infrared image sequences at a frequency of 10Hz. The Physical Information Neural Network-Transformer model is loaded with pre-trained weights, and the loss function is set to λ=0.1. In temperature prediction on the back of a certain battery cluster (where the infrared sensor is obstructed), the error between the pure Transformer model's predicted value and the subsequent measured value after shutdown is 4.8℃, while the error of the PINN-Transformer model of this invention is 1.9℃, and the gradient direction of the predicted temperature field always points towards the known heat source.

[0169] Step S3 Implementation: The four-modal contrastive learning encoder has been pre-trained on a source domain containing 100,000 sets of multimodal data. During deployment, the system calculates the Mahalanobis distance between real-time multimodal features and normal cluster centers in an unsupervised manner. When a slight internal short circuit occurs in a cluster, the gas concentration only increases by 2 ppm, the vibration amplitude increases by 5%, and the temperature change is not significant. The Mahalanobis distance gradually increases from 0.8 to 2.5 (threshold 1.8), and the system issues an "unknown anomaly" warning 8 minutes in advance, while the traditional temperature threshold method does not trigger an alarm at this time. The meta-learning module automatically identifies the system as a lithium-ion battery system based on 20 frames of infrared images during the warm-up period, activating the corresponding hot spot detection branch.

[0170] Step S4 Implementation: The time-series graph attention network constructs a graph structure using 192 battery modules as nodes (one node per module). At a certain moment, the model outputs a directed attention coefficient matrix, showing that the heat flux coefficient of module B-12→B-13 is 0.72 (direction clear), confirming B-12 as the original heat source. Simultaneously, the NeRF module reconstructs the three-dimensional temperature field using infrared images from five different angles (with pose provided by orbital movement), successfully identifying that the temperature of the third cell inside module B-12 is 8.2℃ higher than the average temperature of the module, achieving a positioning accuracy at the cell level.

[0171] Step S5 Implementation: The adversarial simulation server has completed offline training. For the currently detected hotspot, the generator conditionally generates 10 possible thermal runaway evolution trajectories within the next 15 minutes, of which two trajectories are predicted to spread to two adjacent modules after 12 minutes. The risk assessment engine incorporates this information into the risk level calculation, outputting a Level II warning and providing a propagation time prediction.

[0172] Step S6 Implementation: The multi-agent controller comprises 16 liquid-cooled branch agents (one branch corresponding to each battery cluster). The current hot spot is located in cluster B. Upon receiving a Level II warning, its agent increases the local flow rate from 2L / min to 5L / min. Simultaneously, due to the action difference penalty, the agents in adjacent clusters A and C automatically increase their flow rates from 2L / min to 2.5L / min (pre-coordination). The total cooling power consumption increases by only 18%, while under traditional independent control, the flow rate in cluster B increases to 6L / min, while adjacent clusters remain unchanged, resulting in a 30% decrease in the final temperature uniformity index.

[0173] Step S7 implementation: The large language model receives structured data: "Anomaly type: hot spot; Location: 3rd cell of module 12 in cluster B; Risk level: Level II; Temperature trend: 6.2℃ increase in the past 10 minutes; CO concentration: 8ppm; Heat propagation prediction: spread to B-13 and B-14 in 12 minutes." Based on the knowledge graph retrieval results, a report was generated: "Hot spot in cell 3 of module B-12, temperature rise rate 0.62℃ / min, exceeding the safety threshold of 0.4℃ / min. Possible causes: internal micro-short circuit or increased contact resistance at connection points in this cell. Recommendations: 1) Immediately increase the cooling flow rate of cluster B to 5L / min, and pre-cool adjacent clusters to 2.5L / min; 2) Schedule an infrared re-inspection and internal resistance test for this module within 24 hours; 3) Spare parts: cell module model XYZ. Should we perform a counterfactual simulation?" The maintenance personnel asked: "What will happen if I do not increase the flow rate of adjacent clusters?" The system invoked the simulator to simulate: without pre-cooling of adjacent clusters, module B-13 will reach 75℃ 8 minutes after the hot spot spreads. The model returned the results in natural language, and the maintenance personnel made a decision to implement coordinated cooling based on this.

[0174] Federated Learning and Security Defense: Every 24 hours, the edge devices of this energy storage site upload the model gradients to the cloud federated server after homomorphic encryption, and update the global model together with 30 other sites. During this period, the edge LID detection module successfully identified an adversarial perturbation targeting the infrared image (LID value 0.31, below the threshold of 0.45), automatically blocked it, and marked the sample as an attack.

[0175] In one embodiment, the method is used for leak detection in an underground compressed air energy storage tunnel.

[0176] A 300MW compressed air energy storage underground chamber has an inner diameter of 8m, and its sealing layer is made of polymer composite material. (According to...) Figure 12 Deployment: One infrared thermal imaging inspection device is installed every 15° along the circumference of the inner wall of the chamber, totaling 24 units, capable of pitch and rotation. When a minor leak occurs at a certain point of the seal (leakage rate approximately 1 / 15 of the design limit), the Joule-Thomson effect generates a localized cooling zone. The time-series attention network divides the sealing layer into 360 nodes (one per 1°), identifying abnormal heat flow patterns at the leak point through directed attention coefficients. The neural radiation field reconstructs a three-dimensional temperature field using 24 angular images, precisely locating the leak point coordinates (175° longitude, 38° latitude). A large language model generates a report recommending depressurization and scheduling repairs. Leaks are detected 52 hours earlier than traditional pressure drop monitoring methods.

[0177] In one embodiment, the method is used for early warning of flywheel energy storage rotor overheating.

[0178] A 500kW flywheel energy storage system operates at a flywheel speed of 36,000 rpm. An infrared probe is positioned over the rotor endplate, while a vibration sensor monitors the bearing. A zero-sample anomaly detection module, without any fault data, detected a 0.5 harmonic component (early bearing wear characteristic) in the vibration spectrum using a contrastive learning encoder. Simultaneously, the clustering distance of the infrared temperature sequence increased from the normal 0.9 to 2.1. A time-series attention network determined the heat source to be the bearing and predicted a heat transfer time of approximately 15 minutes to the rotor's permanent magnets. A multi-agent controller (with a single agent) implemented reduced power operation (from 500kW to 350kW) and increased the vacuum pump speed to enhance heat dissipation. The system successfully prevented the permanent magnets from overheating and demagnetizing.

[0179] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A thermal management method for an energy storage system, characterized in that, include: Acquire the infrared image sequence acquired by the infrared thermal imaging array in the energy storage system; A hybrid model combining physical information neural network and Transformer is constructed, and the partial differential equation of heat conduction is used as a soft-constraint embedding loss function to extract thermal feature tensors that conform to the laws of thermodynamics from the infrared image sequence. By using four-modal contrastive learning pre-training, multiple modal features, including at least infrared modes, are mapped to a common embedding space to perform zero-sample anomaly detection without fault samples, and model-independent meta-learning is used to adapt to the current energy storage system type. A temporal graph attention network is used to dynamically learn a directed heat propagation path based on the thermal feature tensor, and outputs heat source localization results and heat spread prediction results. Based on the heat source location results and the heat spread prediction results, output thermal management collaborative control commands.

2. The method according to claim 1, characterized in that, The specific method of using the heat conduction partial differential equation as a soft constraint embedding loss function is as follows: A residual term of the heat conduction equation is added to the overall loss function of the hybrid model. This residual term is used to constrain the temperature field prediction results output by the network to satisfy the partial differential equation of heat conduction.

3. The method according to claim 1, characterized in that, The process of mapping multiple modal features, including at least the infrared modality, to a common embedding space through four-modal contrastive learning pre-training is specifically as follows: Data from four sensors—infrared mode, visible light mode, gas mode, and vibration mode—are acquired simultaneously, and feature vectors are extracted from each of the four channels. Using samples from the same modality at different times in the four feature vectors as negative sample pairs and samples from different modalities at the same time as positive sample pairs, the feature distance between positive sample pairs is brought closer and the feature distance between negative sample pairs is pushed apart by the comparison loss function, and the four feature vectors are mapped to the common embedding space. After mapping the feature vector of the sample to be detected to the common embedding space, the distance between it and the cluster center of the normal sample features is calculated. When the distance exceeds a preset threshold, it is determined to be abnormal.

4. The method according to claim 1, characterized in that, The method of using model-independent element learning to adapt to the current energy storage system type specifically includes: Acquire historical data from at least two of the following systems: lithium-ion battery energy storage system, flow battery energy storage system, flywheel energy storage system, and compressed air energy storage system; and learn general initialization parameters during the meta-training phase. During the meta-testing phase, the general initialization parameters are updated one or more times using a small number of labeled samples of the current energy storage system type to obtain model parameters that are adapted to the current energy storage system type.

5. The method according to claim 1, characterized in that, The use of a temporal graph attention network to dynamically learn directed heat propagation paths based on the thermal feature tensor is specifically as follows: A dynamic graph structure is constructed using each monitoring node in the energy storage system as a graph node and the thermal influence relationship between each monitoring node as an edge. For each time step, the temporal attention coefficient between the current monitoring node and its neighboring monitoring nodes is calculated, and the temporal attention coefficient is used as the weight of the directed edge to represent the direction of heat flow. Backpropagation is performed on the temperature change sequence of each monitoring node along the time dimension to trace the earliest starting monitoring node where the temperature anomaly occurred as the heat source location result. Based on the current temperature and its changing trend of each monitoring node, and combined with the weight of the directed edge, the arrival time and temperature peak of heat spread to adjacent monitoring nodes are predicted.

6. The method according to claim 1, characterized in that, The output thermal management collaborative control command is specifically as follows: Construct a multi-agent deep reinforcement learning controller, treating each cooling branch in the energy storage system as an agent, with each agent sharing global state information; With the optimization objectives of minimizing total global energy consumption and maximizing overall temperature field uniformity, the control actions of each agent are output through centralized training and distributed execution strategies. The thermal management coordinated control commands include cooling flow commands for each cooling branch, fan speed commands, and pre-cooling start commands.

7. The method according to claim 6, characterized in that, The reward function R of the multi-agent deep reinforcement learning controller is: ； in, The total energy consumption of all cooling branches. Let be the root mean square deviation of the temperature at each monitoring point within the energy storage system. For the action of the i-th agent, The average value of all agent actions. , , The weighting coefficients are used; the reward function is used to jointly optimize the minimization of total global energy consumption, the maximization of overall temperature field uniformity, and the minimization of the differences in actions of each agent.

8. The method according to claim 1, characterized in that, Following the output heat source location results and heat spread prediction results, the following is also included: Acquire sparse multi-angle infrared images simultaneously captured by infrared thermal imagers from at least two different perspectives deployed in the energy storage system, as well as the pose parameters and internal parameters of each infrared thermal imager. The sparse multi-angle infrared image is input into the neural radiation field network. The pose parameters and internal parameters are used as light sampling parameters. Sampling is performed along each light ray in three-dimensional space. The position coordinates and direction vector of the sampling point are input into the neural radiation field network. The temperature and voxel density of each sampling point are output. A dense three-dimensional temperature voxel field is generated through voxel rendering integration. The spatial resolution of the three-dimensional temperature voxel field is higher than that of a single frame image in the infrared image sequence.

9. The method according to claim 1, characterized in that, Following the output thermal management collaborative control command, the system also includes: The large language model, which has been fine-tuned by instructions from the energy storage field, is accessed. The heat source location result, the heat spread prediction result, and the current operating status parameters of the energy storage system are concatenated into prompt words and input into the large language model. By using a retrieval-enhanced generation mechanism, historical fault cases matching the current anomaly type are retrieved from the knowledge graph. The large language model then combines these historical fault cases to generate and output a natural language fault diagnosis report. In response to a counterfactual hypothesis query input by the user, the system invokes a physical simulator to simulate the temperature evolution process of the energy storage system under the conditions specified in the counterfactual hypothesis query, or the counterfactual hypothesis query is directly input into the large language model for it to infer and generate the expected result, and the expected result is fed back to the user.

10. The method according to claim 1, characterized in that, Also includes: A federated learning architecture is adopted. The local model training gradients of each energy storage site are encrypted and then uploaded to the central server. The central server aggregates the encrypted gradients of each site and distributes the aggregated global model parameters to each site for local updates. An adversarial attack detection module is used to monitor the input data of each modality in real time, calculate the local intrinsic dimension of each input sample in the feature space, and when the local intrinsic dimension exceeds a preset security threshold, the current input sample is determined to be an adversarial sample and is blocked.

Images