An application architecture based on super-computing data platform

Abstract

The application discloses an application framework based on super computing power data platform, comprising: an edge super computing service module for data collection, processing, synchronization, convergence and transmission, including a collection service unit and a data processing unit; a super computing power data platform module for data storage, calculation, management, scheduling and algorithm integration, including a super computing power real-time processing engine unit; an application layer module for data display, analysis, alarm and operation and maintenance management; the collection service unit is used for adjusting the data sampling frequency according to the equipment abnormal signal; the data processing unit is used for compressing data by adopting a method of adaptive differential compression combined with secondary compression; the super computing power real-time processing engine unit is used for task fragmentation according to the equipment level, the data type and the time window, and distributing the task fragmentation to different thread groups for multi-thread parallel calculation based on the equipment correlation degree. The application can improve the data processing efficiency and reduce the storage cost.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to an application architecture based on a supercomputing data platform. Background Technology

[0002] With the rapid development of information technology, the demand for data processing is increasing across various industries. For the energy storage industry, the increasing demand for real-time acquisition and analysis of massive amounts of time-series data from energy storage systems and smart grids presents a significant challenge in real-time processing and risk control of such data. Traditional data processing architectures often struggle to meet the requirements of real-time performance, efficiency, and scalability when dealing with massive amounts of data. They suffer from high data processing latency and high storage costs, failing to effectively achieve efficient data acquisition, storage, computation, and retrieval, and thus failing to meet the needs of modern enterprises for rapid data analysis and decision-making.

[0003] Therefore, there is an urgent need for a supercomputing data platform application architecture that can improve data processing efficiency and reduce storage costs in the energy storage industry, in order to solve the problems of low data processing efficiency and high storage costs in existing technologies. Summary of the Invention

[0004] To address the above technical issues, this invention provides an application architecture based on a supercomputing data platform, which can improve data processing efficiency, reduce storage costs, and enable efficient processing and intelligent application of massive amounts of data in the energy storage industry.

[0005] This invention provides an application architecture based on a supercomputing power data platform, including: An edge supercomputing service module is used for data acquisition, processing, synchronization, aggregation, and secure transmission. The edge supercomputing service module includes at least an acquisition service unit and a data processing unit. The acquisition service unit is used to adjust the data sampling frequency based on detected abnormal device signals. The data processing unit is used to compress the acquired data using an adaptive differential compression combined with secondary compression. The supercomputing data platform module is used for data storage, computation, management, scheduling, algorithm integration, and service provision. The supercomputing data platform module includes at least a supercomputing real-time processing engine unit. The supercomputing real-time processing engine unit is used to perform three-dimensional task sharding based on device level, data type, and time window, and to allocate task shards to different thread groups based on device correlation for multi-threaded parallel computation, wherein each thread group processes an independent shard. The application layer module is used for data display, analysis, alarm handling, and operation and maintenance management.

[0006] As an improvement to the above solution, adjusting the data sampling frequency based on the detected abnormal device signal includes: Acquire device operation data and determine the current state of the device based on the device operation data; if the current state of the device is normal, set the data sampling frequency to a first frequency; if the current state of the device is abnormal, set the data sampling frequency to a second frequency; the second frequency is higher than the first frequency. Based on the device operation data, a preset neural network model is used to predict the changing trend of the device status to obtain the predicted device status; if the predicted device status is abnormal, the data sampling frequency is set to the second frequency.

[0007] As an improvement to the above scheme, the method of compressing the collected data using adaptive differential compression combined with secondary compression includes: Calculate the data volatility index based on the collected data; A differential strategy is determined based on the data volatility index to compress the data and obtain a first compression result; the differential strategy includes real-time sampling differential and window effective value differential; Based on the first compression result and the dynamic dictionary, a secondary compression is performed by selecting an entropy coding method according to the data characteristics. The discrete state data uses Huffman coding, and the continuous difference data uses arithmetic coding. The dynamic dictionary uses the Least Recently Used (LRU) page replacement algorithm to maintain the high-frequency difference pattern.

[0008] As an improvement to the above scheme, the step of determining the difference strategy for data compression based on the data volatility index includes: Select a differencing strategy based on the data volatility index and determine the differencing benchmark value; The dynamic quantization step size is determined based on the data volatility index. In lossy compression mode, a lightweight LSTM model is used to predict the data at the next time point based on the differential reference value, and the prediction residual is calculated as the differential object. The data is then encoded in combination with the dynamic quantization step size to compress the data.

[0009] As an improvement to the above solution, the step of performing three-dimensional task sharding based on device level, data type, and time window, and allocating task shards to different thread groups based on device correlation, includes: Based on the physical structure of the energy storage system, the task is divided into several equipment levels to obtain the first task segment; the equipment level includes power station, container, and battery cell; Based on the first task fragment, the task is divided into three data types: status data, running data, and fault data to obtain the second task fragment. Based on the second task slice, the task is divided into three time window types: real-time window, trend window, and statistics window, to obtain the third task slice; Based on the third task segmentation, the energy storage device correlation factor is calculated, and the task segments corresponding to the energy storage device correlation factors that are greater than the preset correlation threshold are allocated to the same node for operation, so as to achieve dynamic load balancing; wherein, the energy storage device correlation factor is calculated by the ratio of data interaction frequency to the cable length between devices.

[0010] As an improvement to the above solution, the supercomputing data platform module also includes an offline computing engine unit; The offline computing engine unit adopts a distributed computing framework, in which energy storage vectorization operators are set for high-frequency time-series data of energy storage. The energy storage vectorization operators include multi-level time window aggregation operators, parameter correlation coefficient calculation operators, and anomaly detection operators. The multi-level time window aggregation operator is used for multi-level aggregation based on the energy storage device level and the time window. The parameter correlation coefficient calculation operator is used to calculate the Pearson correlation coefficient of two energy storage parameters and is optimized by single instruction multiple data (SIMD) instructions. The anomaly detection operator is used to integrate an energy storage anomaly detection algorithm.

[0011] As an improvement to the above scheme, the offline computing engine unit is equipped with an energy storage protocol parser. The energy storage protocol parser maintains the energy storage protocol feature library through the Hive metadata database and adopts a state machine and regular expression hybrid parsing strategy to support dynamic access of multiple heterogeneous data sources. The offline computing engine unit uses Hive storage, which employs four-level composite partitioning and dynamic bucketing technology, and combines various energy storage-adaptive compression methods to compress data.

[0012] As an improvement to the above solution, the data acquisition service unit is also used to: perform multi-protocol adaptation using dynamic loading technology of protocol stack; and perform data acquisition using a multi-threaded concurrent acquisition architecture and continuous address batch acquisition technology. The supercomputing real-time processing engine unit is also used to: employ an integrated battery cluster consistency KS test acceleration algorithm to perform in-window anomaly detection by pre-computing the mean / variance matrix, and employ a lightweight LSTM network combined with feature compression technology to predict thermal runaway risk; employ incremental memory computing technology to construct a hierarchical device tree index, employ a differential event-driven model to process only the amount of data change, and preload high-frequency access data into local memory through a hot data prediction mechanism.

[0013] As an improvement to the above solution, the edge supercomputing service module also includes a synchronization service unit; The synchronization service unit is equipped with a breakpoint log stored in a hybrid format of memory queue and solid-state drive, so that synchronization can be automatically restored through a backup link after a network interruption. The synchronization service unit divides the table data into multiple shards for distributed parallel reading.

[0014] As an improvement to the above solution, the supercomputing data platform module also includes a time-series data storage unit; The time-series data storage unit is used to compress the data and store the compressed data in a distributed time-series database; The time-series data storage unit adopts a distributed sharding architecture and uses a consistent hashing algorithm to ensure that the data is evenly distributed. Each shard contains several virtual nodes. The time-series data storage unit integrates FPGA vectorized query within the storage node, and performs time series alignment and interpolation operations through vectorized hardware.

[0015] Compared to existing technologies, the beneficial effects of the application architecture based on a supercomputing power data platform provided by this invention are as follows: By constructing an edge supercomputing service module, a supercomputing power data platform module, and an application layer module, it integrates data acquisition, processing, synchronization, aggregation, and secure transmission, as well as data storage, computation, management, scheduling, algorithm integration, and service provision, and data display, analysis, alarm handling, and operation and maintenance management. This constructs a supercomputing power data platform application framework. Through cloud-edge collaborative layered design and multi-module collaboration, it achieves efficient acquisition, intelligent processing, and security early warning of energy storage data, solving the challenges of real-time processing and risk prevention of massive data in the energy storage industry. The edge supercomputing service module includes at least an acquisition service unit and a data processing unit. The acquisition service unit is used to adjust the data processing based on detected abnormal equipment signals. Based on the sampling frequency, the sampling granularity can be automatically increased at critical moments to ensure calculation accuracy. By adaptively adjusting the sampling frequency, unnecessary data is reduced, improving data processing efficiency and lowering storage costs. The data processing unit uses adaptive differential compression combined with secondary compression to compress the collected data, achieving high compression ratios, reducing storage footprint and costs, and improving data transmission efficiency. The supercomputing data platform module includes at least a supercomputing real-time processing engine unit. This unit performs three-dimensional task partitioning based on device level, data type, and time window, and allocates task partitions to different thread groups based on device correlation for multi-threaded parallel computing, effectively improving data processing efficiency and enhancing system scalability. This invention solves the problems of low data processing efficiency, high storage costs, and insufficient system scalability in existing technologies, enabling efficient processing and intelligent applications of massive amounts of data. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an application architecture based on a supercomputing power data platform provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a supercomputing data platform provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To address the challenges of real-time processing and risk control of massive amounts of data in the energy storage industry, and to achieve efficient data collection, intelligent processing, and security early warning, this invention provides an application architecture based on a supercomputing data platform, including: An edge supercomputing service module is used for data acquisition, processing, synchronization, aggregation, and secure transmission. The edge supercomputing service module includes at least an acquisition service unit and a data processing unit. The acquisition service unit is used to adjust the data sampling frequency based on detected abnormal device signals. The data processing unit is used to compress the acquired data using an adaptive differential compression combined with secondary compression. The supercomputing data platform module is used for data storage, computation, management, scheduling, algorithm integration, and service provision. The supercomputing data platform module includes at least a supercomputing real-time processing engine unit. The supercomputing real-time processing engine unit is used to perform three-dimensional task sharding based on device level, data type, and time window, and to allocate task shards to different thread groups based on device correlation for multi-threaded parallel computation, wherein each thread group processes an independent shard. The application layer module is used for data display, analysis, alarm handling, and operation and maintenance management.

[0019] Specifically, this invention, through a cloud-edge collaborative layered design and multi-module collaboration, enables efficient data acquisition, intelligent processing, and safety early warning for energy storage. Please refer to... Figure 1 This is a schematic diagram of an application architecture based on a supercomputing power data platform. The supercomputing power data platform application architecture of this embodiment mainly consists of an edge supercomputing service module, a supercomputing power data platform module, and an application layer module.

[0020] The edge supercomputing service module includes: a data acquisition service unit, a data processing unit, a synchronization service unit, an aggregation service unit, and a security service unit, which are used to realize data acquisition, processing, synchronization, aggregation, and secure transmission. The supercomputing data platform module includes: a time-series data storage unit, a supercomputing real-time processing engine unit, an offline computing engine unit, a data asset management unit, a task scheduling engine unit, an algorithm service unit, and a data service unit, which are used to realize data storage, computing, management, scheduling, algorithm integration, and service provision. The application layer modules include: monitoring center, analysis center, alarm center, handling center, operation and maintenance center, and management center, which are used for data display, analysis, alarm handling, and operation and maintenance management.

[0021] Specifically, for the edge supercomputing service module: The data acquisition service unit is deployed at the site and supports multiple protocol adaptations, such as IEC104 (power system remote communication standard), Modbus (industrial equipment communication protocol), and IEC61850 (substation protocol for power automation communication system). It can acquire full data from devices within the energy storage power station, including battery cells, BMS (battery management system), PCS (power storage DC unit), and fire alarm control panel, with a data acquisition frequency reaching the second level. A single edge supercomputing server can acquire a maximum data volume of ≥100,000 points. The data acquisition service unit utilizes multi-threaded concurrent acquisition and continuous address batch acquisition technologies to reduce the number of acquisitions, improve acquisition efficiency, and ensure full data acquisition is completed within one second.

[0022] The data processing unit can standardize the collected heterogeneous data and convert it into a unified data format. It also uses an intelligent differential compression algorithm combined with secondary compression technology to achieve high compression ratio storage of data, with a lossless compression ratio of ≥10:1 and a lossy compression ratio of ≥150:1, effectively reducing the overhead of data storage and transmission.

[0023] The synchronization service unit can achieve real-time synchronization of cloud and edge data based on database synchronization mechanisms, such as Binlog (binary log). It also supports breakpoint resume function to ensure the stability and integrity of data transmission, making the time to synchronize the full amount of data less than 1 second, and the synchronization traffic only accounts for 5% of the original data size.

[0024] The aggregation service unit can store the collected data in device-level tables to avoid the merging processing overhead during data aggregation. It also reduces the amount of data transmission through a compression and transmission mechanism, supports the aggregation of data from 2000+ devices, and has no additional data merging processing during the aggregation process, thus improving data aggregation efficiency.

[0025] The security service unit uses SSL (Secure Sockets Layer) encryption technology to encrypt data transmission, ensuring the security and confidentiality of data during transmission. Based on Ceph (distributed storage system)'s redundant backup mechanism and snapshot technology, it achieves disaster recovery backup of data, ensuring data reliability and fault tolerance, and effectively improving the system's annual availability.

[0026] For the supercomputing data platform module: The time-series data storage unit adopts a distributed time-series database, supporting 25 years of online storage of second-level data from all equipment and measurement points of the energy storage power station; through intelligent differential algorithms and secondary compression technology, historical data is compressed and stored, with the storage space occupied being only 5% of the original data size, while supporting millisecond-level real-time data query and second-level random association query supercomputing; The supercomputing real-time processing engine unit is capable of performing second-level real-time basic calculations on massive amounts of data based on multi-threaded parallel computing and in-memory computing technologies. It also supports decompression processing of real-time data, making the processing transparent to applications. Its real-time computing response time is in the second range, meeting the real-time display needs of data dashboards and other applications.

[0027] The offline computing engine unit adopts the Spark+Hive distributed computing framework, which supports efficient processing of large-capacity static datasets. Through data locality optimization, computing and storage optimization strategies, it can improve data development efficiency, reduce R&D personnel costs, and support offline and near real-time processing of 30+ heterogeneous data sources.

[0028] The data asset management unit enables full lifecycle management of data, including metadata management, data standard management, data quality monitoring, and data security management. Through data mapping, lineage analysis, and impact analysis, it achieves visualized management and end-to-end tracking of data, thereby improving data availability and security.

[0029] The task scheduling engine unit has two modes: dependency scheduling and timed scheduling. It visualizes and configures task dependencies through a DAG (Directed Acyclic Graph) to achieve automatic task scheduling and monitoring. When a node fails, the downstream node automatically stops execution to ensure the reliability and stability of the task. It also supports task rerunning or rollback, improving the fault tolerance of the system.

[0030] The algorithm service unit integrates multiple types of algorithms (such as machine learning algorithms and deep learning algorithms) and supports deep learning frameworks such as PyTorch 2.0. Through a distributed task processing engine, it realizes parallel computing and efficient execution of algorithms, improves the prediction accuracy and computational efficiency of models, and provides algorithmic support for safety early warning and status assessment of energy storage power stations.

[0031] The data service unit provides a unified data service bus, supporting the generation of data APIs (Application Programming Interfaces) from data tables, tags, and algorithm models, and connecting with the underlying API gateway to achieve one-click API publishing and management. It supports RESTful interfaces and data forwarding services using multiple protocols, providing efficient data services to meet the data needs of different business scenarios.

[0032] For application layer modules: The monitoring center provides a comprehensive large screen and multi-station monitoring functions to display core data of the power station in real time, including multi-level data monitoring such as national, provincial, single station, cabin, cluster, and individual units. It supports marking the location and operating status of the power station on the map and can refresh key indicator parameters of the power station in real time, realizing comprehensive real-time monitoring of the energy storage power station.

[0033] The analysis center is used to perform multi-dimensional analysis of data from energy storage power stations, including charge and discharge analysis, consistency analysis, temperature analysis, SOC (state of charge) analysis, and SOH (state of health) analysis. The analysis results are displayed through curve reports and holograms, providing data support for power station operation optimization and fault prediction.

[0034] The alarm center is used to aggregate all alarm / early warning information from the power plant and display it in the form of lists and statistical charts. It supports setting flexible notification mechanisms (such as applications, SMS, and voice calls) and can configure different notification recipients for different alarm levels and device types to ensure timely response and handling of abnormal situations.

[0035] The processing center is used to comprehensively process alarm and early warning information based on the fault expert database and the early warning expert database. It supports manual issuance of work orders or automatic generation of work orders, realizing a closed-loop management and control system of alarm / early warning-processing-work order-maintenance, thereby improving the efficiency of fault handling.

[0036] The operations and maintenance center includes functions such as work order management, operations and maintenance map, knowledge base, operations and maintenance scheduling, operations and maintenance plan, and intelligent inspection, realizing the online and intelligent operation and maintenance process, assisting operations and maintenance personnel in on-site repair and management, and improving the efficiency and professionalism of operations and maintenance.

[0037] The management center is used to manage power plants, equipment, accounts, roles, logs, etc. It supports the integration of account permission systems and the setting of secondary passwords to ensure the security of the platform and the standardization of management.

[0038] Furthermore, the implementation method of the application architecture based on the embodiments of the present invention includes: data acquisition and preprocessing, data storage and computation, data management and services, and application display and operation and maintenance.

[0039] Specifically, data acquisition and preprocessing: In the edge supercomputing service module, the acquisition service unit collects data from energy storage power station equipment in real time. After format conversion and compression by the data processing unit, the data is transmitted to the supercomputing power data platform module through the synchronization service unit and the aggregation service unit. Data storage and computation: The time-series data storage unit of the supercomputing power data platform module compresses and stores the data. The supercomputing real-time processing engine unit and the offline computing engine unit perform computation processing on the real-time data and offline data, respectively, to generate various analysis results and early warning data. Data management and services: The data asset management unit of the supercomputing power data platform module performs full lifecycle management of the data. The algorithm service unit uses the data to train and run algorithm models. The data service unit provides data and algorithm results to the application layer module in the form of APIs, etc. Application display and operation and maintenance: The application layer module displays data and analysis results through functional modules such as the monitoring center and the analysis center. The operation and maintenance center realizes the operation and maintenance management of the power station, and the disposal center handles abnormal situations, forming a complete business closed loop.

[0040] In a specific implementation of this invention, an edge supercomputing server (i.e., an edge supercomputing service module) is deployed at the energy storage power station site. The edge supercomputing server, through a data acquisition service unit, utilizes multi-threaded concurrent acquisition technology to collect 7 million measurement point data points per second. The data processing unit converts the data into a standard format and compresses it, achieving a compression ratio of 20:1. The synchronization service unit, based on Binlog incremental synchronization technology, synchronizes the compressed data to the cloud in real time, with a synchronization time of less than 1 second. Furthermore, a distributed time-series database and a Spark+Hive computing cluster are deployed in the cloud. The time-series data storage unit performs secondary compression on the data transmitted from the edge, reducing storage space to 5% of the original data and supporting millisecond-level real-time queries. The supercomputing power real-time processing engine unit utilizes multi-threaded parallel computing to perform second-level calculations on the real-time data, generating real-time statistical results for display in the monitoring center. The offline computing engine unit analyzes historical data and trains a battery status assessment model. The data asset management unit manages metadata, data standards, and data quality to ensure data accuracy and consistency. Furthermore, in terms of application layer modules, the monitoring center displays core power plant data in real time on a comprehensive large screen, while the analysis center performs multi-dimensional analysis of the data and generates curve reports. When the alarm center detects abnormal data, it triggers an alarm notification. The handling center generates a handling work order based on the fault expert database, and the operation and maintenance center receives the work order and performs operation and maintenance processing, realizing closed-loop management from alarm to handling. Therefore, the supercomputing data platform application architecture of this embodiment effectively solves the problems in the prior art through efficient data processing, low storage costs, powerful scalability, and intelligent operation and maintenance management, demonstrating significant economic and social benefits.

[0041] In a specific application example, please refer to Figure 2 This is an application demonstration of a supercomputing data platform. The supercomputing data center integrates multiple modules, including a task scheduling engine, algorithm services, a data development platform, data asset management, a real-time supercomputing processing engine, and time-series data storage. The data development platform provides functions such as data synchronization, data development, operation and maintenance monitoring, data management, and access control. The real-time supercomputing processing engine enables supercomputing data querying, processing, real-time computation, and data export. Data asset management encompasses data mapping, data standards, data models, and data quality. At the edge, it integrates various functions such as acquisition services, data processing, compression services, aggregation services, synchronization services, and security services. At the application end, it provides API gateway, data services, and visualization capabilities. Through collaborative processing and scheduling among these modules, data processing efficiency is effectively improved, storage costs are reduced, and system scalability is enhanced.

[0042] Furthermore, the data acquisition service unit adopts a dynamic adjustment mechanism for the acquisition frequency, specifically adjusting the data sampling frequency based on the detected abnormal signals from the equipment.

[0043] As one optional embodiment, adjusting the data sampling frequency based on the detected device anomaly signal includes: Acquire device operation data and determine the current state of the device based on the device operation data; if the current state of the device is normal, set the data sampling frequency to a first frequency; if the current state of the device is abnormal, set the data sampling frequency to a second frequency; the second frequency is higher than the first frequency. Based on the device operation data, a preset neural network model is used to predict the changing trend of the device status to obtain the predicted device status; if the predicted device status is abnormal, the data sampling frequency is set to the second frequency.

[0044] Specifically, under normal conditions, the device data acquisition frequency is set to a first frequency for basic data monitoring during stable device operation. Under abnormal conditions, i.e., when device operating data indicates an abnormal device status (e.g., battery cluster voltage deviation exceeding a threshold (e.g., ±200mV)) or abnormal temperature (e.g., >55℃), a dynamic adjustment algorithm for the acquisition frequency is triggered, adjusting the data acquisition frequency to a second frequency. For example, the first frequency is set to 1 second / time, and the second frequency is set to 0.1 seconds / time.

[0045] Furthermore, in addition to adjusting the sampling frequency when an abnormal device state is detected based on real-time data, the sampling frequency is also adjusted in advance based on predicted device state data. Specifically, based on an event-driven state machine model, combined with deep learning to predict the changing trends of device state, such as using an LSTM neural network model to predict battery consistency degradation, the data sampling frequency is increased to a second frequency in advance when an abnormal device state is predicted.

[0046] In a specific example, the dynamic sampling frequency adjustment process includes: first, status monitoring is performed, and AI inference is used at the edge to analyze parameters such as battery cluster consistency and SOH (state of health) in real time, and to determine whether the device has entered an abnormal state based on these parameters; when the abnormal state lasts for more than 500ms, the edge controller sends a frequency adjustment command to the acquisition service unit; the acquisition service unit dynamically modifies the timer interrupt frequency of the data acquisition device and reallocates thread resources to adapt to the high-frequency acquisition requirements.

[0047] As one optional embodiment, the data acquisition service unit is further configured to: perform multi-protocol adaptation using dynamic loading technology of protocol stack; and perform data acquisition using a multi-threaded concurrent acquisition architecture and continuous address batch acquisition technology.

[0048] Specifically, the data acquisition service unit adopts dynamic protocol stack loading technology to support the parsing and data interaction of the following protocols: IEC 61850 Protocol: Implements the MMS (Manufacturing Message Specification) protocol stack through the libiec61850 library to parse the logical nodes (LN) and data objects (such as the switching status of XCBR circuit breakers and the voltage values ​​of MMXU measurement units) of power equipment. BMS CAN Protocol: Based on the STM32 microcontroller, the CAN controller realizes data acquisition of the battery management system, supports the CAN FD protocol, and can parse parameters such as cell voltage, temperature, and SOC; Modbus RTU / TCP protocol: Communicates with the energy storage converter (PCS) via serial port or Ethernet interface, and uses CRC checksum and register mapping table to parse data frames.

[0049] It should be noted that the dynamic loading mechanism of the protocol stack allows for automatic upgrades of the protocol version by adding or updating the protocol parsing module through the configuration file without restarting the system.

[0050] In a specific example, the protocol parsing process includes: after the edge supercomputing server starts up, it scans for devices supporting the IEC 61850 protocol within the local area network via broadcast frames, obtains the device description file (ICD), and parses its data model; then, it performs protocol initialization, dynamically loading the corresponding protocol parsing module according to the device type, such as loading the CANopen library for the BMS CAN protocol, configuring communication parameters such as baud rate, data bits, and parity bits; finally, it performs data mapping, establishing a mapping relationship between the device register address and the internal data structure of the system.

[0051] Furthermore, the multi-threaded concurrent data acquisition architecture specifically includes thread pool management, secure data transmission, and exception handling. In thread pool management, a TaskPool concurrent task scheduling mechanism is used to create 16 independent acquisition threads, each responsible for data acquisition from 12 device channels, supporting dynamic load balancing. For secure data transmission, the acquired data is encapsulated using the Sendable data structure, and asynchronous locks (AsyncLock) are used to achieve data synchronization between threads to avoid data contention. In exception handling, a retry mechanism is implemented; automatic retransmission occurs when a single acquisition fails, and if three consecutive failures occur, a log is logged and an alarm is triggered, while a backup communication link is switched.

[0052] Specifically, the execution logic of multi-threaded concurrent acquisition includes: in task allocation, the thread pool allocates acquisition tasks according to the device type and the number of channels, for example, each thread is responsible for acquiring data from 12 CAN channels or 24 Modbus registers; in concurrent acquisition, threads read data from multiple devices simultaneously through asynchronous I / O interfaces, for example, using the libiec61850's MMS_ReadVariable service to read data from multiple logical nodes in batches; in data caching, the acquired data is first stored in the thread's local buffer, and then merged into the global shared buffer through atomic operations to ensure data consistency.

[0053] Furthermore, the continuous address batch acquisition technology utilizes the GPMC parallel interface of the TI AM62x processor in terms of hardware configuration, supporting 8-bit / 16-bit data buses to achieve high-speed access to the continuous address space, with a rate ≥100MB / s. For memory management, a dual-buffer mechanism (BUF1 and BUF2) is used. When BUF1 is full, it automatically switches to BUF2 for acquisition, simultaneously triggering DMA (Direct Memory Access) transfer to write BUF1 data to the storage module, ensuring uninterrupted acquisition. Regarding data compression, differential compression + dictionary compression algorithms are applied to the continuously acquired data at the edge, achieving a compression ratio ≥20:1, reducing data transmission bandwidth consumption.

[0054] Specifically, the implementation process of continuous address batch acquisition technology includes: first, configuring the address range, where the user specifies the continuous address range through the management interface, and the system automatically calculates the required acquisition cycle and buffer size; starting the DMA controller, directly transferring the data acquired by the GPMC interface to the memory buffer to avoid CPU intervention and improve acquisition efficiency; after the batch acquisition is completed, verifying the data integrity through the CRC32 check algorithm, and if the check fails, re-acquiring the data of the address segment.

[0055] The data acquisition service unit of this invention supports multi-protocol adaptation, with a data acquisition frequency reaching the second level. The maximum data acquisition capacity of a single edge supercomputing server is ≥100,000 points. Furthermore, it adopts multi-threaded concurrent acquisition and continuous address batch acquisition technology, which effectively improves data processing efficiency.

[0056] As one optional embodiment, the method of compressing the collected data using adaptive differential compression combined with secondary compression includes: Calculate the data volatility index based on the collected data; A differential strategy is determined based on the data volatility index to compress the data and obtain a first compression result; the differential strategy includes real-time sampling differential and window effective value differential; Based on the first compression result and the dynamic dictionary, a secondary compression is performed by selecting an entropy coding method according to the data characteristics. The discrete state data uses Huffman coding, and the continuous difference data uses arithmetic coding. The dynamic dictionary uses the Least Recently Used (LRU) page replacement algorithm to maintain the high-frequency difference pattern.

[0057] As one optional embodiment, determining the differential strategy for data compression based on the data volatility index includes: Select a differencing strategy based on the data volatility index and determine the differencing benchmark value; The dynamic quantization step size is determined based on the data volatility index. In lossy compression mode, a lightweight LSTM model is used to predict the data at the next time point based on the differential reference value, and the prediction residual is calculated as the differential object. The data is then encoded in combination with the dynamic quantization step size to compress the data.

[0058] The data volatility index is calculated as the ratio of the standard deviation to the mean of the differences between adjacent sampling points.

[0059] In a specific example, a data volatility index is calculated based on the collected data; when the data volatility index is greater than a preset volatility threshold, real-time sampling differential is used for data compression to obtain a first compression result; when the data volatility index is not greater than the preset volatility threshold, window effective value differential is used for data compression to obtain a first compression result.

[0060] Specifically, the data processing unit's compression technology adopts a three-level processing architecture: a data preprocessing layer, an intelligent differential compression layer, and a secondary enhanced compression layer. High compression ratios are achieved through dynamic feature analysis, adaptive differential strategies, and a combination of multiple algorithms. It employs an adaptive differential compression algorithm combined with secondary compression technology to achieve a lossless compression ratio ≥10:1 and a lossy compression ratio ≥150:1. The adaptive differential compression algorithm includes data volatility index calculation, specifically based on short-time Fourier transform and the 3σ principle to determine the differential baseline value switching strategy. The secondary compression technology includes a dynamic dictionary and adaptive entropy coding. The dynamic dictionary uses the Least Recently Used (LRU) algorithm to maintain high-frequency differential patterns, and the adaptive entropy coding automatically selects Huffman coding or arithmetic coding based on the data type.

[0061] Furthermore, in the implementation of the adaptive differential compression algorithm, the data type is first identified through the signal processing module, such as slowly varying voltage / current data, abruptly changing temperature / pressure data, and discrete state position data, thereby constructing data feature labels, such as volatility index, periodicity index, and effective value range. The data processing also includes analyzing the data frequency characteristics using Short-Time Fourier Transform (STFT) and calculating the data volatility threshold using the 3σ principle. Regarding the selection of the differential strategy, real-time sampling value differential is used for data with large volatility indices, with the reference value set as the previous sampling point; for data with small volatility indices, the most recent effective value differential is used, with the reference value set as the effective value of the previous window. The selection of the differential strategy is specifically implemented through a state machine algorithm to achieve automatic switching of the differential mode. In a specific example, when the data volatility exceeds the threshold, such as when the difference between three consecutive points is >5% of the full scale, real-time differential is triggered; when the volatility of ten consecutive windows is <1%, it switches to effective value differential.

[0062] In differential compression, the differential sequence is first quantized, and then encoded. The dynamic quantization step size is set based on data feature labels. For slowly varying data, the quantization step size is set to 0.5% of full scale; for abruptly changing data, hierarchical quantization is used. After quantization, the differential values ​​are converted into integer codes, reducing data precision redundancy. During differential sequence encoding, variable word length encoding is used in lossless compression scenarios, while run-length encoding is used in lossy compression scenarios. Continuous repetitions are encoded as "count + value".

[0063] Furthermore, for lossy compression scenarios, the differential compression algorithm incorporates an error compensation mechanism. Specifically, it introduces prediction residual correction, employs a lightweight LSTM (Long Short-Term Memory) model to predict the data at the next time point, and calculates the residual as the difference object. The calculation is as follows: in, The data is sampled at time t. A lightweight neural network model is used to generate the predicted residual sequence, and combined with a dynamic quantization step size, a balance between data accuracy and compression ratio can be achieved.

[0064] Furthermore, the second-level compression includes dictionary construction and matching, as well as entropy coding enhancement. In dictionary construction and matching, a dynamic dictionary library is maintained for high-frequency scenarios. Based on the most recent 10 window data points, high-frequency differential patterns are statistically analyzed (e.g., in the equalization charging phase of energy storage batteries, the differential sequence is mostly 0±1). Dictionary replacement is enabled when the dictionary matching rate is ≥60%: compressed data = dictionary index + unmatched remainder. The dictionary is updated using the LRU algorithm to ensure that hot patterns are stored first. In entropy coding enhancement, for the differentially compressed data, the entropy coding method is selected based on the data characteristics. Huffman coding is used for discrete-state data to utilize its symbol frequency distribution characteristics; algorithmic coding is used for continuous differential data to efficiently compress the 0-probability interval.

[0065] As one optional embodiment, the step of performing three-dimensional task sharding based on device level, data type, and time window, and allocating task shards to different thread groups based on device correlation, includes: Based on the physical structure of the energy storage system, the task is divided into several equipment levels to obtain the first task segment; the equipment level includes power station, container, and battery cell; Based on the first task fragment, the task is divided into three data types: status data, running data, and fault data to obtain the second task fragment. Based on the second task slice, the task is divided into three time window types: real-time window, trend window, and statistics window, to obtain the third task slice; Based on the third task segmentation, the energy storage device correlation factor is calculated, and the task segments corresponding to the energy storage device correlation factors that are greater than the preset correlation threshold are allocated to the same node for operation, so as to achieve dynamic load balancing; wherein, the energy storage device correlation factor is calculated by the ratio of data interaction frequency to the cable length between devices.

[0066] Specifically, this invention designs a data-driven task partitioning strategy and constructs a three-dimensional task sharding model. This model can perform three-dimensional task partitioning based on the physical structure of the energy storage system (power station → container → cell), data type (status data / operational data / fault data), and time window (1s real-time window / 10s trend window / 1min statistical window). Each thread group is responsible for an independent "device-data-time" task sharding. For example, the cell voltage fluctuation analysis task is divided into 200 containers × 16 cells × 1s real-time window, totaling 3200 parallel subtasks.

[0067] Furthermore, in load balancing, an energy storage device correlation factor is introduced, calculated as: Energy Storage Device Correlation Factor = Data Interaction Frequency / Cable Length Between Devices. This prioritizes allocating computational tasks from strongly correlated devices to CPU cores on the same node, reducing cross-core data transfer latency. Specifically, this is achieved through an OpenMP dynamic thread pool, binding 8 logical threads to each physical core, resulting in thread switching overhead ≤1μs.

[0068] As one optional embodiment, the supercomputing real-time processing engine unit is also used to: employ an integrated battery cluster consistency KS test acceleration algorithm to perform in-window anomaly detection by pre-computing the mean / variance matrix, and employ a lightweight LSTM network combined with feature compression technology to predict thermal runaway risk; employ incremental memory computing technology to construct a hierarchical device tree index, employ a differential event-driven model to process only the amount of data change, and preload high-frequency access data into local memory through a hot data prediction mechanism.

[0069] Specifically, the supercomputing real-time processing engine unit adopts a combination of algorithm and hardware acceleration: it integrates a battery cluster consistency KS test acceleration algorithm, which realizes anomaly detection within a preset window by pre-computing the mean / variance matrix, making the processing time of a single cluster ≤80μs; ​​and it adopts a lightweight LSTM network combined with feature compression technology to achieve a single sample inference time of ≤2μs for thermal runaway risk prediction, and supports concurrent computing of tens of thousands of battery clusters when deployed on a GPU cluster.

[0070] The supercomputing real-time processing engine unit employs incremental in-memory computing technology: it constructs a hierarchical device tree index, uses a differential event-driven model to process only data changes, dynamically allocates computing resources through event queue priority scheduling (e.g., fault events have higher priority than status events), and preloads frequently accessed data into local memory through a hotspot data prediction mechanism, achieving a memory access latency reduction of over 40%. Specifically, the hotspot data prediction mechanism utilizes an energy storage device operating cycle model to predict hotspot data.

[0071] Furthermore, the supercomputing real-time processing engine unit is also equipped with a heterogeneous computing acceleration architecture, which uses CPU-GPU collaborative scheduling: routine computing (data verification, protocol parsing) is handled by the CPU, while intensive computing (matrix operations, neural network inference) is offloaded to the GPU. By grouping multiprocessors, the computing tasks of each energy storage container are bound to an independent computing unit to avoid resource contention and improve computing efficiency.

[0072] As one optional embodiment, the supercomputing data platform module also includes an offline computing engine unit; The offline computing engine unit adopts a distributed computing framework, in which energy storage vectorization operators are set for high-frequency time-series data of energy storage. The energy storage vectorization operators include multi-level time window aggregation operators, parameter correlation coefficient calculation operators, and anomaly detection operators. The multi-level time window aggregation operator is used for multi-level aggregation based on the energy storage device level and the time window. The parameter correlation coefficient calculation operator is used to calculate the Pearson correlation coefficient of two energy storage parameters and is optimized by single instruction multiple data (SIMD) instructions. The anomaly detection operator is used to integrate an energy storage anomaly detection algorithm.

[0073] As one of the optional embodiments, the offline computing engine unit is equipped with an energy storage protocol parser. The energy storage protocol parser maintains an energy storage protocol feature library through the Hive metadata database and adopts a state machine and regular expression hybrid parsing strategy to support dynamic access of multiple heterogeneous data sources. The offline computing engine unit uses Hive storage, which employs four-level composite partitioning and dynamic bucketing technology, and combines various energy storage-adaptive compression methods to compress data.

[0074] Specifically, the offline computing engine unit adopts the Spark+Hive distributed computing framework, supporting offline and near real-time processing of 30+ heterogeneous data sources. To address the strong temporal correlation and computational timeliness of energy storage data, the offline computing engine unit customizes the Spark task scheduler: a dynamic task priority adjustment mechanism and vectorized computation operator extensions.

[0075] For the dynamic task priority adjustment mechanism, industry characteristic tags are added to each computing task, which are synchronized to the Spark task scheduler through Hive metadata. A rule-based scheduling strategy is designed to prioritize the allocation of computing resources to urgent tasks, allocate the remaining resources to regular tasks, and dynamically adjust the task interval for near real-time tasks based on the device status.

[0076] For the extension of vectorized computation operators, new vectorized operators for energy storage have been added to Spark SQL for high-frequency time-series data: Multi-level time window aggregation operator: supports multi-level aggregation by energy storage device level and time window, which improves efficiency compared to ordinary window functions; Parameter correlation coefficient calculation operator: calculates the Pearson correlation coefficient of two energy storage parameters (such as voltage and temperature), which can improve the calculation speed through SIMD instruction optimization; Anomaly detection operator: integrates energy storage anomaly detection algorithm, which can be directly called in Spark DataFrame.

[0077] Furthermore, based on the Hive storage engine, and considering the lifecycle characteristics and query patterns of energy storage data, storage partitioning strategies and energy storage data compression encoding adaptations are implemented. Storage partitioning strategy optimization includes a four-level composite partition design and dynamic bucketing technology. The four-level composite partitioning specifically includes power plant ID, container ID, date, and time period. Dynamic bucketing technology buckets frequently related fields and leverages Hive's bucketed table features to achieve join queries.

[0078] The offline computing engine unit in this embodiment of the invention employs a multi-protocol adaptive parsing engine, including an energy storage protocol parser (ESS-Adapter). This ESS-Adapter maintains an energy storage protocol feature library through the Hive metadata database and uses a hybrid parsing strategy of state machine + regular expression, supporting dynamic access from 30+ heterogeneous data sources. It adopts composite partitioning and hierarchical storage. The Hive storage engine uses four-level composite partitioning and dynamic bucketing technology, combined with energy storage adaptation compression schemes such as ZSTD+Delta encoding and RLE+dictionary encoding, reducing storage usage to less than 5% of the original data.

[0079] As one optional embodiment, the edge supercomputing service module further includes a synchronization service unit; The synchronization service unit is equipped with a breakpoint log stored in a hybrid format of memory queue and solid-state drive, so that synchronization can be automatically restored through a backup link after a network interruption. The synchronization service unit divides the table data into multiple shards for distributed parallel reading.

[0080] Specifically, the synchronization service unit adopts a three-tier synchronization architecture, combining real-time log parsing and intelligent flow control technology to achieve highly reliable, low-latency data synchronization. This includes a data capture layer, a data transmission layer, and a data replay layer. The data capture layer is used to capture incremental logs in real time and perform full scan optimization. The full scan optimization uses a lock-free snapshot algorithm to divide table data into multiple shards based on primary or unique keys, achieving second-level synchronization of all data through distributed parallel reading. The data transmission layer performs flow control. The data replay layer ensures transaction consistency through distributed transaction sorting, write-ahead logging, and a dual-write buffer mechanism.

[0081] The synchronization service unit implements breakpoint resumption functionality based on dual logging, an intelligent retry mechanism, and a parallel recovery algorithm. Dual logging includes a synchronization location log and a data verification log. The intelligent retry mechanism automatically switches to a backup link when the network is interrupted, with retry intervals increasing linearly. If five consecutive retries fail, a global consistency check is triggered, automatically repairing data differences by comparing the verification values ​​of the source and target databases. The parallel recovery algorithm restarts sharded parallel reading during full synchronization interruptions, skipping synchronized data blocks using historical verification logs. During incremental synchronization interruptions, it re-retrieves the binary log from the breakpoint and continues replaying it using cached uncommitted transactions.

[0082] The synchronization service unit of this invention divides table data into multiple shards and adopts distributed parallel reading technology to achieve a data throughput of ≥10GB / s during the full synchronization phase. The synchronization service unit sets up a breakpoint log that is stored in a hybrid memory queue and SSD (solid-state drive) and automatically resumes synchronization through a backup link after a network interruption, with a recovery time of ≤200ms. The synchronization service unit uses ZSTD dictionary compression combined with field-level differential transmission to ensure that the synchronization traffic accounts for less than 5% of the original data size.

[0083] As one optional embodiment, the supercomputing data platform module further includes a time-series data storage unit; The time-series data storage unit is used to compress the data and store the compressed data in a distributed time-series database; The time-series data storage unit adopts a distributed sharding architecture and uses a consistent hashing algorithm to ensure that the data is evenly distributed. Each shard contains several virtual nodes. The time-series data storage unit integrates FPGA vectorized query within the storage node, and performs time series alignment and interpolation operations through vectorized hardware.

[0084] Specifically, the time-series data storage unit adopts a three-tiered distributed architecture, combining intelligent data scheduling and hardware acceleration technologies to achieve a balance between ultra-long-term storage and high-performance querying. The time-series data storage unit includes a data preprocessing layer, an intelligent compression layer, and a hardware acceleration layer. The data preprocessing layer divides the data into several shards based on the device ID, with each shard containing multiple virtual nodes, and uses a consistent hashing algorithm to ensure even data distribution. It also includes delta encoding of the data. The intelligent compression layer includes dynamic dictionary generation and a hybrid compression algorithm. The time-series data storage unit contains a distributed query engine that can break down query tasks into multiple subtasks, process them in parallel through data sharding, and use pipeline aggregation to reduce intermediate result storage. The distributed query engine includes a vector computation unit that uses an FPGA chip to implement time series alignment and interpolation operations, supports the SIMD instruction set, and can perform 16 floating-point operations per cycle.

[0085] The time-series data storage unit in this embodiment of the invention adopts a distributed sharding architecture, achieving balanced data distribution through a consistent hashing algorithm. Each shard contains multiple virtual nodes, supporting automatic load balancing and fault recovery. The intelligent compression layer includes a dynamic dictionary generation module that generates a ZSTD compression dictionary based on historical data statistics of high-frequency field values, combined with delta encoding to achieve a lossless compression ratio ≥15:1. The vector computation unit uses an FPGA chip to implement time series alignment and interpolation operations, supports the SIMD instruction set, and processes 16 floating-point operations per cycle. This embodiment supports online storage of 25 years of data, occupying less than 5% of the original data size, and supports millisecond-level real-time data queries and second-level random association queries in supercomputing.

[0086] The beneficial effects of the application architecture based on a supercomputing power data platform provided by this invention are as follows: By designing an edge supercomputing service module, a supercomputing power data platform module, and an application layer module, it integrates data acquisition, processing, synchronization, aggregation, and secure transmission, as well as data storage, computation, management, scheduling, algorithm integration, and service provision, and data display, analysis, alarm handling, and operation and maintenance management. This constructs a supercomputing power data platform application framework. Through cloud-edge collaborative layered design and multi-module collaboration, it achieves efficient acquisition, intelligent processing, and security early warning of energy storage data, solving the challenges of real-time processing and risk prevention of massive data in the energy storage industry. The edge supercomputing service module includes at least an acquisition service unit and a data processing unit. The acquisition service unit is used to adjust the data sampling frequency according to detected abnormal equipment signals. The system can automatically increase sampling granularity at critical moments to ensure computational accuracy. By adaptively adjusting the sampling frequency, it reduces unnecessary data, improves data processing efficiency, and lowers storage costs. The data processing unit uses adaptive differential compression combined with secondary compression to compress the collected data, achieving high compression ratios, reducing storage footprint and costs, and improving data transmission efficiency. The supercomputing data platform module includes at least a real-time supercomputing engine unit. This unit performs three-dimensional task partitioning based on device level, data type, and time window, and allocates task partitions to different thread groups based on device correlation for multi-threaded parallel computing, effectively improving data processing efficiency and enhancing system scalability. This invention solves the problems of low data processing efficiency, high storage costs, and insufficient system scalability in existing technologies, enabling efficient processing and intelligent applications of massive amounts of data.

[0087] The databases involved in the various embodiments provided by this invention may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the various embodiments provided by this invention may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0088] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An application architecture based on a supercomputing power data platform, characterized in that, include: The edge supercomputing service module is used for data collection, processing, synchronization, aggregation, and secure transmission. The edge supercomputing service module includes at least a data acquisition service unit and a data processing unit; the data acquisition service unit is used to adjust the data sampling frequency according to the detected abnormal device signals; the data processing unit is used to compress the acquired data using an adaptive differential compression combined with a secondary compression method. The supercomputing data platform module is used for data storage, computation, management, scheduling, algorithm integration, and service provision. The supercomputing data platform module includes at least a supercomputing real-time processing engine unit. The supercomputing real-time processing engine unit is used to perform three-dimensional task sharding based on device level, data type, and time window, and to allocate task shards to different thread groups based on device correlation for multi-threaded parallel computation, wherein each thread group processes an independent shard. The application layer module is used for data display, analysis, alarm handling, and operation and maintenance management.

2. The application architecture based on a supercomputing power data platform as described in claim 1, characterized in that, The step of adjusting the data sampling frequency based on the detected abnormal device signals includes: Acquire device operation data and determine the current state of the device based on the device operation data; if the current state of the device is normal, set the data sampling frequency to a first frequency; if the current state of the device is abnormal, set the data sampling frequency to a second frequency; the second frequency is higher than the first frequency. Based on the device operation data, a preset neural network model is used to predict the changing trend of the device status to obtain the predicted device status; if the predicted device status is abnormal, the data sampling frequency is set to the second frequency.

3. The application architecture based on a supercomputing power data platform as described in claim 1, characterized in that, The method of compressing the collected data using adaptive differential compression combined with secondary compression includes: Calculate the data volatility index based on the collected data; A differential strategy is determined based on the data volatility index to compress the data and obtain a first compression result; the differential strategy includes real-time sampling differential and window effective value differential; Based on the first compression result and the dynamic dictionary, a secondary compression is performed by selecting an entropy coding method according to the data characteristics. The discrete state data uses Huffman coding, and the continuous difference data uses arithmetic coding. The dynamic dictionary uses the Least Recently Used (LRU) page replacement algorithm to maintain the high-frequency difference pattern.

4. The application architecture based on a supercomputing power data platform as described in claim 3, characterized in that, The step of determining a differencing strategy for data compression based on the data volatility index includes: Select a differencing strategy based on the data volatility index and determine the differencing benchmark value; The dynamic quantization step size is determined based on the data volatility index. In lossy compression mode, a lightweight LSTM model is used to predict the data at the next time point based on the differential reference value, and the prediction residual is calculated as the differential object. The data is then encoded in combination with the dynamic quantization step size to compress the data.

5. The application architecture based on a supercomputing power data platform as described in claim 1, characterized in that, The step of performing three-dimensional task fragmentation based on device level, data type, and time window, and allocating task fragments to different thread groups based on device correlation, includes: Based on the physical structure of the energy storage system, the task is divided into several equipment levels to obtain the first task segment; the equipment level includes power station, container, and battery cell; Based on the first task fragment, the task is divided into three data types: status data, running data, and fault data to obtain the second task fragment. Based on the second task slice, the task is divided into three time window types: real-time window, trend window, and statistics window, to obtain the third task slice; Based on the third task segmentation, the energy storage device correlation factor is calculated, and the task segments corresponding to the energy storage device correlation factors that are greater than the preset correlation threshold are allocated to the same node for operation, so as to achieve dynamic load balancing; wherein, the energy storage device correlation factor is calculated by the ratio of data interaction frequency to the cable length between devices.

6. The application architecture based on a supercomputing power data platform as described in claim 1, characterized in that, The supercomputing data platform module also includes an offline computing engine unit; The offline computing engine unit adopts a distributed computing framework, in which energy storage vectorization operators are set for high-frequency time-series data of energy storage. The energy storage vectorization operators include multi-level time window aggregation operators, parameter correlation coefficient calculation operators, and anomaly detection operators. The multi-level time window aggregation operator is used for multi-level aggregation based on the energy storage device level and the time window. The parameter correlation coefficient calculation operator is used to calculate the Pearson correlation coefficient of two energy storage parameters and is optimized by single instruction multiple data (SIMD) instructions. The anomaly detection operator is used to integrate an energy storage anomaly detection algorithm.

7. The application architecture based on a supercomputing power data platform as described in claim 6, characterized in that, The offline computing engine unit is equipped with an energy storage protocol parser. The energy storage protocol parser maintains an energy storage protocol feature library through the Hive metadata database and adopts a state machine and regular expression hybrid parsing strategy to support dynamic access of multiple heterogeneous data sources. The offline computing engine unit uses Hive storage, which employs four-level composite partitioning and dynamic bucketing technology, and combines various energy storage-adaptive compression methods to compress data.

8. The application architecture based on a supercomputing power data platform as described in claim 1, characterized in that, The data acquisition service unit is also used to: perform multi-protocol adaptation using dynamic loading technology of protocol stack; and perform data acquisition using a multi-threaded concurrent acquisition architecture and continuous address batch acquisition technology. The supercomputing real-time processing engine unit is also used to: employ an integrated battery cluster consistency KS test acceleration algorithm, perform in-window anomaly detection by pre-computing mean / variance matrix, and perform thermal runaway risk prediction by employing a lightweight LSTM network combined with feature compression technology. Incremental in-memory computing technology is used to build a hierarchical device tree index. A differential event-driven model is used to process only the amount of data change, and a hot data prediction mechanism is used to preload frequently accessed data into local memory.

9. The application architecture based on a supercomputing power data platform as described in claim 1, characterized in that, The edge supercomputing service module also includes a synchronization service unit; The synchronization service unit is equipped with a breakpoint log stored in a hybrid format of memory queue and solid-state drive, so that synchronization can be automatically restored through a backup link after a network interruption. The synchronization service unit divides the table data into multiple shards for distributed parallel reading.

10. The application architecture based on a supercomputing power data platform as described in claim 1, characterized in that, The supercomputing data platform module also includes a time-series data storage unit; The time-series data storage unit is used to compress the data and store the compressed data in a distributed time-series database; The time-series data storage unit adopts a distributed sharding architecture and uses a consistent hashing algorithm to ensure that the data is evenly distributed. Each shard contains several virtual nodes. The time-series data storage unit integrates FPGA vectorized query within the storage node, and performs time series alignment and interpolation operations through vectorized hardware.

Images