Hybrid energy storage power station operation and maintenance system and method based on energy storage operation and maintenance large model

Abstract

The application discloses a kind of based on energy storage operation and maintenance large model hybrid energy storage power station operation and maintenance system and method, belong to electrochemistry energy storage operation and maintenance technical field.The method includes: data processing module, knowledge graph module, model training module, adaptive deployment module and inference engine module;Data processing module is used to obtain multi-source data and carries out semantic alignment, generates triadic data;Knowledge graph module is used to construct triadic-time sequence hybrid knowledge graph;Model training module is used to construct energy storage operation and maintenance large model;And to energy storage operation and maintenance large model is compressed, and extraction low rank adaptation parameter;Adaptive deployment module is used for local deployment and fault prediction;Inference engine module is used to generate four-dimensional time sequence, and combined with the energy storage operation and maintenance large model of local deployment carries out interpretable inference and fault prediction.The application based on energy storage operation and maintenance large model hybrid energy storage power station operation and maintenance system and method improve applicability, reliability and operation and maintenance precision.

Description

Technical Field

[0001] This application belongs to the field of electrochemical energy storage operation and maintenance technology, and particularly relates to the operation and maintenance system and method, storage medium and equipment of hybrid energy storage power station based on a large-scale energy storage operation and maintenance model. Background Technology

[0002] Currently, electrochemical energy storage power stations are rapidly evolving from a single construction model of "single manufacturer, unified model" to a hybrid model of "multiple manufacturers, multiple models, and multiple operating conditions". In scenarios such as grid-side megawatt-hour-level power stations, industrial and commercial "photovoltaic-storage-charging" parks, and new energy distribution and storage bases, battery compartments, energy storage converters (PCS), air conditioning, and fire protection systems from different brands are often connected simultaneously.

[0003] For the operation and maintenance of such complex systems, existing technologies mainly adopt the following solutions: 1. Data Layer Solution: Each manufacturer independently deploys its Battery Management System (BMS) and Energy Management System (EMS), collecting time-series data according to their own proprietary protocols (such as CAN, Modbus, IEC 61850-MMS, etc.). This data is converted to CSV or XML format via a gateway and then uploaded to a local Supervisory Control and Data Acquisition (SCADA) system or a cloud-based big data platform. Meanwhile, unstructured records from the maintenance process are still scattered across each manufacturer's after-sales systems in the form of PDFs, audio files, etc., lacking a unified semantic standard.

[0004] 2. Knowledge Layer Solution: The mainstream approach is to establish a two-dimensional relationship database of "device-alarm" in the cloud and use machine learning algorithms such as expert rules or random forests to generate labels for single-point faults. A few solutions introduce static knowledge graphs, but these are limited to constructing triples from structured information such as device ledgers, alarm codes, and maintenance manuals. The edge weights in this type of graph are fixed and do not contain a time dimension, nor can they express the non-linear aging chain of batteries.

[0005] 3. Model Layer Solution: Lightweight machine learning models, such as XGBoost or one-dimensional convolutional neural networks (1-D CNN), are commonly used at the field end for estimating battery state of charge (SOC) and state of health (SOH). These models are limited to measurement sequences provided by a single vendor. When applied to cross-site scenarios, data needs to be collected again and the model retrained offline, a process that can take 2 to 4 weeks. Recently, some vendors have attempted to introduce a general-purpose language model with 1 billion parameters for work order question answering, but this model is completely disconnected from real-time time-series data, making closed-loop control impossible.

[0006] 4. Inference and Deployment Scheme: BMS and EMS only provide threshold over-limit alarms. The digital twin on the EMS side is mostly a "display-level" 3D visualization model, and its boundary conditions need to be manually set. When a new power plant is expanded, engineers need to recalibrate the system parameters on-site, lacking an online update mechanism.

[0007] Because the existing technical solutions mentioned above are still at the stage of "single-point alarm, offline modeling, and centralized training", the following problems urgently need to be solved in the operation and maintenance of hybrid energy storage power stations: (1) Data format fragmentation: Each manufacturer's communication protocol is proprietary and lacks a unified semantic registry, which means that the platform needs to write independent parsing scripts for each type of device. Connecting to a hybrid power station containing devices from multiple manufacturers often takes 3 to 6 months, making it impossible to achieve "plug and play" for new devices.

[0008] (2) Knowledge fragmentation: Static knowledge graphs lack time decay factors and cannot depict aging chains that lag by several months, such as "temperature rise → SEI film growth → internal resistance increase → thermal runaway". At the same time, the edge weights in the rule base are fixed, resulting in opaque cascading failure paths, and maintenance personnel can only perform post-event analysis and remediation.

[0009] (3) Model updates are cumbersome and slow to be implemented: Cross-site data collection requires the full amount of data to be collected again and the model to be retrained offline, which causes the model parameter size to expand linearly with the number of connected vendors. In addition, the general large model is not integrated with the real-time time series signal, and the inference results lack an interpretable causal graph, making it difficult for on-site operation and maintenance personnel to trust and directly perform key operations such as switching machines or reducing power.

[0010] (4) Privacy and compliance bottlenecks: The existing cloud-based centralized training model requires manufacturers to upload data in plaintext, which involves sensitive information such as battery process parameters. This leads to manufacturers refusing to cooperate due to data security concerns, forming "data silos" and making it difficult to maintain the accuracy of the model due to insufficient training samples.

[0011] (5) Poor edge real-time performance: The computing power of traditional 32-bit BMS chips is usually less than 1 TOPS, which makes it difficult to support the local deployment of floating-point models. If all sampled data is sent back to the cloud for processing, the sampling stream of hundreds of megabytes per second will have a delay of more than 500 milliseconds under the 4G link, which cannot meet the requirements of millisecond-level shutdown window.

[0012] In summary, current technologies still face five major challenges for hybrid energy storage power stations: data fragmentation, knowledge fragmentation, model retraining, privacy compliance, and lack of transparency in cascading failures. These challenges make it difficult to support the urgent needs of new power systems for highly reliable, adaptive, and explainable operation and maintenance. Summary of the Invention

[0013] To address the aforementioned issues, this application provides a hybrid energy storage power station operation and maintenance system based on a large-scale energy storage operation and maintenance model, which improves the system's applicability, reliability, and operation and maintenance accuracy.

[0014] This application is achieved through the following technical solution: An operation and maintenance system for a hybrid energy storage power station based on a large-scale energy storage operation and maintenance model includes: a data processing module, a knowledge graph module, a model training module, an adaptive deployment module, and an inference engine module; The data processing module includes a global data lake and an energy storage field model registration center, which is used to acquire multi-source data, perform semantic alignment on the multi-source data, and generate triple data. The knowledge graph module is used to construct a triple-time series hybrid knowledge graph based on the triple data; The model training module is used to construct a large-scale energy storage operation and maintenance model based on the multi-source data and the triple-time series hybrid knowledge graph; and to compress and quantize the large-scale energy storage operation and maintenance model to obtain a lightweight large-scale energy storage operation and maintenance model, and to extract the low-rank adaptation parameters of the large-scale energy storage operation and maintenance model. The adaptive deployment module includes: a federated learning coordinator and an edge inference box, used for local deployment based on the lightweight energy storage operation and maintenance large model and low-rank adaptation parameters, combining local data for fault prediction, and outputting fault probability and feature vector; as well as updating the energy storage operation and maintenance large model and its low-rank adaptation parameters. The inference engine module includes: a causal discovery unit, a thermal-electric coupling model, and a graph neural ODE solver, used to generate a four-dimensional time series based on the multi-source data, and combine it with the locally deployed energy storage operation and maintenance large model to perform interpretable inference and fault prediction, and output the prediction results.

[0015] Optional, The data processing module is also configured to: Acquire multi-source data and store it in a global data lake. The multi-source data includes: multi-source time-series data and multi-source unstructured data. According to the energy storage field model registration center, the multi-source data is semantically aligned to generate triple data with timestamps.

[0016] Optional, The knowledge graph module is also configured to: Define node types, edge types, and weights; Nodes and edges are stored in a graph database to form a queryable network, thus constructing a knowledge graph. Temporal embedded knowledge graphs employ temporal relation graph convolutional networks to add a time decay function to each edge, forming a triplet-temporal hybrid knowledge graph.

[0017] Optional, The model training module also includes: A multimodal encoder, including a timing encoder, a text / speech encoder, and an image encoder, is used to uniformly vectorize the multi-source data.

[0018] Optional, The model training module is also configured to: For the four modalities of time series, text, speech, and image data in the multi-source data, corresponding pre-trained encoders are set up to transform the multi-source data into feature vectors in a unified manner; A predetermined amount of corpus is generated based on the triplet-temporal hybrid knowledge graph, and a domain model is obtained by distillation of the feature vectors through a general large model. A large-scale energy storage operation and maintenance model is trained by weighting triplet loss and masked language model loss according to a preset ratio. The large-scale energy storage operation and maintenance model is compressed and quantized to obtain a lightweight large-scale energy storage operation and maintenance model, and the low-rank adaptation parameters of the large-scale energy storage operation and maintenance model are extracted.

[0019] Optional, The adaptive deployment module includes: an edge inference box and a federated learning coordinator. The federated learning coordinator is used for central-side parameter aggregation and differential privacy policy management. The edge inference box is used to perform local deployment based on the lightweight energy storage operation and maintenance model and low-rank adaptation parameters to obtain a local lightweight energy storage operation and maintenance model; based on local data, it performs fault prediction through the local lightweight energy storage operation and maintenance model and outputs fault probability and feature vector.

[0020] Optional, The federated learning coordinator is used for central-side parameter aggregation and differential privacy policy management, including: Lightweight energy storage operation and maintenance large model and low-rank adaptation parameters are issued to the edge inference box; Based on the incremental low-rank adaptation parameters uploaded by each edge inference box, the incremental low-rank adaptation parameters of multiple sites are aggregated to generate a new large-scale energy storage operation and maintenance model and new low-rank adaptation parameters. The new energy storage operation and maintenance model is compressed and quantized to obtain an updated lightweight energy storage operation and maintenance model. The new low-rank adaptation parameters are then sent back to the edge inference box to realize the dynamic iterative update of the energy storage operation and maintenance model.

[0021] Optional, The edge inference box is also configured to: Based on the lightweight energy storage operation and maintenance model and low-rank adaptation parameters, complete the deployment of the local lightweight energy storage operation and maintenance model. Using local data, the low-rank adaptation parameters are incrementally fine-tuned using AdaLoRA to complete the deployment and update of a local, lightweight energy storage operation and maintenance model. Based on a local, lightweight energy storage operation and maintenance model, the model outputs fault probability and feature vectors, and works in conjunction with the inference engine module to perform fault prediction.

[0022] Optional, The inference engine module includes: a causal discovery unit, a thermal-electric coupling model, and a graph neural ODE solver. The causal discovery unit is used to generate a four-dimensional time series sequence based on the multi-source data; and to establish a directed acyclic graph based on the four-dimensional time series sequence, and update the nodes at a preset period. A thermo-electric coupling model is used to perform thermo-electric coupling simulations using the directed acyclic graph as boundary conditions and the feature vectors output by the local lightweight energy storage operation and maintenance model, and output simulation data; simulation parameters are calibrated based on the simulation data and real-time sensor data from the field. The graph neural network ODE solver is used to propagate the fault probability / feature vector output by the directed acyclic graph, simulation data, and the local lightweight energy storage operation and maintenance large model. It propagates the fault probability on the subgraph of the triple-time series hybrid knowledge graph through the graph neural network ODE, integrates causal constraints, physical simulation, and large model probability, performs interpretable reasoning and fault prediction, and outputs the prediction results.

[0023] Optional, The system also includes: a data quality governance unit, an infrared image super-resolution preprocessing unit, a knowledge verification unit, an incremental expert head hot-swappable base, and a watchdog rollback unit.

[0024] This application also provides a hybrid energy storage power station operation and maintenance method based on a large-scale energy storage operation and maintenance model, the method comprising: Acquire multi-source data, perform semantic alignment on the multi-source data, and generate triplet data; Construct a triplet-temporal hybrid knowledge graph based on the triplet data; A large-scale energy storage operation and maintenance model is constructed based on the multi-source data and the triple-time series hybrid knowledge graph; the large-scale energy storage operation and maintenance model is compressed and quantized to obtain a lightweight large-scale energy storage operation and maintenance model, and the low-rank adaptation parameters of the large-scale energy storage operation and maintenance model are extracted. Based on the lightweight energy storage operation and maintenance model and low-rank adaptation parameters, local deployment is carried out, and fault prediction is performed by combining local data to output fault probability and feature vector. A four-dimensional time series is generated based on the multi-source data, and interpretable reasoning and fault prediction are performed in conjunction with the locally deployed energy storage operation and maintenance model, and the prediction results are output.

[0025] This application also provides a computer-readable storage medium storing one or more programs, which, when executed, can implement the aforementioned operation and maintenance method for hybrid energy storage power stations based on a large-scale energy storage operation and maintenance model.

[0026] This application also provides a device, including a processor, a communication interface, a computer-readable storage medium, and a communication bus; wherein the processor, the communication interface, and the computer-readable storage medium communicate with each other through the communication bus; The processor is used to execute programs stored in a computer-readable storage medium.

[0027] Compared with the prior art, this application has the following advantages: 1. The hybrid energy storage power station operation and maintenance system method based on a large-scale energy storage operation and maintenance model proposed in this application uses a global data lake and an energy storage domain pattern registration center to semantically align multi-source data, generate triple data, and construct a triple-time series hybrid knowledge graph. Based on the multi-source data and the triple-time series hybrid knowledge graph, a large-scale energy storage operation and maintenance model is obtained using federated knowledge distillation. Based on the compressed large-scale energy storage operation and maintenance model, local deployment is carried out at the edge, and interpretable reasoning and fault prediction are performed, which improves the access efficiency, operation and maintenance accuracy, fault early warning transparency, and system robustness of the hybrid energy storage power station.

[0028] 2. By constructing a global, multi-source, heterogeneous data lake and a schema registry for the energy storage field, plug-and-play and semantic unification of devices from multiple vendors are achieved. Through a triple-time-series hybrid knowledge graph, unstructured O&M experience is transformed into computable and decaying causal edges, explicitly characterizing the nonlinear aging chain of batteries and the coupling relationship between devices. A domain-enhanced energy storage O&M model is trained using a multimodal general model and KG-T federated knowledge distillation, and local deployment at the edge can be completed through incremental fine-tuning at a single station, enabling the model to adaptively evolve as the power station expands. Through a causal discovery-digital twin dual-loop inference engine, an interpretable Top-K cascading failure path is output, upgrading traditional single-point alarms to predictive O&M driven by a dual-engine approach of "industry knowledge + local real-time data."

[0029] 3. By using edge-cloud collaborative federated learning and lightweight local model deployment, inference latency pressure is reduced, millisecond-level safety shutdown requirements are met, and a flexible and customizable implementation path is provided for energy storage power stations of different sizes and maturity levels while maintaining a minimum number of operable kernels.

[0030] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A schematic block diagram of the operation and maintenance system of a hybrid energy storage power station based on a large-scale energy storage operation and maintenance model is shown. Figure 2 A flowchart illustrating the operation and maintenance method of a hybrid energy storage power station based on a large-scale energy storage operation and maintenance model is shown. Figure 3 A schematic diagram of the operation and maintenance system of the hybrid energy storage power station according to an embodiment of this application is shown; Figure 4 This paper presents a schematic diagram of the overall system architecture of the hybrid energy storage power station operation and maintenance system according to an embodiment of this application. Figure 5 A schematic diagram illustrating the function of the model training module in an embodiment of this application is shown. Figure 6 This is a schematic diagram of the structure of a device according to an embodiment of this application. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] See appendix Figure 1 The system of this application includes: a data processing module, a knowledge graph module, a model training module, an adaptive deployment module, and an inference engine module; The data processing module includes a global data lake and an energy storage domain schema registry, which is used to acquire multi-source data, perform semantic alignment on the multi-source data, and generate triple data. The knowledge graph module is used to construct a triple-time hybrid knowledge graph (KG-T) based on the triple data. The model training module is used to construct a large-scale energy storage operation and maintenance model based on the multi-source data and the triple-time series hybrid knowledge graph; and to compress and quantize the large-scale energy storage operation and maintenance model to obtain a lightweight large-scale energy storage operation and maintenance model, and to extract the low-rank adaptation parameters (LoRA parameters) of the large-scale energy storage operation and maintenance model. The adaptive deployment module includes: a federated learning coordinator and an edge inference box, used for local deployment based on the lightweight energy storage operation and maintenance large model and low-rank adaptation parameters, combining local data for fault prediction, and outputting fault probability and feature vector; as well as updating the energy storage operation and maintenance large model and its low-rank adaptation parameters. The inference engine module includes: a causal discovery unit, a thermal-electric coupling model, and a graph neural ODE solver, which are used to generate a four-dimensional time series based on the multi-source data, and combine it with the locally deployed energy storage operation and maintenance large model to perform interpretable inference and fault prediction, and output an interpretable fault chain.

[0035] I. Data Processing Module.

[0036] The data processing module, including a global data lake and an energy storage schema registry, is used to acquire multi-source data, perform semantic alignment on the multi-source data, and generate triplet data.

[0037] The data processing module is also configured for; Acquire multi-source data and store it in a global data lake. The multi-source data includes: multi-source time-series data and multi-source unstructured data. According to the energy storage field model registration center, the multi-source data is semantically aligned to generate triple data with timestamps.

[0038] 1. Access data from multiple sources and store it in the global data lake.

[0039] Acquiring multi-source time-series data includes sampling time-series data such as cell voltage, current, temperature, PCS power, air conditioning supply and return air temperature, and fire-fighting combustible gas concentration (e.g., sampling frequency ≤ 1s), and entering the Kafka message queue through the Industrial Internet of Things protocol (MQTT / OPC-UA).

[0040] Multi-source unstructured data includes unstructured data such as PDF logs, WAV audio, JPG infrared images, and TXT work orders generated during operation and maintenance, which are stored and deployed through an object storage platform built on MinIO.

[0041] 2. Semantic alignment.

[0042] Based on the energy storage schema registry, the multi-source data is semantically aligned to generate triplet data with timestamps, including: Define a schema registry for the energy storage field as a unified rule; Multi-source data is converted into triple data that conforms to the Schema Registry by using a mapping script written according to the energy storage domain schema Registry, and then timestamped.

[0043] In this embodiment, a schema registry for the energy storage domain is defined, including: taking the top-level entities of the four major domains—battery, PCS, cooling, and fire protection—as the root, and defining the URIs, units, ranges, and sampling frequencies of measurement signals, fault modes, and operational actions downwards.

[0044] Based on the energy storage schema registry, mapping scripts are written for each data format to convert multi-source data into RDF triples that conform to the schema, and add a timestamp of UTC + time zone + device SN, which can achieve "plug and play".

[0045] In this embodiment, multi-source data of different formats can be unified into a triple form of "subject-verb-object", including: writing a mapping script for each data format and preset mapping rules; reading multi-source data and extracting "subject" (device), "verb" (attribute or relation), and "object" (value or target) according to the rules to form triple data; and attaching a timestamp containing UTC + time zone + device serial number to each triple.

[0046] II. Knowledge Graph Module.

[0047] The knowledge graph module is used to construct a triple-time hybrid knowledge graph (KG-T) based on the triple data.

[0048] The knowledge graph module is also configured for: Define node types, edge types, and weights; Nodes and edges are stored in a graph database to form a queryable network, thus constructing a knowledge graph. The temporal embedded knowledge graph uses a temporal relation graph convolutional network (Temporal-RGCN) to add a time decay function to each edge, forming a triplet-temporal hybrid knowledge graph.

[0049] In this embodiment, representation learning can also be performed based on triplet-temporal hybrid knowledge graph (KG-T), including: performing self-supervised comparative learning on subgraphs within a preset time period on the same device, so that the node feature vectors can simultaneously retain spatial topology and long-term temporal evolution information.

[0050] In this embodiment, a queryable network is formed by storing nodes and edges in a graph database. A temporal decay function is added to the edges using a temporal relation graph convolutional network to construct a triplet-temporal hybrid knowledge graph. Simultaneously, self-supervised comparative learning is carried out on the subgraphs within a preset time period of the same device to achieve representation learning. This allows the node feature vectors to retain both spatial topology and long-term temporal evolution information, forming a dynamic and evolvable temporal knowledge system that can express battery aging and fault hysteresis chains. This provides core knowledge support for subsequent large-scale model knowledge distillation, causal inference, and cascading failure prediction. The generated node feature vectors can be used for subsequent large-scale energy storage operation and maintenance model training, causal discovery and fault probability propagation in the inference engine, as well as similarity calculation, missing value completion, and fault tracing within the knowledge graph.

[0051] In this embodiment, four types of nodes are defined: ① Device instance (IMEI / SN); ② Operating condition range (SOC, rate of operation, and temperature threshold); ③ Fault mode (thermal runaway, IGBT failure, fan stall); ④ Maintenance action (balancing, replacement, dust removal). Among them, the device instance can be directly extracted from the data by the device ID; the operating condition range is divided according to data rules (e.g., SOC>90% is defined as "high charge state"); the fault mode can be extracted from historical alarms and maintenance records; and the maintenance action can be extracted from the work order text (e.g., "replace fan").

[0052] In this embodiment, three types of edges are defined: ① "Cause-Effect" edge weights are jointly learned by the prior probability of the fault chain and the SHAP value; ② "Coupling" edge weights are automatically generated by mutual information I>θ; ③ "Solution-Solution" edges are derived from the "Problem-Action" co-occurrence extracted from the work order NLP. Among them, the weights of "Cause-Effect" edges are obtained by historical data statistics (fault A often leads to fault B) and algorithm analysis; "Coupling" edges are constructed by calculating the correlation between two device parameters, and if the correlation exceeds a threshold; "Solution" edges can be extracted from the "Problem-Handling Measures" work order pairs.

[0053] In this embodiment, the time decay function formula is: Edge weight(t) = Initial weight exp(-λ Δt), Where Δt is the time length during which the distance relationship occurs; λ is the attenuation coefficient, which can be set based on experience (e.g., the battery aging process is slow, so the value of λ is small), meaning that the impact of a fault that occurred a long time ago on the present will be smaller.

[0054] III. Model Training Module.

[0055] The model training module is used to construct a large-scale energy storage operation and maintenance model based on the multi-source data and the triple-time series hybrid knowledge graph; and to compress and quantize the large-scale energy storage operation and maintenance model to obtain a lightweight large-scale energy storage operation and maintenance model, and to extract the low-rank adaptation parameters (LoRA parameters) of the large-scale energy storage operation and maintenance model.

[0056] The model training module further includes a multimodal encoder, comprising a temporal encoder, a text / speech encoder, and an image encoder, used to uniformly vectorize the multi-source data.

[0057] The model training module is also configured to: For the four modalities of time series, text, speech, and image data in the multi-source data, corresponding pre-trained encoders are set up to transform the multi-source data into feature vectors in a unified manner; A predetermined amount of corpus is generated based on the triplet-temporal hybrid knowledge graph, and a domain model is obtained by distillation of the feature vectors through a general large model. A large-scale energy storage operation and maintenance model is trained by weighting triplet loss and masked language model loss according to a preset ratio. The large-scale energy storage operation and maintenance model is compressed and quantized to obtain a lightweight large-scale energy storage operation and maintenance model, and the low-rank adaptation parameters (LoRA parameters) of the large-scale energy storage operation and maintenance model are extracted.

[0058] In this embodiment, INT8 quantization can be used to compress and quantize the large energy storage operation and maintenance model to obtain a lightweight large energy storage operation and maintenance model.

[0059] In this embodiment, a multimodal encoder is used to uniformly vectorize four types of heterogeneous data—time series, text, speech, and image—from multi-source data. This vectorized feature vector is then used to eliminate modal differences and form a unified input, providing fundamental feature support for subsequent training of the large-scale energy storage operation and maintenance model and prediction by the inference engine. The multimodal encoder is responsible for transforming multi-source heterogeneous data into features of a unified dimension, thereby realizing the feature extraction and knowledge distillation process from raw data to a large-scale domain model.

[0060] In this embodiment, the triplet loss is obtained by calculating the representation learning error of nodes and edges in the knowledge graph; the mask language model loss is obtained by performing mask prediction on the operation and maintenance text corpus and calculating the prediction error; the triplet loss and the mask language model loss are weighted and jointly trained according to a preset ratio, which enables the large energy storage operation and maintenance model to take into account both the knowledge graph structure constraints and language modeling capabilities, thereby improving the model's accuracy and interpretability.

[0061] IV. Adaptive Deployment Module.

[0062] The adaptive deployment module includes: an edge inference box and a federated learning coordinator; The federated learning coordinator is used for central-side parameter aggregation and differential privacy policy management; The edge inference box is used to perform local deployment based on the lightweight energy storage operation and maintenance model and LoRA parameters to obtain a local lightweight energy storage operation and maintenance model; based on local data, it performs fault prediction through the local lightweight energy storage operation and maintenance model and outputs fault probability and feature vector.

[0063] The federated learning coordinator is used for central-side parameter aggregation and differential privacy policy management, including: Deploy lightweight energy storage operation and maintenance large model and LoRA parameters to the edge inference box; Based on the LoRA parameter increments uploaded by each edge inference box, the LoRA parameter increments of multiple sites are aggregated to generate a new large-scale energy storage operation and maintenance model and new LoRA parameters. The new energy storage operation and maintenance model is compressed and quantized to obtain an updated lightweight energy storage operation and maintenance model. The new LoRA parameters are then distributed to the edge inference box to realize the dynamic iterative update of the energy storage operation and maintenance model.

[0064] Edge reasoning box, used for: Based on the lightweight energy storage operation and maintenance model and LoRA parameters, complete the deployment of the local lightweight energy storage operation and maintenance model. Using local data, LoRA parameters are incrementally fine-tuned using AdaLoRA, enabling the deployment and updating of a local, lightweight energy storage operation and maintenance model. Based on a local, lightweight energy storage operation and maintenance model, the model outputs fault probability and feature vectors, and works in conjunction with the inference engine module to perform fault prediction.

[0065] In this embodiment, the edge inference box is also used for: online real-time discrimination of the device's operating status, fault risk, and model drift; providing calculation support for INT8 quantization parameters, electrothermal parameters, inference thresholds, etc., and providing real-time feature, probability, and causal confidence calculations for fault prediction and status discrimination, serving the entire process of local inference and incremental training; after completing local incremental fine-tuning, only uploading the weight increment of LoRA parameters; receiving the latest global quantization model from the cloud, combining it with the local LoRA parameters, and completing the non-stop online update of the local lightweight energy storage operation and maintenance model.

[0066] In this embodiment, the federated learning coordinator can adopt the Flower framework, which distributes lightweight modeling through the central server, and each power station uploads gradients locally using differential privacy. The communication volume for 3 rounds of convergence is less than 100 MB.

[0067] In this embodiment, after the model training module completes knowledge distillation, model training, compression quantization, and LoRA matrix extraction, it obtains a lightweight energy storage operation and maintenance (O&M) model and LoRA parameters. The federated learning coordinator distributes the lightweight O&M model and LoRA parameters to the edge inference box. The edge inference box loads the lightweight O&M model and LoRA parameters, performs AdaLoRA incremental fine-tuning based on local data, completes the deployment of the local lightweight O&M model, performs inference prediction based on local data using the local lightweight O&M model, and returns the local LoRA parameters. The federated learning coordinator aggregates the increments of local LoRA parameters from all power plants to generate a new lightweight O&M model, which is then distributed again. The edge inference box uses the newly distributed lightweight O&M model and its own LoRA parameters to update the local lightweight O&M model, achieving dynamic adjustment of the local lightweight O&M model without changing the quantization structure.

[0068] V. Inference Engine Module.

[0069] The inference engine module includes a causal discovery unit, a thermal-electric coupling model, and a graph neural ODE solver. It is used to generate a four-dimensional time series based on the multi-source data, and combine it with the local lightweight energy storage operation and maintenance model to perform interpretable inference and fault prediction, and output the prediction results.

[0070] The causal discovery unit is used to generate a four-dimensional time series based on the multi-source data; and to mine the causal relationships between variables from the four-dimensional time series using the PC algorithm + DoWhy, and to establish a directed acyclic graph (DAG) and update the nodes at a preset period. A thermo-electric coupling model is used to perform thermo-electric coupling simulations by using the parent node of a directed acyclic graph as the boundary condition and combining it with the feature vectors output by the local lightweight energy storage operation and maintenance model, and outputting simulation data; simulation parameters are calibrated based on the simulation data and real-time sensor data from the field. The graph neural network ODE solver is used to propagate the fault probability / feature vector output by the directed acyclic graph, simulation data, and the local lightweight energy storage operation and maintenance large model. It propagates the fault probability on the subgraph of the triple-time series hybrid knowledge graph through the graph neural network ODE, integrates causal constraints, physical simulation, and large model probability, performs interpretable reasoning and fault prediction, and outputs prediction results. The prediction results include: interpretable fault chains, cascading fault prediction results, and full-link fault probability distribution.

[0071] In this embodiment, the inference engine module uses the fault probability and feature vector output by the local lightweight energy storage operation and maintenance large model to perform interpretable inference and fault prediction. At the same time, the directed acyclic graph of the inference engine module and the simulation parameters of the thermal-electric coupling model will also back-calibrate the energy storage operation and maintenance large model, and the two form a closed-loop collaboration. The input data of the causal discovery unit is the temperature-voltage-PCS power-cooling power four-dimensional time series obtained after processing the multi-source data.

[0072] In this embodiment, the inference engine module uses a local lightweight energy storage operation and maintenance large model as the data-driven foundation, and integrates causal reasoning, physical simulation, and graph neural network ODE technology to achieve interpretable, traceable, and physically self-consistent fault prediction and reasoning. It is the interpretability enhancement engine and physical calibrator of the energy storage operation and maintenance large model, and the two form a two-way closed-loop collaboration.

[0073] In this embodiment, the causal discovery unit uses the PC algorithm + DoWhy to build a directed acyclic graph (DAG) based on the four-dimensional time series of "temperature-voltage-PCS power-cooling power" within a preset time window in real time. The nodes are updated according to a preset period. The thermal-electric coupling model uses the parent node of the DAG as the boundary condition input and combines the feature vector output by the local lightweight energy storage operation and maintenance model to use UKF to correct the electrical-thermal parameters (internal resistance, convection coefficient) online. Chain prediction is performed through the graph neural ODE solver. The graph neural ODE propagates the failure probability on the triplet-time series hybrid knowledge graph (KG-T) subgraph. The time step is set, and the Top-K cascading failure paths and probabilities for the future preset time are output.

[0074] Furthermore, to achieve better results, the system in this embodiment may also include the following enhanced structures: The data processing module may also include: a data quality management unit and an infrared image super-resolution preprocessing unit; The knowledge graph module may also include: a knowledge verification unit; The adaptive deployment module may also include: an incremental expert head hot-swap base and a watchdog rollback unit.

[0075] The data quality governance unit is used to clean outliers and complete missing values ​​in multi-source data. It employs a sliding window + 3σ + isolated forest joint cleaning method, and uses collaborative filtering based on triple-time series hybrid knowledge graph (KG-T) similar devices to complete missing values. Furthermore, it uses local differential privacy noise addition before uploading sensitive process parameters to meet vendor confidentiality requirements.

[0076] Outlier cleaning includes: Using a sliding window with +3σ, the mean and standard deviation of data within a preset time period are calculated; data exceeding the range of "mean ± 3 times standard deviation" are considered outliers and deleted. The Isolation Forest algorithm identifies "isolated points" that are significantly different from other data patterns. These are considered anomalous data and are deleted.

[0077] Missing value completion includes: Based on the triplet-temporal hybrid knowledge graph (KG-T), the "neighboring device" most similar to the device under test is found, and the normal data of the neighboring device at the same time is used to calculate and fill in the missing values.

[0078] Differential privacy noise addition involves adding random noise to the data to protect the original sensitive information from leakage. Specifically, it includes: determining the privacy protection strength parameter ε (the smaller ε is, the greater the added noise and the stronger the privacy, but the lower the data accuracy); adding random noise following a Laplace distribution to the original data that needs to be protected; and uploading the noise-added data.

[0079] The knowledge verification unit is used to verify the knowledge consistency of the triple-time sequence hybrid knowledge graph through the "causal + consistency" dual gate.

[0080] In this embodiment, in order to ensure that newly discovered knowledge does not conflict with existing knowledge and to update the credibility of knowledge based on new evidence, knowledge verification can be performed on the triple-time series hybrid knowledge graph, including: causal consistency verification and Bayesian dynamic update.

[0081] Causal consistency verification: When a new potential causal edge (such as "A causes B") is discovered through the PC algorithm, the conflict degree between the new causal edge and the existing relationship in the triple-temporal hybrid knowledge graph is calculated. If the conflict degree is lower than a preset threshold (such as 0.2), the new causal edge is merged into the triple-temporal hybrid knowledge graph; otherwise, manual review is triggered to ensure the logical consistency of the knowledge system and prevent the introduction of contradictory information. Bayesian dynamic update: The weight of each edge in the triplet-time series hybrid knowledge graph is regarded as a quantitative representation of its credibility. When new operation and maintenance feedback is obtained (such as the closed-loop result of maintenance work order), the posterior weight of the edge is updated using the Bayesian formula. If the weight change exceeds the set threshold (such as 30%), it indicates that the credibility of the knowledge has changed significantly. At this time, manual review is triggered to ensure that the dynamic evolution of knowledge is traceable and auditable.

[0082] In this embodiment, the "causal + consistency" dual-gate verification mechanism can remove causal conflict edges and dynamically and reliably update the edge weights of the knowledge graph, further reducing the false alarm rate.

[0083] Incremental expert head hot-swappable base: The incremental expert head hot-swappable base adopts a hybrid expert (MoE) architecture and a dynamic routing mechanism. When new battery systems such as sodium-ion batteries and solid-state batteries are added, only an independent expert head and corresponding routing branch are added to the model. There is no need to modify the original model backbone parameters. The routing mechanism directs the data of the new battery system to the new expert head for training and inference, thereby enabling the rapid access of new systems without destroying the original model knowledge, avoiding full retraining, and significantly reducing computing power consumption and training time.

[0084] In this embodiment, when a new battery system (sodium-ion, solid-state battery) is connected, the incremental expert head hot-swap base eliminates the need to retrain the entire model, effectively saving GPU computing power.

[0085] The watchdog rollback unit is used to automatically roll back to the previous version in the event of model drift, avoiding erroneous chain predictions that could lead to unintended shutdowns.

[0086] In this embodiment, if the probability of a real fault falling into the Top-K path for a consecutive preset number of days is less than a preset threshold, model drift is determined, and the system automatically rolls back to the previous version and issues an alarm to ensure reliability. When a new fault mode (such as sodium ion precipitation at low temperature) appears, the Schema Registry automatically expands its nodes, triggering incremental expert head training. Retraining time is short and no downtime is required. The Top-K path probability is calculated by the graph neural ODE solver in the inference engine module. The graph neural ODE solver propagates fault probability based on a subgraph of a triple-temporal hybrid knowledge graph, outputting the K most likely cascaded failure paths in the future time period and their corresponding probabilities in descending order of confidence. By comparing this probability with the actual occurrence of the real fault, the probability of a real fault falling into the Top-K path can be obtained.

[0087] The infrared image super-resolution preprocessing unit uses the Real-ESRGAN model to perform super-resolution reconstruction of infrared images, improving the quality of infrared images and enabling the system to detect early signs of battery thermal runaway earlier and more accurately.

[0088] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0089] Figure 3 This is a schematic diagram of the operation and maintenance system of a hybrid energy storage power station according to an embodiment of this application. It shows the relationship between "data lake, KG-T, distillation, federated evolution, and inference", highlighting that after the multi-source data in the whole domain is semantically aligned once by the Schema Registry, it directly drives all subsequent links. Any new manufacturer / new model of equipment only needs to complete the registration in the data lake to automatically flow into the subsequent knowledge distillation and inference closed loop, realizing "plug and play" expansion.

[0090] Figure 4 This is a schematic diagram of the overall system architecture of the hybrid energy storage power station operation and maintenance system according to an embodiment of this application. In this embodiment, a two-level deployment of "centralized training in the cloud and edge inference in the power station" is adopted: the cloud is responsible for heavy computing knowledge distillation, causal discovery and parameter aggregation, while the edge inference box (Atlas 300I) only stores LoRA parameters and completes millisecond-level local inference and incremental updates; through the Flower federation framework and differential privacy strategy, it is ensured that the original data of the manufacturer does not leave the domain, meeting compliance requirements while realizing the continuous evolution of cross-site models.

[0091] Figure 5 This is a functional diagram of the model training module in an embodiment of this application. In this embodiment, the temporal encoder adopts Transformer-MoE (hybrid expert model) with a total of 8 expert heads, grouped and gated according to "manufacturer + battery type", and the parameter quantity is set to 0.8B; the text / speech encoder adopts BERT-wwm-ext 110M + Conformer streaming encoding, which supports direct conversion of 16kHz speech into speech feature vectors; the image encoder can use EfficientNet-V2-S as an infrared hotspot encoder, outputting 1280-dimensional vectors.

[0092] Temporal-RGCN transforms time-decayed causal edges into "fact-question-answer" corpora. It uses a 7B general large model as the teacher, distilling triple knowledge and multimodal features together into a 1.3B domain-enhanced energy storage operation and maintenance large model. The student model, through a pluggable LoRA structure, completes incremental fine-tuning at the edge in a short time. It can quickly absorb new failure modes without returning the original data, realizing a lightweight, interpretable, and sustainably evolving energy storage operation and maintenance brain.

[0093] The following are specific implementation examples of this application.

[0094] Implementation Case 1 Background: A large-scale independent energy storage power station on the grid side employs a hybrid deployment of batteries from multiple manufacturers and of various types. Equipment such as PCS, air conditioning, and fire protection systems come from different suppliers. The existing operation and maintenance system only supports single-point alarms, and has experienced cascading failures involving multiple equipment levels. This resulted in long troubleshooting cycles, significant power loss, and an inability to meet the requirements for safe operation and maintenance and rapid response.

[0095] Implement according to the technical solution of this application: ① Plug and play data layer: Edge master nodes are deployed in the station control room, equipped with a unified semantic registration module for the energy storage field. After maintenance personnel upload the equipment point tables of various manufacturers, the system automatically generates protocol mapping scripts to quickly complete the semantic conversion of heterogeneous communication protocols from multiple manufacturers, unifying the measurements of various equipment into standard domain identifiers; it collects time-series data of all dimensions such as cells, PCS, air conditioning, and fire protection in a high-frequency sampling mode, and connects to the multi-source heterogeneous data lake of the entire domain through a message queue.

[0096] ② KG-T Time Series Knowledge Graph Fast Cold Start: Based on unstructured data such as historical operation and maintenance work orders and maintenance documents of power plants, fault-related triple knowledge is extracted through a text encoding model; combined with a time-series graph neural network, mutual information of equipment parameters is calculated within a sliding time window to generate causal relationship edges of equipment faults with time decay weights; for new battery systems, dedicated fault nodes are added, the semantic registration module automatically expands the identifiers, the knowledge graph is upgraded and the on-site business is not interrupted.

[0097] ③ Federal knowledge distillation training: Using a large cloud-based model as the teacher model, each power station completes calculations locally and uploads the noise gradient. The cloud aggregates the data to obtain a large energy storage operation and maintenance model. At the edge, a low-rank adaptation method is used to inject local operating data to complete incremental fine-tuning and generate a lightweight energy storage operation and maintenance model adapted to the local station, enabling rapid local deployment.

[0098] ④ Causal-digital twin dual-cycle operation: When an abnormal increase in cabin temperature occurs on-site, the causal discovery algorithm constructs a directed acyclic graph of equipment status in real time to lock the causal chain of the fault; the thermoelectric coupling digital twin model uses the causal graph as boundary conditions to correct key parameters of the model online and predict the equipment status and failure risk after the fault continues to develop; the failure probability is propagated on the knowledge graph subgraph through graph god ordinary differential equations to output high-probability cascaded failure paths in the future period.

[0099] ⑤ Operation and maintenance intervention and results: After receiving a cascading failure warning, the operation and maintenance platform performs remote power reduction control and dispatches a team to handle the situation on-site, quickly restoring the equipment to normal status, successfully blocking the chain of failures, significantly shortening the handling time, and intervening much earlier than the traditional alarm mode, thus avoiding power loss and equipment damage.

[0100] ⑥ Online model evolution: The fault samples and handling data were sent back to the cloud, updating the knowledge graph association weights and triggering cloud-based federated learning iterations. The performance of the energy storage operation and maintenance big model continued to improve, and the model update was distributed to each edge node, allowing the on-site inference equipment to complete the upgrade without restarting.

[0101] Implementation effect The integration cycle of equipment from multiple vendors has been significantly shortened; the hit rate of cascaded failure prediction has been significantly improved, and explainable fault paths can be output in advance; the annual operation and maintenance cost of power plants has been significantly reduced, the frequency of on-site manual visits has been reduced, and the safety and efficiency of operation and maintenance have been comprehensively improved.

[0102] Implementation Case 2 Background: A certain photovoltaic-storage-charging integrated system uses recycled batteries for energy storage. It lacks complete factory records, has a small operation and maintenance team, and lacks professional battery technical support. As a result, it has problems such as unknown battery aging characteristics, high risk of thermal runaway under high temperature environment, and difficulty in predicting failure in advance.

[0103] Implemented according to the technical solution of this invention: ① Cold start with zero historical data: The edge inference box is deployed on-site, the zero-sample initialization mode is started, and the pre-trained model of the same type of battery is loaded to complete the basic deployment; after the access device communication point table is connected, the unified semantic registration module automatically identifies the key operating parameters of the cascade battery and generates exclusive identifiers; the system collects on-site time-series operating data, and the knowledge graph automatically extracts the potential causal relationship edges between battery status and faults.

[0104] ② Automated accumulation of voice-based operation and maintenance knowledge: Maintenance personnel upload inspection and repair information via voice. After voice encoding and text decoding, the information is automatically converted into faults in a knowledge graph. The solution is linked to form a closed loop of operational and maintenance knowledge accumulation.

[0105] ③ Summer high-temperature fault early warning and handling: When an air conditioning system malfunctions and causes the cabin temperature to rise continuously, the edge device quickly completes graph neural network simulation calculations, outputs high-probability cascaded failure paths and targeted intervention suggestions. Maintenance personnel then follow the system's suggestions to perform operations such as power reduction, activation of backup equipment, and on-site cleanup, rapidly stabilizing the cabin temperature, eliminating the risk of thermal runaway, and ensuring the normal operation of charging piles and park loads.

[0106] ④ Model adaptive evolution: The edge inference box performs low-rank incremental training based on the fault and handling data of the day, adds new equipment coupling and correlation edges, upgrades the model version, and ensures zero interruption of on-site business; the same batch of tiered battery projects can directly reuse the model of this site to achieve rapid deployment and plug-and-play.

[0107] Implementation Results: In scenarios where there are no factory records for cascaded batteries, the accuracy of battery status estimation meets the requirements of park scheduling; high-temperature environment cascading faults can be warned in advance, avoiding losses from fire malfunctions and power outages; the frequency of on-site visits by battery experts is greatly reduced, annual operation and maintenance costs are lowered, and the overall project benefits and operational reliability are improved.

[0108] See appendix Figure 2This application also provides an operation and maintenance method for a hybrid energy storage power station based on a large-scale energy storage operation and maintenance model, including: Acquire multi-source data, perform semantic alignment on the multi-source data, and generate triplet data; Construct a triplet-temporal hybrid knowledge graph based on the triplet data; A large-scale energy storage operation and maintenance model is constructed based on the multi-source data and the triple-time series hybrid knowledge graph; the large-scale energy storage operation and maintenance model is compressed and quantized to obtain a lightweight large-scale energy storage operation and maintenance model, and the low-rank adaptation parameters of the large-scale energy storage operation and maintenance model are extracted. Based on the lightweight energy storage operation and maintenance model and low-rank adaptation parameters, local deployment is carried out, and fault prediction is performed by combining local data to output fault probability and feature vector. A four-dimensional time series is generated based on the multi-source data, and interpretable reasoning and fault prediction are performed in conjunction with the locally deployed energy storage operation and maintenance model, and the prediction results are output.

[0109] Furthermore, this application also provides an operation and maintenance device for a hybrid energy storage power station based on a large-scale energy storage operation and maintenance model, including: The data processing module includes a global data lake and an energy storage field model registration center, which acquires multi-source data, performs semantic alignment on the multi-source data, and generates triplet data. The knowledge graph module constructs a triple-time series hybrid knowledge graph based on the triple data. The model training module constructs a large-scale energy storage operation and maintenance model based on the multi-source data and the triple-time-series hybrid knowledge graph; and compresses and quantizes the large-scale energy storage operation and maintenance model to obtain a lightweight large-scale energy storage operation and maintenance model, and extracts the low-rank adaptation parameters of the large-scale energy storage operation and maintenance model. The adaptive deployment module includes: a federated learning coordinator and an edge inference box, which perform local deployment based on the lightweight energy storage operation and maintenance model and low-rank adaptation parameters, combine local data to perform fault prediction, and output fault probability and feature vector; as well as update the energy storage operation and maintenance model and its low-rank adaptation parameters. The inference engine module includes: a causal discovery unit, a thermal-electric coupling model, and a graph neural ODE solver. It generates a four-dimensional time series based on the multi-source data and combines it with the locally deployed energy storage operation and maintenance large model to perform interpretable inference and fault prediction, and outputs the prediction results.

[0110] Based on the same concept, this application also provides a computer-readable storage medium storing one or more programs, which, when executed, can realize the aforementioned operation and maintenance method for hybrid energy storage power stations based on a large-scale energy storage operation and maintenance model.

[0111] like Figure 6As shown in the illustration, this application also provides a device including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus.

[0112] The memory is a computer-readable storage medium used to store one or more programs.

[0113] The processor is configured to execute a program stored in a computer-readable storage medium.

[0114] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus.

[0115] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A hybrid energy storage power station operation and maintenance system based on a large-scale energy storage operation and maintenance model, characterized in that, include: The module includes a data processing module, a knowledge graph module, a model training module, an adaptive deployment module, and an inference engine module. The data processing module includes a global data lake and an energy storage field model registration center, which is used to acquire multi-source data, perform semantic alignment on the multi-source data, and generate triple data. The knowledge graph module is used to construct a triple-time series hybrid knowledge graph based on the triple data; The model training module is used to construct a large-scale energy storage operation and maintenance model based on the multi-source data and the triple-time series hybrid knowledge graph. The energy storage operation and maintenance model is compressed and quantized to obtain a lightweight energy storage operation and maintenance model, and the low-rank adaptation parameters of the energy storage operation and maintenance model are extracted. The adaptive deployment module includes: a federated learning coordinator and an edge inference box, used for local deployment based on the lightweight energy storage operation and maintenance large model and low-rank adaptation parameters, combining local data for fault prediction, and outputting fault probability and feature vector; as well as updating the energy storage operation and maintenance large model and its low-rank adaptation parameters. The inference engine module includes: a causal discovery unit, a thermal-electric coupling model, and a graph neural ODE solver, used to generate a four-dimensional time series based on the multi-source data, and combine it with the locally deployed energy storage operation and maintenance large model to perform interpretable inference and fault prediction, and output the prediction results.

2. The system according to claim 1, characterized in that, The data processing module is also configured to: Acquire multi-source data and store it in a global data lake. The multi-source data includes: multi-source time-series data and multi-source unstructured data. According to the energy storage field model registration center, the multi-source data is semantically aligned to generate triple data with timestamps.

3. The system according to claim 1, characterized in that, The knowledge graph module is also configured to: Define node types, edge types, and weights; Nodes and edges are stored in a graph database to form a queryable network and construct a knowledge graph. Temporal embedded knowledge graphs employ temporal relation graph convolutional networks to add a time decay function to each edge, forming a triplet-temporal hybrid knowledge graph.

4. The system according to claim 1, characterized in that, The model training module also includes: A multimodal encoder, including a timing encoder, a text / speech encoder, and an image encoder, is used to uniformly vectorize the multi-source data.

5. The system according to claim 4, characterized in that, The model training module is also configured to: For the four modalities of time series, text, speech, and image data in the multi-source data, corresponding pre-trained encoders are set up to transform the multi-source data into feature vectors in a unified manner; A predetermined amount of corpus is generated based on the triplet-temporal hybrid knowledge graph, and a domain model is obtained by distillation of the feature vectors through a general large model. A large-scale energy storage operation and maintenance model is trained by weighting triplet loss and masked language model loss according to a preset ratio. The large-scale energy storage operation and maintenance model is compressed and quantized to obtain a lightweight large-scale energy storage operation and maintenance model, and the low-rank adaptation parameters of the large-scale energy storage operation and maintenance model are extracted.

6. The system according to claim 1, characterized in that, The adaptive deployment module includes: an edge inference box and a federated learning coordinator. The federated learning coordinator is used for central-side parameter aggregation and differential privacy policy management. The edge inference box is used to perform local deployment based on the lightweight energy storage operation and maintenance model and low-rank adaptation parameters to obtain a local lightweight energy storage operation and maintenance model; based on local data, it performs fault prediction through the local lightweight energy storage operation and maintenance model and outputs fault probability and feature vector.

7. The system according to claim 6, characterized in that, The federated learning coordinator is used for central-side parameter aggregation and differential privacy policy management, including: Lightweight energy storage operation and maintenance large model and low-rank adaptation parameters are issued to the edge inference box; Based on the incremental low-rank adaptation parameters uploaded by each edge inference box, the incremental low-rank adaptation parameters of multiple sites are aggregated to generate a new large-scale energy storage operation and maintenance model and new low-rank adaptation parameters. The new energy storage operation and maintenance model is compressed and quantized to obtain an updated lightweight energy storage operation and maintenance model. The new low-rank adaptation parameters are then sent back to the edge inference box to realize the dynamic iterative update of the energy storage operation and maintenance model.

8. The system according to claim 6, characterized in that, The edge inference box is also configured to: Based on the lightweight energy storage operation and maintenance model and low-rank adaptation parameters, complete the deployment of the local lightweight energy storage operation and maintenance model. Using local data, the low-rank adaptation parameters are incrementally fine-tuned using AdaLoRA to complete the deployment and update of a local, lightweight energy storage operation and maintenance model. Based on a local, lightweight energy storage operation and maintenance model, the model outputs fault probability and feature vectors, and works in conjunction with the inference engine module to perform fault prediction.

9. The system according to claim 1, characterized in that, The inference engine module includes: a causal discovery unit, a thermal-electric coupling model, and a graph neural ODE solver. The causal discovery unit is used to generate a four-dimensional time series sequence based on the multi-source data; and to establish a directed acyclic graph based on the four-dimensional time series sequence, and update the nodes at a preset period. A thermo-electric coupling model is used to perform thermo-electric coupling simulations using the directed acyclic graph as boundary conditions and the feature vectors output by the local lightweight energy storage operation and maintenance model, and output simulation data; simulation parameters are calibrated based on the simulation data and real-time sensor data from the field. The graph neural network ODE solver is used to propagate the fault probability / feature vector output by the directed acyclic graph, simulation data, and the local lightweight energy storage operation and maintenance large model. It propagates the fault probability on the subgraph of the triple-time series hybrid knowledge graph through the graph neural network ODE, integrates causal constraints, physical simulation, and large model probability, performs interpretable reasoning and fault prediction, and outputs the prediction results.

10. The system according to any one of claims 1-9, characterized in that, The system also includes: a data quality governance unit, an infrared image super-resolution preprocessing unit, a knowledge verification unit, an incremental expert head hot-swappable base, and a watchdog rollback unit.

11. A hybrid energy storage power station operation and maintenance method based on a large-scale energy storage operation and maintenance model, characterized in that, The method includes: Acquire multi-source data, perform semantic alignment on the multi-source data, and generate triplet data; Construct a triplet-temporal hybrid knowledge graph based on the triplet data; A large-scale energy storage operation and maintenance model is constructed based on the multi-source data and the triple-time series hybrid knowledge graph; the large-scale energy storage operation and maintenance model is compressed and quantized to obtain a lightweight large-scale energy storage operation and maintenance model, and the low-rank adaptation parameters of the large-scale energy storage operation and maintenance model are extracted. Based on the lightweight energy storage operation and maintenance model and low-rank adaptation parameters, local deployment is carried out, and fault prediction is performed by combining local data to output fault probability and feature vector. A four-dimensional time series is generated based on the multi-source data, and interpretable reasoning and fault prediction are performed in conjunction with the locally deployed energy storage operation and maintenance model, and the prediction results are output.

12. A computer-readable storage medium storing one or more programs, characterized in that, When one or more of these programs are executed, the hybrid energy storage power station operation and maintenance method based on the large-scale energy storage operation and maintenance model described in claim 11 can be implemented.

13. An electronic device comprising a processor, a communication interface, a computer-readable storage medium as described in claim 12, and a communication bus; wherein, The processor, communication interface, and computer-readable storage medium communicate electronically with each other via a communication bus; characterized in that, The processor is used to execute programs stored in a computer-readable storage medium.

Images