A Distributed Storage and Processing Method for Energy Big Data

By generating a unified semantic tensor and building a hybrid index at edge nodes, combined with hot and cold data layering and adaptive synchronization, the cross-modal semantic gap of multimodal data in power grid field and the write blocking and false data injection problems in weak network high-concurrency environments are solved, achieving efficient and secure data processing and storage.

CN123115232APending Publication Date: 2026-11-17XIAN BINGCHENG RUIXUAN GREEN ENERGY TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611423242.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for multimodal energy big data storage in power grid sites suffer from cross-modal semantic gaps, write blocking in weak network high-concurrency environments, and fake data injection attacks, making it difficult to meet real-time and security requirements.

Method used

Streaming immune cleaning and reputation scoring are performed at edge nodes to generate a unified semantic tensor and build a hybrid index. Combined with hot and cold data layering and an adaptive synchronization mechanism, false data is identified and repaired in real time, achieving cross-modal data fusion alignment and efficient synchronization.

Benefits of technology

It enables low-latency cross-modal data fusion querying, avoids write blocking and loss of critical data, and ensures the authenticity and security of data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115232A_ABST
    Figure CN123115232A_ABST
Patent Text Reader

Abstract

This invention discloses a distributed storage and processing method for energy big data, relating to the fields of energy big data processing and distributed storage technology. The method performs streaming immune cleaning on the raw collected data, uses reputation scoring and Kirchhoff current residuals to jointly determine and repair suspicious data, maps and concatenates verified data into a unified semantic tensor, constructs a hybrid index key based on topological spatial coordinates, collection timestamps, and modal identifiers, and writes it into the storage engine. Data is divided into hot and cold data based on business value scoring. Cold data is subjected to lossy compression followed by weak consistency asynchronous synchronization, while hot data is subjected to preemptive strong consistency synchronization. The sending rate is adjusted according to link bandwidth. This invention reduces cross-modal fusion query latency, avoids write blocking and data loss in weak network environments, and achieves real-time identification and repair of fake data injection attacks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of energy big data processing and distributed storage technology, specifically a method for distributed storage and processing of energy big data. Background Technology

[0002] With the development of new power systems and the integration of power generation, grid, load, and storage, phasor measurement units, inspection image acquisition devices, and equipment log acquisition systems deployed at the power grid site continuously generate massive, high-frequency, and multimodal energy big data. These data types exhibit significant differences in acquisition frequency, data structure, and semantic granularity. The closest existing technical solution to this invention is: deploying heterogeneous dedicated databases at the power grid site to store different modal data in a modal isolation manner. Specifically, a time-series database is used to store time-series measurement data, an object storage service is used to store image and video data, and a retrieval engine is used to store equipment log text data. Each dedicated database establishes an independent index. When the upper-layer system needs to perform cross-modal joint analysis, data is retrieved from each dedicated database through application-layer join table queries and aligned and stitched according to time and spatial dimensions. Regarding data synchronization between the edge and the cloud, a unified consistency protocol is used to perform multi-replica strong consistency synchronization of all data without distinction. To prevent data tampering, blockchain notarization or fully homomorphic encryption is introduced during the data writing process to encrypt the data before storing it on disk.

[0003] The above-mentioned schemes suffer from high response latency due to the isolation of data indexes for each modality and the reliance on manual alignment at the application layer for cross-modal joint queries. This makes it difficult to meet the real-time requirements under power grid fault scenarios. Furthermore, the use of a unified strong consistency protocol for synchronization of all data indiscriminately can easily lead to edge write blocking or even loss of critical data in remote renewable energy power plants with weak networks and high concurrency environments. Additionally, the use of heavy cryptographic methods such as blockchain or fully homomorphic encryption to prevent data tampering results in a sharp increase in computation and storage overhead with the amount of data, making it unsuitable for high-concurrency write scenarios of high-frequency time-series data. Therefore, we propose a distributed storage and processing method for energy big data. Summary of the Invention

[0004] The purpose of this invention is to provide a distributed storage and processing method for energy big data.

[0005] The technical problems to be solved by this invention include the following three aspects: First, how to eliminate the cross-modal semantic gap between millisecond-level phasor measurement time-series data, inspection image and video data and equipment log text data at the power grid site, so that the distributed storage system can complete the fusion and alignment query of multimodal data with low latency; Second, how to overcome the edge write blocking and critical data loss problems caused by existing distributed consensus protocols in the weak network and high concurrency environment that is common in remote new energy power stations; Third, how to achieve real-time identification and repair of fake data injection attacks at the underlying layer of data storage without introducing heavy security mechanisms such as blockchain or fully homomorphic encryption.

[0006] To address the aforementioned technical problems, this invention provides a distributed storage and processing method for energy big data, comprising: acquiring raw data uploaded by a data acquisition terminal at an edge node in the power grid field; the raw data including phasor measurement data, inspection image and video data, and equipment log text data; performing streaming immune cleaning on the raw data at the edge node to obtain verified data; the streaming immune cleaning includes feature encoding of the raw data to obtain feature vectors, calculating a reputation score based on the feature vectors, and jointly judging the raw data by combining the Kirchhoff current residuals calculated based on the power grid topology connection relationship; when the data is judged to be suspicious, retrieving historical reliable data from adjacent times within a sliding time window. The suspicious data is interpolated and repaired to obtain the verified data. At the edge node, the verified data is mapped and concatenated into a unified semantic tensor. A hybrid index key is constructed based on the topological space coordinates, collection timestamp, and modality identifier of the verified data. The unified semantic tensor is used as the index value and written to the storage engine according to the hybrid index key. At the edge node, a business value score is calculated based on the verified data. Based on the business value score, the verified data is divided into hot data and cold data. The cold data is lossily compressed and asynchronously synchronized to the cloud storage layer using a weak consistency protocol. The hot data is inserted into a high-priority synchronization queue in a preemptive manner and then synchronized to the cloud storage layer using a strong consistency protocol.

[0007] The present invention also provides a distributed storage and processing system for energy big data corresponding to the above method, for implementing the functions described in each step of the above method.

[0008] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0009] 1. This invention maps and concatenates multimodal data into a unified semantic tensor locally at edge nodes and establishes a hybrid index that combines topological coordinates and timestamps. Cross-modal data fusion alignment queries can be completed in one go without cross-system table joins, which can significantly reduce the response latency of cross-modal fusion queries and meet the real-time requirements of automated joint decision-making in power grid fault scenarios.

[0010] 2. This invention scores data by business value and distinguishes between hot and cold data. For low-value cold data, lossy compression and weak consistency asynchronous synchronization are used, while for high-value hot data, preemptive strong consistency synchronization is used. The synchronization rate is adjusted in combination with link bandwidth. This can ensure that critical data is not lost while avoiding write blocking in weak network environments.

[0011] 3. This invention deploys a lightweight reputation scoring mechanism based on contrastive learning and physical law constraints at the data inflow layer. It identifies and interpolates to repair false injected data in real time before the data is written to the disk. This can ensure the authenticity and security of power grid dispatch and virtual power plant transaction data without introducing heavy cryptographic mechanisms, while keeping the throughput loss at a low level. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall process of distributed storage and processing of energy big data in an embodiment of the present invention;

[0013] Figure 2 This is a schematic diagram of the flow cytometry immune cleaning process in an embodiment of the present invention;

[0014] Figure 3 This is a schematic diagram of the cold and hot stratification and adaptive synchronous scheduling process in an embodiment of the present invention. Detailed Implementation

[0015] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0016] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] I. System Overall Architecture:

[0018] The cloud-edge collaborative distributed storage and processing system provided by this invention comprises four layers from bottom to top: a data acquisition layer, an edge processing layer, a transmission link layer, and a cloud storage and service layer. The data acquisition layer is deployed at the power grid site and includes phasor measurement devices, inspection image and video acquisition devices, and equipment log acquisition devices, corresponding to the electrical measurement, image monitoring, and operation log information sources at the power grid site, respectively. The edge processing layer is deployed within edge gateways or intelligent fusion terminals at power grid sites such as renewable energy plants and substations, and belongs to the edge nodes described in this invention. Internally, it includes a business feature extractor, a reputation scoring and streaming cleaning module, a spatiotemporal semantic tensor construction module, a hybrid index writing module, and a cold / hot tiering and adaptive synchronization scheduling module. These modules are interconnected via a local data bus. The raw data generated by the acquisition layer is first processed... The system enters the reputation scoring and streaming cleaning module to complete security verification. After obtaining the verified data, it enters the spatiotemporal semantic tensor construction module to complete multimodal fusion and hybrid index writing. At the same time, the cold and hot stratification and adaptive synchronization scheduling module continuously monitors the business value of the data and the bandwidth status of the transmission link. The transmission link layer includes a synchronization mechanism based on bidirectional streaming channels and an adaptive rate limiting controller to connect the edge processing layer and the cloud storage layer. The cloud storage and business layer includes a distributed unified index storage cluster, a scheduling decision system, a virtual power plant trading system, and a data analysis system. The cloud storage cluster receives unified semantic tensor data synchronized from edge nodes and maintains a globally unified index for direct use by the upper-layer business system. The reputation scoring module and the cold and hot stratification module also provide feedback updates to the historical sample library, forming a closed loop for continuous model optimization.

[0019] II. Overall Methodology and Flow:

[0020] The method provided by this invention starts from the acquisition end and includes the following steps in sequence: acquiring multimodal raw acquisition data at the edge node and marking the acquisition timestamp and the corresponding power grid topology coordinates; performing streaming immune cleaning and reputation scoring verification on the raw acquisition data; interpolating and repairing the data that fails the verification to obtain verified data; mapping and concatenating the verified data into a unified semantic tensor and writing it into the storage engine according to the hybrid index key; scoring the business value of the verified data and determining hot and cold stratification on the edge node side; performing lossy compression weak consistency asynchronous synchronization or strong consistency preemptive synchronization according to the stratification results; adaptively adjusting the sending rate of the synchronization queue according to the link bandwidth status; synchronizing the data to the cloud storage layer and maintaining a globally unified index; and initiating cross-modal joint queries as needed by the upper-layer business system and directly hitting the unified index to complete the retrieval. In the above steps, the reputation scoring verification step is located before the data flows into the storage engine, reflecting the processing sequence of the present invention of storage and computing integration and writing and cleaning simultaneously.

[0021] III. Multimodal Fusion Storage and Hybrid Indexing Based on Spatiotemporal Unified Semantic Tensors;

[0022] This section of the technical solution is used to address the cross-modal semantic gap problem between multimodal data. In this embodiment, the edge node performs time-aligned slicing on the phasor measurement data stream, inspection image and video stream, and equipment log text stream with a preset time window as the granularity, to obtain phasor measurement data slices, image slices, and log slices within the same spatiotemporal unit. The preset time window is set to one hundred milliseconds in this embodiment to adapt to the time granularity requirements of power grid transient fault analysis. In other embodiments, it can be set to other values ​​according to the real-time requirements of the business scenario.

[0023] The temporal feature mapping function, image feature mapping function, and text feature mapping function are called respectively to map phasor measurement data slices, image slices, and log slices to a feature subspace of the same dimension, obtaining the corresponding feature representations. These representations are then concatenated according to the feature dimensions to generate a unified semantic tensor corresponding to the spatiotemporal unit. The generation method is represented as follows:

[0024] ;

[0025] in, Indicates the first A unified semantic tensor generated by spatiotemporal alignment units , , These represent the phasor measurement data slice, inspection image data slice, and equipment log text data slice obtained by time alignment within the spatiotemporal unit, respectively. , , These represent the feature mapping functions for the corresponding modalities, used to map the original heterogeneous data to a feature subspace of a unified dimension. In this embodiment, they are implemented by pre-trained temporal coding networks, image convolutional coding networks, and text coding networks, respectively. This is an operator that concatenates multiple feature representations along the feature dimension.

[0026] For each unified semantic tensor, the edge node constructs a hybrid index key based on the topological spatial coordinates of the spatiotemporal unit, the acquisition timestamp, and the modality identifier. The construction method is represented as follows:

[0027] ;

[0028] in, This indicates the composite index key written to the storage engine. This indicates the spatial coordinate encoding of the power grid equipment or line node to which the data belongs within the power grid topology. The timestamp represents the moment of data collection, with a precision down to the millisecond level. This represents a modality identifier used to distinguish whether the data originates from a time series, image, or text modality. This indicates a byte-level concatenation operation performed in a predetermined order of topological coordinates, timestamps, and modal identifiers.

[0029] Edge nodes use the unified semantic tensor as an index value and write it into the improved LSM-Tree storage engine according to the above hybrid index key. The storage engine sorts the memory and writes the data sequentially to disk according to the prefix order of the topological coordinates and timestamps in the hybrid index key, forming a contiguous memory layout. When the upper-layer scheduling or analysis system initiates a cross-modal joint query request, it directly locates the corresponding unified index interval according to the topological coordinate interval and time interval carried in the request, reads the unified semantic tensor in the interval at once, and completes the cross-modal data fusion and alignment query without relying on the application layer to perform cross-system table joining and splicing, thereby keeping the query latency at a low level.

[0030] In another implementation, in addition to using the improved LSM-Tree storage engine, the hybrid index structure can also use a composite index method that combines R-tree-based spatial indexes with skip list-based time indexes. Independent indexes are created for the topological spatial coordinate dimension and the timestamp dimension, and then joint query optimization is performed. This can also achieve fast retrieval of topological coordinates and timestamps, serving as an equivalent alternative to the aforementioned storage engine.

[0031] IV. Business Semantic Aware Cloud-Edge Collaborative Adaptive Cold / Hot Layering and Scheduling Mechanism:

[0032] This section of the technical solution addresses the write blocking issue caused by consistency protocols in low-concurrency, high-network environments. A service feature extractor is deployed on the edge node side to perform sliding window feature statistics on real-time electrical measurement data such as power and frequency within a preset data window. This extracts the active power change rate and frequency fluctuation amplitude, and, combined with the indication of whether alarm events exist within the window, calculates the service value score as follows:

[0033] ;

[0034] in, Indicates the first Business value score for each data window This indicates the rate of change of active power within the window. This indicates the frequency fluctuation range within the window. This is an indicator value that shows whether there are any alarm events in the window; its value is either 0 or 1. , , These are the weight coefficients for the corresponding features. In this embodiment, they are statistically calibrated using historical fault samples, and their values ​​are 0.4, 0.3, and 0.3, respectively. Not less than the preset scoring threshold In this embodiment, data is classified as hot data if it is hot and cold data otherwise. The value is 0.6;

[0035] For windows identified as cold data, edge nodes use a lossy compression algorithm to compress the data and asynchronously synchronize it to the cloud storage layer using a weak consistency protocol, allowing for a certain range of latency and packet loss. For windows identified as hot data, edge nodes immediately trigger a multi-replica strong consistency synchronization process, inserting the window data into a high-priority synchronization queue in a preemptive manner and pausing the currently ongoing cold data synchronization task. The cloud storage layer confirms and writes the arriving cold and hot data to disk according to their respective consistency levels. In the case of hot data synchronization failure, retransmission processing and alarm prompts are triggered to ensure that critical data is not lost.

[0036] Edge nodes also continuously monitor the available network bandwidth of the current transmission link and the actual bandwidth used in the current period, and dynamically update the target transmission rate of the synchronization queue for the next period in the following manner:

[0037] ;

[0038] in, This indicates the target transmission rate of the synchronization queue in the next cycle. This indicates the transmission rate for the current period. This indicates the currently detected available network bandwidth. This indicates the actual bandwidth used in the current period. This represents the adjustment step size coefficient, used to control the smoothness of the rate adjustment and avoid rate oscillation. In this embodiment, the value is 0.5. Through the above-mentioned bandwidth adaptive adjustment mechanism, edge nodes can avoid queue accumulation and write blocking caused by excessively high write rates in weak network environments.

[0039] V. In-store computing integrated streaming data immunization and lightweight cleaning mechanism:

[0040] This section of the technical solution addresses the issue of spoofed data injection attacks on distributed nodes. A lightweight feature encoder is deployed at the data inflow interface of the edge nodes to encode the features of each newly arrived raw data entry, obtaining a corresponding feature vector. Based on the topological connection relationship of the power grid node to which the data belongs, the Kirchhoff current residual of that node is calculated as a preliminary verification index for the consistency of physical laws:

[0041] ;

[0042] in, Indicates the first Kirchhoff current residuals at each grid node Represents nodes The set of directly connected adjacent nodes. Indicates adjacent nodes Flow to Node The measured current value, Represents a node Its own current measurement value, when When the preset tolerance threshold is exceeded, it is considered that the data of the node has an abnormal physical law. In this embodiment, the preset tolerance threshold is five percent of the rated current.

[0043] The feature encoder uses historical normal data samples as positive samples and artificially constructed anomalous samples that violate Kirchhoff's laws as negative samples. It is trained using a self-supervised contrastive learning approach, which ensures that the similarity of normal data samples in the feature space is significantly higher than that of anomalous samples. After training, for each piece of data to be detected, its reputation score is calculated as follows:

[0044] ;

[0045] in, Indicates the first The credibility score for each piece of collected data ranges from 0 to 1, with higher values ​​indicating more reliable data. Indicates the first The feature vector obtained by mapping data through a feature encoder Indicates and Positive sample feature vectors that are consistent in physical laws are constructed from historical normal data of the same topological node. Indicates the first The negative sample feature vectors are the feature vectors corresponding to artificially constructed outlier samples that violate Kirchhoff's laws. Indicates the number of negative samples. Represents the cosine similarity function. This represents the temperature coefficient, used to control the smoothness of the probability distribution; in this embodiment, it is set to 0.1.

[0046] Edge nodes combine reputation scores and Kirchhoff current residuals to jointly determine the original collected data: when the reputation score is lower than a preset reputation threshold, the original collected data is determined to be suspicious data; when the Kirchhoff current residual exceeds a preset tolerance threshold, the original collected data is determined to be suspicious data. In this embodiment, the preset reputation threshold is 0.5. For data determined to be suspicious, historical reliable data at adjacent times within the sliding time window is called, and an interpolation repair algorithm is used to repair the abnormal data, generating corrected data. The corrected data or the original data that has passed the verification is written to the disk queue and uniformly stored by the aforementioned multimodal spatiotemporal tensor construction and hybrid index process. At the same time, the determination results and correction records are written to the audit log module for subsequent business system traceability. In addition, edge nodes regularly supplement newly confirmed false injection samples to the negative sample library, incrementally retrain the feature encoder, continuously improve the reputation scoring mechanism's ability to identify new attack patterns, and form a closed loop for continuous evolution of detection capabilities.

[0047] In another implementation, in addition to using a feature encoder based on contrastive learning, the reputation scoring network can also use a graph neural network to model the topological connection relationship of the power grid. By directly learning the physical consistency relationship between electrical measurements of adjacent nodes through the message passing mechanism between nodes, the node anomaly score output by the graph neural network can replace the aforementioned reputation score. This can also achieve real-time identification and joint judgment of false data injection attacks, serving as an equivalent alternative to the aforementioned reputation scoring mechanism.

[0048] VI. Specific Application Examples:

[0049] Taking an edge node of a new energy power station as an example, the implementation process of this invention is illustrated. The edge node receives in real-time voltage, current, and frequency data uploaded by the phasor measurement device within the station, image data uploaded by the inspection camera device, and alarm text data uploaded by the equipment log collection system. The aforementioned raw data first enters the reputation scoring and streaming cleaning module. After feature encoding, the reputation score and Kirchhoff current residual are calculated respectively. When the current measurement data at a certain moment experiences abnormal fluctuations due to communication interference, and its Kirchhoff current residual exceeds a preset tolerance threshold, this data is determined to be suspicious. The edge node then calls historical reliable data from adjacent moments before and after the measurement point to interpolate and repair it, generating corrected data and recording it in the audit log. The corrected data, along with other verified data, enters the spatiotemporal semantic tensor construction module. It performs feature mapping and splicing with image slices and log slices within the same time window to generate a unified semantic tensor and constructs a hybrid index key according to topological coordinates, timestamps, and modal identifiers. The data is then written to the storage engine. Simultaneously, the hot / cold tiering and adaptive synchronization scheduling module calculates the service value score based on the active power change rate, frequency fluctuation amplitude, and the presence of alarm events within the time window. If the score is not less than the preset score threshold, it is determined to be hot data and inserted into the high-priority synchronization queue in a preemptive manner, immediately triggering a strong consistency synchronization process to send it to the cloud storage layer. If the score is less than the preset score threshold, it is determined to be cold data, and after lossy compression, it is added to the weak consistency asynchronous synchronization queue. Edge nodes continuously monitor the available and used bandwidth of the current link and dynamically adjust the sending rate of the synchronization queue to avoid write blocking caused by limited network bandwidth in the area where the site is located. After receiving the synchronization data, the cloud storage layer maintains a globally unified index. The scheduling decision system can directly hit the unified index interval based on the topology coordinate interval and time interval, and read the phasor measurement data, image data, and log data corresponding to the fault time of the site at one time to complete cross-modal joint analysis and provide support for power grid scheduling decisions.

[0050] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A distributed storage and processing method for energy big data, characterized in that, Includes the following steps: Raw data uploaded from the acquisition terminal is acquired at the edge node of the power grid site. This raw data includes phasor measurement data, inspection image and video data, and equipment log text data. Streaming immune cleaning is performed on the raw data at the edge node to obtain verified data. Streaming immune cleaning includes feature encoding of the raw data to obtain feature vectors, and calculating a reputation score based on these feature vectors. The reputation score is calculated based on a feature encoder trained using a contrastive learning method. The feature encoder is trained using historical normal data samples as positive samples and artificially constructed anomalous samples violating Kirchhoff's laws as negative samples, ensuring that the similarity of normal data samples in the feature space is higher than that of anomalous samples. This is combined with Kirchhoff current residuals calculated based on the power grid topology to jointly determine the raw data. When data is determined to be suspicious, historical reliable data from adjacent moments within a sliding time window is used to further verify its validity. The suspicious data is interpolated and repaired to obtain the verified data. At the edge node, the verified data is mapped and concatenated into a unified semantic tensor. A hybrid index key is constructed based on the topological coordinates, collection timestamp, and modality identifier of the verified data. The unified semantic tensor is used as the index value and written to the storage engine according to the hybrid index key. At the edge node, a business value score is calculated based on the verified data. Based on the business value score, the verified data is divided into hot data and cold data. The cold data is lossily compressed and asynchronously synchronized to the cloud storage layer using a weak consistency protocol. The hot data is inserted into a high-priority synchronization queue in a preemptive manner and synchronized to the cloud storage layer using a strong consistency protocol. The available network bandwidth and the bandwidth used in the current period are continuously monitored. Based on the difference between the available network bandwidth and the bandwidth used, the target transmission rate of the synchronization queue in the next period is adjusted according to a preset step size coefficient.

2. The distributed storage and processing method for energy big data according to claim 1, characterized in that: The step of mapping and concatenating the verified data into a unified semantic tensor includes: performing time-aligned slicing on the phasor measurement data, inspection image and video data, and equipment log text data in the verified data with a preset time window as the granularity, to obtain phasor measurement data slices, image slices, and log slices within the same spatiotemporal unit; calling the temporal feature mapping function, image feature mapping function, and text feature mapping function respectively to map the phasor measurement data slices, image slices, and log slices to a feature subspace of a unified dimension to obtain the corresponding feature representations; and concatenating the feature representations along the feature dimension to obtain the unified semantic tensor corresponding to the spatiotemporal unit.

3. The distributed storage and processing method for energy big data according to claim 1, characterized in that: The hybrid index key is composed of the topological space coordinates, the acquisition timestamp, and the modality identifier concatenated byte-level in a predetermined order. The storage engine is an improved LSM-Tree storage engine. The storage engine sorts memory and writes data sequentially to disk based on the prefix order of the topological space coordinates and the acquisition timestamp in the hybrid index key to form a contiguous memory layout. When a cross-modal joint query request is received, the corresponding unified index interval is directly located based on the topological coordinate interval and time interval carried in the query request, and the unified semantic tensor within that interval is read to complete the cross-modal data fusion and alignment query.

4. The distributed storage and processing method for energy big data according to claim 1, characterized in that: The business value score is calculated by weighting the active power change rate, frequency fluctuation amplitude, and alarm event indication quantity of the verified data within a preset data window. The alarm event indication quantity is used to indicate whether there is an alarm event within the data window. When the business value score is not less than a preset score threshold, the verified data corresponding to the data window is determined to be the hot data. When the business value score is less than the preset score threshold, it is determined to be the cold data.

5. The distributed storage and processing method for energy big data according to claim 1, characterized in that: The step of inserting the hot data into the high-priority synchronization queue in a preemptive manner further includes: pausing the currently ongoing cold data synchronization task, and the cloud storage layer triggering retransmission processing and alarm prompts in the event of failure to synchronize the arriving hot data.

6. The distributed storage and processing method for energy big data according to claim 1, characterized in that: The reputation score is calculated based on a feature encoder trained using a contrastive learning method. The feature encoder is trained using historical normal data samples as positive samples and artificially constructed abnormal samples that violate Kirchhoff's laws as negative samples, so that the similarity of normal data samples in the feature space is higher than that of abnormal samples in the feature space.

7. The distributed storage and processing method for energy big data according to claim 6, characterized in that, Also includes: Write the judgment result and the correction record of interpolation repair into the audit log, add the confirmed suspicious data to the negative sample library, and perform incremental retraining on the feature encoder.

8. A distributed storage and processing system for energy big data, characterized in that, The system comprises an edge processing layer and a cloud storage layer. The edge processing layer is deployed at edge nodes in the power grid field and includes a reputation scoring and streaming cleansing module, a spatiotemporal semantic tensor construction module, a hybrid index writing module, and a hot / cold stratification and adaptive synchronization scheduling module. The reputation scoring and streaming cleansing module performs feature encoding on the raw collected data uploaded by the acquisition terminal to obtain feature vectors, calculates a reputation score based on the feature vectors, and performs joint judgment on the raw collected data in conjunction with Kirchhoff current residuals calculated based on the power grid topology connection relationship. For data judged as suspicious, historical reliable data at adjacent times within the sliding time window is used for interpolation and repair to obtain verified data. The spatiotemporal semantic tensor construction module maps and concatenates the verified data into a unified semantic tensor. The hybrid index writing module is used to determine the verification data based on the topology of the verified data. A hybrid index key is constructed using spatial coordinates, acquisition timestamps, and modal identifiers. The unified semantic tensor is then written into the storage engine as the index value according to the hybrid index key. The hot / cold tiering and adaptive synchronization scheduling module is used to calculate the service value score based on the verified data and continuously monitor the available network bandwidth of the transmission link and the bandwidth used in the current period. The target transmission rate of the synchronization queue in the next period is adjusted according to a preset step size coefficient. Based on the service value score, the verified data is divided into hot data and cold data. The cold data is lossily compressed and then asynchronously synchronized to the cloud storage layer using a weak consistency protocol. The hot data is inserted into a high-priority synchronization queue in a preemptive manner and then synchronized to the cloud storage layer using a strong consistency protocol. The cloud storage layer is used to receive the unified semantic tensor data synchronized by the edge processing layer and maintain a globally unified index.