An ai power load blockchain prediction method

Abstract

This invention relates to the field of data processing technology in artificial intelligence and blockchain, and discloses an AI-based blockchain prediction method for power load. The method includes: acquiring and cleaning multi-source data based on hierarchical collection points and time anchors; calculating hashes and packaging them to generate on-chain evidence pointers; then verifying the source signature, time anchor, unit, and numerical boundaries based on the evidence pointers to provide a gating pass certificate; constructing time-series features and holiday tags under gating constraints, establishing electrical and geographical topologies and aggregating messages, conducting dual-channel fusion inference to obtain the segment load prediction results and reliable intervals; finally, generating a prediction summary and model version registration, writing them on-chain, establishing a traceable index, and combining meter reading receipts to form parameter update annotations. This scheme constructs an evidence-first and gating-into-model link, integrates time-series and topological contexts, and forms a prediction-registration-on-chain-callback closed loop, improving data credibility and prediction stability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology in artificial intelligence and blockchain, and in particular to an AI-based blockchain method for predicting electricity load. Background Technology

[0002] In the field of data processing technology in artificial intelligence and blockchain, existing solutions for power load forecasting typically build centralized models around multi-source historical and real-time data. These models generate input records for forecasting through conventional preprocessing and unified clock mapping, and complete result registration and business traceability on the business system side. However, these solutions suffer from limitations such as uneven reliability of raw data batches, inconsistent time anchor configurations across domains, and difficulty in unifying unit definitions and numerical boundaries. Existing methods often rely on offline archiving or platform-side verification lists for post-event comparison, lacking a source signature verification and time anchor alignment link in conjunction with hierarchical collection point lists. Furthermore, they fail to establish an on-chain registration path for the evidentiary association between cleaned data batches and original data batches. In scenarios requiring access control at the forecast entry point, this can easily lead to missing gating strategies and fragmented exception handling, making it difficult to achieve stable results and reliable intervals for segmented load forecasting. For the joint processing of on-chain evidence pointer structure, gating strategy judgment and dual-channel fusion reasoning, existing technologies generally have fragmented links in the following aspects: synchronous reference of data batch package after cleaning and on-chain evidence pointer structure; pre-judgment of source signature verification, unit specification verification and numerical boundary verification; collaborative fusion of time series feature construction and holiday tag binding and graph topology construction and message aggregation; closed-loop governance of prediction summary generation and model version registration and on-chain writing and traceable index generation and deviation calculation and parameter update proposal preparation. It is difficult to form a consistent process of collection-alignment-judgment-fusion-recording-callback in the application scenario of power load forecasting, resulting in insufficient source traceability and process governance of section load forecasting results and reliable intervals. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an AI-based blockchain-based method for predicting electricity load, comprising:

[0004] The raw data batch package is obtained from the hierarchical collection point list and time anchor configuration through collection session initialization and data retrieval processing. The format standardization, including the unified conversion of units and dimensions, the abnormal fragment labeling including out-of-bounds fragment labeling, the packet-level and segment-level hash calculation and evidence packaging processing are performed to generate the cleaned data batch package and the on-chain evidence pointer structure.

[0005] Based on the on-chain evidence pointer structure and the original data batch package, source signature verification and time anchor alignment, unit specification verification and numerical boundary verification, and gate policy determination operations hosted by the on-chain smart contract are performed to obtain the gate pass certificate structure.

[0006] The gating system obtains the voucher structure and cleaned data batch package, performs time-series feature construction and holiday label binding, graph topology construction and message aggregation, and dual-channel fusion inference processing through fusion gate reference neighborhood consistency label and rhythm label overview to generate segment load prediction results and reliable intervals;

[0007] The system extracts correlation information from the segment load forecast results and the reliable interval, performs forecast summary generation and model version registration, on-chain writing and traceable index generation, deviation calculation based on real meter reading receipts and parameter update proposal preparation and processing including parameter whitelist, and constructs on-chain forecast record and parameter update annotation structure.

[0008] Furthermore, the tiered data collection point list and time anchor configuration include:

[0009] The hierarchical data collection point list includes the unique identifier of the data collection point, the geographical location information of the data collection point, the electrical topology location, the equipment category and maintenance status, the sampling channel and sampling frequency, the data access interface and authentication method, the data collection agent type and deployment medium; the time anchor configuration includes the subnet alignment parameters based on the precision time protocol, the timing channel and timing alarm threshold based on the global satellite navigation system, the sampling window and sliding step size, the time zone and daylight saving time policy, the cross-domain boundary buffer duration, the playback window and fault tolerance delay.

[0010] Furthermore, the on-chain evidence pointer structure and the original data batch package include:

[0011] The on-chain evidence pointer structure includes evidence pointer, block location, transaction index, batch number, session number, time window, hierarchical source range, evidence digest verification value, and trusted storage domain access guide; the raw data batch package includes batch number, collection range, number of fragments, fragment order index, hierarchical source details, time window, timing status digest, and health digest.

[0012] Furthermore, the gating system includes the following components in the voucher structure and the cleaned data batch package:

[0013] The gated pass credentials structure includes pass tags, batch numbers, policy version fingerprints, approval chain summaries, disposal log indexes, available fragment list references, available source set references, time window mapping table references, and evidence references; the cleaned data batch package includes batch numbers, source batch references, normalized versions, labeled versions, segment counts, abnormal fragment counts, and abnormal category distributions.

[0014] Furthermore, the process of performing source signature verification and time anchor alignment, unit specification verification and numerical boundary checks, and gating strategy determination operations hosted by on-chain smart contracts also includes:

[0015] The consistency verification record is used as the input for judgment. The batch number, time window, unit mapping overview and boundary verification overview are read, and the compliance factor calculation process defined in the gating strategy is called to complete the batch compliance judgment. The judgment process completes the compliance comparison by enumerating and counting the compliance factor triggering situations. When there is a conflict set or review set in the batch, the batch is put into a suspended state according to the exception list in the strategy and the manual review interface, and the reason for suspension and pending actions are recorded. The gating strategy provides a fine-grained judgment path at the fragment level. When the compliance at the batch level meets the strategy requirements but there are exceptions at the fragment level, a fragment exception list is generated on the contract side, and the gating judgment is required to be submitted again after the exception list is handled. The handling actions include fragment removal, fragment demotion, fragment revision and fragment review. The execution record of the handling actions and the executing entity are written to the handling log, and the handling receipt is recorded by the contract side.

[0016] Furthermore, the gating strategy process also includes:

[0017] The gating policy supports a multi-party approval process. When cross-domain or cross-layer sources are involved, the system initiates approval requests in the order of the policy, and the data owner, the operation and maintenance party and the predictive side responsible party complete the approval in sequence.

[0018] If the approval chain is not completed at any stage, the system will place the batch into the approval timeout queue and prioritize processing the queue in subsequent judgments.

[0019] Gating is generated by marking after the batch completes compliance determination, exception list handling, and multi-party approval. The system generates a unique pass mark, gating strategy version fingerprint, approval chain summary, handling log index, and evidence reference for the batch. The contract side returns a pass receipt and registers the receipt location.

[0020] Furthermore, the process of constructing time-series features and binding holiday labels, constructing graph topology and aggregating messages, and performing dual-channel fusion inference processing by referencing neighborhood consistency tags and rhythmic label overviews through fusion gates also includes:

[0021] Within the inference scheduler, a dual-channel processing pipeline is established for each batch. The first channel of the pipeline is the temporal branch, which receives the sample sequence and label of the temporal feature set and performs intra-segment encoding, cross-segment aggregation, and temporal context injection.

[0022] Intra-segment encoding maps the analyzed segment to a fixed-length representation and applies a masking strategy to the gap-derived placeholders. The masking strategy skips the corresponding positions and records the masking bitmap during intra-segment encoding.

[0023] The second channel of the pipeline is the graph topology branch, which receives the node features, edge features and neighborhood summaries of the graph feature set, and performs multi-hop message propagation, edge weight mapping and hierarchical restoration.

[0024] Multi-hop message propagation propagates node status and neighborhood summaries along adjacency relationships on the graph in units of time points. The number of propagation steps is determined by the configuration parameters indicated by the disposal log index.

[0025] Furthermore, the dual-channel fusion inference process also includes:

[0026] A fusion gate is constructed between the two channels. The fusion gate refers to the neighborhood consistency marker and rhythm label overview to align the time-series representation and graph representation at the same time point and generate a fusion weight. The fusion weight and the fusion strategy indicated by the disposal log index jointly determine the channel contribution at different source levels, different rhythm segments, and different neighborhood consistency states.

[0027] During the fusion process, the system outputs a fused time location representation for each acquisition point, and attaches a masking bitmap, neighborhood consistency reference, and rhythm injection index to the representation for subsequent reliable interval estimation and sample tracing.

[0028] Furthermore, the process of generating the section load forecast results and the confidence interval also includes:

[0029] For each time point, segment-level load forecasts are generated, and a confidence interval is generated outside the forecast values;

[0030] The generation of credible intervals relies on two types of information: one is the quality and uncertainty indication provided by the masking bitmap and the anomaly annotation summary, which is used to characterize the completeness and stability of the input at that time point; the other is the external consistency indication provided by the neighborhood consistency marker and the rhythm label overview, which is used to characterize the cooperative state of the same or neighboring domains at that time point.

[0031] Given upper and lower bounds under the constraints of two types of information, and record the source summary of the credible interval.

[0032] Furthermore, the process of extracting correlation information from the segment load forecast results and the reliable interval, generating forecast summaries and registering model versions, writing on-chain and generating traceable indexes, and calculating deviations based on real meter reading receipts and compiling parameter update proposals including parameter whitelists also includes:

[0033] Write the session number, batch number, model version fingerprint, channel configuration summary, fusion strategy identifier and trace reference into the output record to form the segment load prediction result and the confidence interval;

[0034] The segment load prediction results and trusted intervals are referenced to complete on-chain writing and traceable index generation, and are also referenced for deviation calculation and parameter update proposal preparation, constructing an on-chain prediction record and parameter update annotation structure.

[0035] The key innovations of this invention include:

[0036] (1) Construct an evidence-first input access link: Obtain the original data batch package from the hierarchical collection point list and time anchor configuration, and form an on-chain evidence pointer structure by format standardization, abnormal fragment labeling, hash calculation and evidence packaging. Based on the on-chain evidence pointer structure and the original data batch package, perform source signature verification and time anchor alignment, unit standard verification and numerical boundary verification and gate control strategy determination to obtain the gate pass certificate structure, which is used to limit the data batch package after cleaning to enter the subsequent processing.

[0037] (2) Complete time-series-graph dual-channel fusion reasoning within the same pipeline: obtain the gated pass certificate structure and cleaned data batch package, construct time-series features and bind holiday tags, construct graph topology and aggregate messages, and perform dual-channel fusion reasoning processing on this basis to generate section load prediction results and reliable intervals.

[0038] (3) Form a traceable closed loop of prediction-registration-on-chain-callback: Extract related information from the segment load prediction results and the reliable interval, generate prediction summaries and register model versions, write on-chain and generate traceable indexes, and carry out deviation calculation and parameter update proposal preparation based on on-chain prediction records and real meter reading receipts, and construct on-chain prediction record and parameter update annotation structure.

[0039] The following are its main beneficial effects:

[0040] (1) It acts on the input link before entering the model. Through source signature verification and time anchor alignment, unit specification verification and numerical boundary verification and gating strategy determination, it establishes a one-to-one verifiable relationship between the original data batch package and the on-chain evidence pointer structure. It completes the admission control at the prediction entry point. In view of the uneven source credibility and inconsistent clock mapping in the background technology, it avoids the transmission of unverified data to subsequent links.

[0041] (2) It is applied to the model processing stage. It is constructed by time series features and bound to holiday tags, constructed by graph topology and aggregated by message. In the dual-channel fusion inference processing, it utilizes both time series context and topological context to output the segment load prediction results containing the credible interval. It addresses the single path modeling and context missing problem in the background technology, and reduces the risk of cross-regional linkage and structural fluctuations being ignored.

[0042] (3) It acts on the link of result governance and continuous optimization. By generating prediction summaries and registering model versions, writing on the chain and generating traceable indexes, the segment load prediction results and the trusted interval are recorded on the chain. Then, combined with real meter reading receipts, deviation calculation and parameter update proposal are carried out, and the parameter update annotation structure is output. In response to the link fragmentation and unverifiable update problem in the background technology, a continuous process from registration to callback is formed, which is convenient for review and iteration under the same scenario boundary. Attached Figure Description

[0043] Figure 1 A flowchart illustrating an AI-based blockchain-based power load prediction method provided in this application embodiment;

[0044] Figure 2 A flowchart illustrating step S100 provided in an embodiment of this application;

[0045] Figure 3 A flowchart illustrating step S200 provided in an embodiment of this application;

[0046] Figure 4 A flowchart illustrating step S300 provided in an embodiment of this application;

[0047] Figure 5 This is a flowchart illustrating step S400 provided in an embodiment of this application. Detailed Implementation

[0048] Example 1: Refer to Figure 1 This is a flowchart illustrating an AI-based electricity load blockchain prediction method provided in an embodiment of the present invention. The process may include at least steps S100-S400:

[0049] S100: Obtain the original data batch package from the hierarchical collection point list and time anchor configuration through collection session initialization and data retrieval processing; perform format standardization including unified conversion of units and dimensions, abnormal fragment labeling including out-of-bounds fragment labeling, and packet-level and segment-level hash calculation and evidence packaging processing to generate the cleaned data batch package and on-chain evidence pointer structure.

[0050] S200, based on the on-chain evidence pointer structure and the original data batch package, performs source signature verification and time anchor alignment, unit specification verification and numerical boundary check, and gate policy determination operation hosted by the on-chain smart contract to obtain the gate pass certificate structure;

[0051] S300: Obtain the gating through the voucher structure and cleaned data batch package, perform time series feature construction and holiday label binding, graph topology construction and message aggregation, and generate segment load prediction results and reliable intervals through dual-channel fusion inference processing of fusion gate reference neighborhood consistency label and rhythm label overview.

[0052] S400: Extract correlation information from the segment load prediction results and the reliable interval, perform prediction summary generation and model version registration, on-chain writing and traceable index generation, deviation calculation based on real meter reading receipts and parameter update proposal preparation and processing including parameter whitelist, and construct on-chain prediction record and parameter update annotation structure.

[0053] like Figure 2 As shown, Figure 2 This is a flowchart illustrating step S100 provided in an embodiment of this application. Step S100 includes at least steps S110-S130:

[0054] S110. Obtain the list of hierarchical collection points and time anchor configuration, perform collection session initialization and data retrieval processing, and obtain the raw data batch package;

[0055] The input sources for this section are the hierarchical data collection point list and time anchor configuration generated in the previous configuration phase. The hierarchical data collection point list is a collection list of substation busbar outgoing terminals, transformer secondary sides, and key large user metering points. It includes fields such as unique identifier of the data collection point, geographical location information of the data collection point, electrical topology location, equipment category and maintenance status, sampling channel and sampling frequency, data access interface and authentication method, data collection agent type and deployment medium, etc. The time anchor configuration is a unified description of the time reference, including subnet alignment parameters based on the Precision Time Protocol, timing channel and timing alarm threshold based on the Global Navigation Satellite System, sampling window and sliding step size, time zone and daylight saving time strategy, cross-domain boundary buffer duration, playback window and fault tolerance delay, etc. Specifically, the aforementioned hierarchical collection point list and time anchor configuration are used as input. The collection session initialization process generates session numbers, requests session-level access tokens, performs collection agent health self-checks and link connectivity detection, and performs channel handshakes, permission verifications, and rate distribution for collection agents at different levels. The collection agent has a built-in Trusted Execution Environment and Trusted Platform Module, which are used to complete key desealing, startup measurement, and startup evidence reporting locally to ensure that the integrity measurement of the collection agent during startup and operation can be recorded. After the session initialization is completed, the sampling plan is allocated to each collection agent according to the hierarchy, triggering data pull processing, including configuring the pull mode (active push mode or passive pull mode), data buffer size and overflow policy, breakpoint resumption and retry count, collection window start and end time, channel priority, and reserved channels. When a collection agent returns a health abnormality or a link packet loss mark, the collection session will automatically fall back to the backup channel or backup sampling plan and record the recovery action. Furthermore, for data channels at different levels, the raw data is first aggregated into micro-batch fragments by sampling timestamp at the edge side, and then aligned by a unified time anchor at the session control side to form a segmented and ordered sequence of raw fragments. Fragments that arrive later than the configured threshold are marked as delayed fragments, duplicate fragments are marked as duplicate fragments, and missing fragments are recorded as missing fragments. All fragments retain the original timestamp and the proxy-side timestamp when they arrive for subsequent time consistency verification.Understandably, during data retrieval, if a cross-domain clock offset exceeds a threshold, the acquisition session will append a time offset tag to the segment in that domain and write the offset tag, along with the timing channel status on the agent side, into the session record. When multi-site parallel acquisition is triggered, the session control side assigns a higher buffer priority to the segments at the bus outgoing end based on electrical topology priority, thereby avoiding time imbalance caused by downstream transformer segment arriving before upstream segment. Finally, all segments accessed and minimally verified by the session control side are merged into a raw data batch packet for one session in the control side's buffer, including fields such as batch number, acquisition range, number of segments, segment order index, hierarchical source details, time window, timing status summary, and health summary. The above output field is named "Raw Data Batch Packet," and its subsequent input position is the raw data batch packet of S120. It also serves as a parallel input source for the raw data batch packet of S210 for source signature verification and time anchor alignment processing.

[0056] S120. Extract fields and labels from the original data batch package, perform format standardization and abnormal fragment labeling, and generate a cleaned data batch package;

[0057] The input source for this section is the raw data batch package output by S110. The range of fields and tags extracted includes device identifiers and channel identifiers, sampling timestamps and sampling intervals, units and dimensions, measured values ​​and quality bits, device operating status tags and maintenance status tags, environmental and holiday tags, metering loop topology tags and electricity price period tags, etc. The tags are jointly generated by the status reporting on the acquisition agent side and the business configuration on the session control side. Specifically, the format standardization first performs a unified conversion of units and dimensions, aligning the current, voltage, active power, reactive power, apparent power, power factor, phase angle, and other measurement dimensions reported by different devices to the system's predetermined standard dimensions, and writing the dimensions before conversion, the conversion factor, and the dimensions after conversion into the conversion record; the timestamps are standardized, converting the time recorded by the edge side using the local clock to the absolute time under a unified time anchor, retaining the original time and generating a time mapping relationship; the device and channel identifiers are dictionary-mapped, mapping the supplier-defined identifiers to the system's unified identifiers, and recording the before-and-after mapping comparison table and the mapping version number. Furthermore, the abnormal segment labeling process performs quality bit resolution and value range detection on all segments within the original data batch. Values ​​exceeding the physically reasonable range are labeled as out-of-bounds segments, sequences exhibiting non-physical jumps within a short period are labeled as jump segments, segments that are continuous and constant but do not conform to the equipment characteristic range are labeled as frozen segments, and intervals with missing samples or abnormally widened sample time intervals are labeled as missing sample segments. During the labeling process, for each type of abnormal segment, the abnormality category, start and end times, involved channels, involved equipment, abnormality intensity, and labeling source are written, and a labeling evidence summary is generated for subsequent evidence packaging. For time arrangement imbalances between cross-domain associated transformer substations and buses, transformer substation segments are labeled as arrangement imbalance segments by referring to the event point list on the bus side. When encountering duplicate segments caused by sampling retransmission, the original segments are retained and the duplicate segments are labeled as redundant segments, while the source of redundancy and the number of retransmissions are recorded. Understandably, format normalization also unifies the data record structure, transcribing raw records in different transmission formats into structured records within the system, constructing data segment headers, data segment bodies, and data segment footers. The segment header includes source, channel, time window, and quality bit summary; the segment body includes measurement fields and extended tags; and the segment footer includes segment checksums and record counts. After completing the above processing, all records that have passed normalization and annotation are reassembled into cleaned data batch packets, and batch number, source batch reference, normalization version, annotation version, segment count, abnormal fragment count, and abnormal category distribution are written at the packet level. The above output field is named "cleaned data batch packet," and its subsequent input position is the cleaned data batch packet of S130. It also serves as the input source for the cleaned data batch packet of S310 for time-series feature construction and holiday tag binding processing. In addition, the original data batch packet is still retained as a read-only reference on the session control side for S130 to reference when performing dual-stream evidence packaging.

[0058] S130. Perform hash calculation and evidence packaging on the original data batch package and the cleaned data batch package to generate an on-chain evidence pointer structure.

[0059] This section takes as input the raw data batch packets output from S110 and the cleaned data batch packets output from S120. Hash calculation is performed on the two types of batch packets at the packet level and the segment level, respectively. The packet-level summary describes the overall characteristics of the batch content at the packet scale, while the segment-level summary describes the fine-grained characteristics of a single record or segment at the segment scale. To ensure the consistency of the arrangement of order-sensitive data in the summary, the records are stably sorted before hash calculation. The sorting key consists of timestamp, channel identifier, and hierarchical source, and the original order index and the stable sorting index are retained in the sorted records. After the calculation is completed, the raw data batch packets are given a set of raw packet-level summaries and a set of raw segment-level summaries, and the cleaned data batch packets are given a set of cleaned packet-level summaries and a set of cleaned segment-level summaries. The batch number, sorting rule fingerprint, normalized version, and labeled version are associated in the evidence summary table. Specifically, the evidence packaging process involves placing the aforementioned four types of digests and key context records into an evidence container. The key context records include the session number, session time window, session control side timing status digest and link health digest, collection agent startup metric and local root certificate chain digest, mapping dictionary and mapping version, and anomaly annotation category dictionary and annotation source. To ensure that the evidence container does not expose plaintext content when referenced on-chain, the evidence packaging process only outputs the evidence pointer and necessary index information. The evidence content is stored in segments and sealed in a trusted storage domain, located on edge devices or central trusted storage nodes that have undergone trusted execution environment measurement. After evidence packaging is complete, the on-chain registration interface is called to submit the evidence pointer, batch number, session number, time window, submitter identifier, and evidence digest verification value. The on-chain registration interface returns the successfully registered transaction index and block position, and generates an evidence registration receipt. Under boundary conditions, if the on-chain registration interface returns a failure or the delay exceeds a threshold, the evidence packaging process will enter a retry queue, recording the reason for failure and the next attempt time in the retry queue, while retaining a read-only access entry for the evidence container until registration is complete.Furthermore, the evidence packaging process also writes two reference relationships into the evidence index table: the reference relationship between the original data batch package and the original package-level digest and the original segment-level digest set, and the reference relationship between the cleaned data batch package and the cleaned package-level digest and the cleaned segment-level digest set. The index table records version evolution and rollback pointers, allowing for tracing back to any historical evidence when normalization or annotation rules are upgraded. After the above processing, the evidence packaging process generates an on-chain evidence pointer structure, which includes evidence pointers, transaction indexes and block positions, batch numbers and session numbers, time windows and hierarchical source ranges, evidence digest verification values, and trusted storage domain access guidelines. The output field is named "On-Chain Evidence Pointer Structure," and its subsequent input position is the on-chain evidence pointer structure of S210, used for source signature verification and time anchor alignment, while also providing a query entry point for evidence association in S220 and S230. Regarding cross-main step connections, the on-chain evidence pointer structure also serves as an evidence reference when writing prediction records in the S400 stage, forming a continuous index from input evidence to prediction records. In summary, the technical effects of this step are as follows: By performing digest calculations and evidence packaging on the original data batch package and the cleaned data batch package respectively, and registering referable evidence pointers on the chain, a verifiable credential chain and version evolution of the input data are established, providing a directly referable evidence entry point for subsequent consistency verification and gating determination.

[0060] like Figure 3 As shown, Figure 3 This is a flowchart illustrating step S200 provided in an embodiment of this application. Step S200 includes at least steps S210-S230:

[0061] S210. Obtain the on-chain evidence pointer structure and the original data batch packet, perform source signature verification and time anchor alignment processing, and obtain the source verification record;

[0062] The input sources for this section are the on-chain evidence pointer structure output from the preceding step S130 and the raw data batch package output from S110. The on-chain evidence pointer structure includes the evidence pointer, block location, transaction index, batch number, session number, time window, hierarchical source range, evidence digest verification value, and trusted storage domain access guide. The raw data batch package includes the batch number, collection range, number of fragments, fragment order index, hierarchical source details, time window, timing status digest, and health digest. Specifically, the on-chain evidence pointer structure is used as the evidence entry point. An on-chain read command is invoked to obtain the corresponding registration receipt and registration metadata. The collection agent certificate digest, registration sequence, submitter identifier, and registration verification value are read from the registration metadata. To complete the source signature verification, the system verifies the signature submitted by the collection agent, covering the batch number, fragment order index, and registration verification value simultaneously, and cross-validates the evidence digest verification value. To ensure the reliability of the signature chain, this section introduces operational constraints of the Public Key Infrastructure (PKI) and the Hardware Security Module (HSM) when it first appears. The former provides certificate issuance, revocation, and trust chain verification, while the latter provides key escrow and signature operator isolation. When a certificate is revoked, expired, or has an incomplete issuance path, the system marks the source as an abnormal certificate source in the source signature verification process and adds a certificate abnormality label and certificate verification snapshot to the source side of the original data batch packet. Furthermore, to complete the time anchor alignment process, this section references the timing records of the Precision Time Protocol (PMT) and the Global Navigation Satellite System (GNSS) from the previous configuration. It maps the edge device time and session control side time in the original data batch packets to absolute time under a unified time anchor, and writes the mapping relationship and offset into the time alignment snapshot. When the cross-domain offset exceeds the configured threshold, the system generates a time drift label and a source domain alarm fragment, and attaches them to the hierarchical source details. Understandably, source signature verification and time anchor alignment require concurrent execution. This section constructs parallel verification channels at the session layer and performs a combined verification of the sub-results output by each channel. Only when the signature chain is complete, the evidence digest is consistent, and the time anchor mapping is completed, is the corresponding source included in the available source set. When a source is in a semi-available state (e.g., the signature is valid but the time drift label exists), the system includes the source in the pending verification set and records the reason for the pending verification.After the processing link is completed, the system summarizes the available source set, the set to be reviewed, and the unavailable source set. Simultaneously, it organizes certificate verification snapshots, time alignment snapshots, and anomaly tags to form a source verification record. The source verification record includes fields such as set partitioning result, set count, source list, evidence citation, alignment offset distribution, review reason summary, and review suggestion. The above output field is named "Source Verification Record," and its subsequent input position is the starting input item for the consistency verification process of S220, i.e., the source verification record of S220. Simultaneously, in cross-main-step scenarios, it writes source snapshots that provide indirect evidence citations to the prediction records of the S400 stage.

[0063] S220. Extract the time window and boundary fields from the source verification record, perform unit specification verification and numerical boundary check, and generate consistency verification record;

[0064] The input source for this section is the source verification record generated by S210. The time window field comes from the time-aligned snapshot and session time window of the source verification record. The boundary field covers unit and dimension definitions, measurement upper and lower limits, equipment status boundaries, measurement accuracy level, anomaly label dictionary, and boundary rule version. Specifically, each segment in the available source set is expanded one by one, and the segment's time window, unit and dimension, measured value, quality position, and equipment status label are read. The unit specification is verified according to the system's unified unit mapping dictionary. The verification process records the unit value before mapping, the value after mapping, and the mapping dictionary version. When a unit mapping is missing or a mapping conflict is found, the system generates a unit anomaly item and places the segment into the unit review queue. After completing the unit specification verification, the system calls the numerical boundary verification sub-process. This sub-process sets differentiated boundaries for different layered sources: the busbar outgoing end of the substation uses the busbar-level boundary, the secondary side of the transformer in the distribution area uses the distribution area-level boundary, and the metering point of key large users uses the user-level boundary. The verification process sequentially executes value range verification, jump verification, freeze verification, and arrangement stability verification. Value range verification is used to determine whether the segment value falls within the rated range and extended safety range of the equipment. Jump verification is used to identify non-physical sudden changes in a short period of time. Freeze verification is used to identify behaviors that remain constant for a long time and are not compatible with the equipment operating conditions. Arrangement stability verification is used to check the orderliness and gaps of the segments on the aligned time axis. Furthermore, for records marked as delayed segments, duplicate segments, missing segments, or imbalanced segments by S120, this section merges the abnormal label dictionary during verification, cross-compares the cleaned-side labels with the local verification conclusions, and outputs three results: consistent label, appended label, or conflict label. When a conflict label appears, the system adds the segment to the conflict review set and uses it as a separate factor in the subsequent gating strategy. Understandably, time window processing is present throughout this section. While verifying units and boundaries, the system compares the time window in which the segment is located with the holiday label. If there is a contradiction between the holiday label and the time window, a holiday conflict label is added to the segment and the source of the label is recorded. When performing numerical boundary verification on segments in adjacent transformer areas or on the same feeder, the system can call the neighborhood reference service to pull a statistical overview of the nearest segments and write this overview into the verification evidence summary. Through the above processing, the system summarizes the unit specification verification list, numerical boundary check list, anomaly review list and neighborhood reference summary at the batch scale. At the same time, it aggregates the fragment layer results into batch layer statistics to form a consistency verification record. The consistency verification record includes fields such as batch number, time window, boundary rule version, unit mapping overview, boundary check overview, anomaly label overview, review set and conflict set, evidence citation and reference summary.The above output field is named Consistency Verification Record. Its subsequent input position is the gating strategy judgment input item of S230, that is, the consistency verification record of S230. At the same time, this record can provide the basis for selecting compliant fragments for the construction of time series features of S300 in cross-main step scenarios, and prevent non-compliant fragments from entering the feature construction channel.

[0065] S230. Determine the gating strategy and generate the pass mark for the consistency verification record, and generate the gating pass certificate structure;

[0066] The input source for this section is the consistency verification record output by S220. The gating policy is managed by an on-chain smart contract. The gating policy specifies the policy version, judgment threshold, compliance factor, exception list, manual review interface, and multi-party approval process. The multi-party approval process records the approval role, approval order, and approval time limit. Specifically, the system uses the consistency verification record as the judgment input. First, it reads the batch number, time window, unit mapping overview, and boundary check overview, and then calls the compliance factor calculation process defined in the gating policy to complete the batch compliance judgment. The judgment process does not introduce new quantitative formulas, but completes the compliance comparison by enumerating and counting the compliance factor triggering situations. When a batch has a conflict set or review set, the system puts the batch into a suspended state according to the exception list and manual review interface in the policy, and records the suspension reason and pending actions. Furthermore, the gating strategy provides fine-grained judgment paths at the fragment level. When batch-level compliance meets the strategy requirements but there are exceptions at the fragment level, the system generates a fragment exception list on the contract side and requires a resubmission of the gating judgment after the exception list is processed. Processing actions include fragment removal, fragment deweighting, fragment revision, and fragment review. The execution records and executing entities of these actions are written to the processing log, and the contract side records the processing receipt. Understandably, the gating strategy supports multi-party approval processes. When cross-domain or cross-layer sources are involved, the system initiates approval requests according to the strategy order, with the data owner, operations and maintenance party, and prediction-side responsible party completing approvals in sequence. If the approval chain is not completed within the allotted time, the system places the batch in an approval timeout queue and prioritizes processing this queue in subsequent judgments. Gating is performed after a batch completes compliance assessment, exception list processing, and multi-party approval. The system generates a unique pass token, gating strategy version fingerprint, approval chain summary, processing log index, and evidence reference for that batch. The contract returns a pass receipt and registers its location. Simultaneously with pass token generation, the system prepares a list of available fragments, a set of available sources, and a time window mapping table for the subsequent prediction phase. Evidence references are written for these three types of products at the batch level. These evidence references are linked to the aforementioned on-chain evidence pointer structure and consistency verification records, ensuring traceability of input evidence and verification basis during the prediction phase. Through this process, the system generates a gating pass credential structure. This structure includes fields such as pass token, batch number, strategy version fingerprint, approval chain summary, processing log index, available fragment list reference, available source set reference, time window mapping table reference, and evidence reference. The above output field is named Gated Pass Certificate Structure, and its subsequent input position is the Gated Pass Certificate Structure of S310, which is used to drive the construction of time series features and the binding of holiday labels. In terms of cross-main step connection, the Gated Pass Certificate Structure also provides a gated basis reference for the prediction record writing in the S400 stage, thereby associating input evidence, verification conclusions and gated receipts at the record layer.In summary, the technical effects of this step are as follows: a credible foundation for the source is established through source signature verification and time anchor alignment; a conclusion on the consistency between fragments and batches is formed through unit specification verification and numerical boundary checks; and an access certificate for the prediction pipeline is generated through the gating strategy of contract custody and multi-party approval. This ensures that the inputs entering the prediction stage undergo pre-verification in terms of source, time, and content consistency and form an interlocking reference relationship with on-chain evidence.

[0067] In one embodiment, based on the on-chain evidence pointer structure and the original data batch package, source signature verification and time anchor alignment, unit specification verification and numerical boundary check, and gating policy determination operations are performed to obtain the gating pass certificate structure. The input sources of this main step are the on-chain evidence pointer structure output by the preceding step S130 and the original data batch package output by S110. The on-chain evidence pointer structure includes evidence pointer, block position, transaction index, batch number, session number, time window, hierarchical source range, evidence digest verification value, and trusted storage domain access guide. The original data batch package includes batch number, collection range, number of fragments, fragment order index, hierarchical source details, time window, timing status digest, and health digest. Specifically, the on-chain evidence pointer structure is used as the evidence entry point. On-chain read commands are invoked to obtain the corresponding registration receipt and registration metadata. The system then reads the acquisition agent certificate digest, registration sequence, submitter identifier, and registration verification value from the registration metadata. To complete source signature verification, the system verifies the signature submitted by the acquisition agent, covering three elements: batch number, fragment sequence index, and registration verification value, and cross-validates them with the evidence digest verification value. To ensure the reliability of the signature chain, this section introduces operational constraints from public key infrastructure and hardware security modules. The former provides certificate issuance, revocation, and trust chain verification, while the latter provides key escrow and signature operator isolation. When a certificate is revoked, expired, or has an incomplete issuance path, the system marks the source as an abnormal certificate source during the source signature verification process and adds a certificate abnormality tag and certificate verification snapshot to the source side of the original data batch packet. Furthermore, to complete the time anchor alignment process, this section references the timing records of the precise time protocol and the Global Navigation Satellite System from the previous configuration. It maps the edge device time and session control side time in the original data batch packets to absolute time under a unified time anchor, and writes the mapping relationship and offset into the time alignment snapshot. When the cross-domain offset exceeds the configured threshold, the system generates a time drift label and a source domain alarm fragment, and attaches them to the hierarchical source details. Understandably, source signature verification and time anchor alignment require concurrent execution. This section constructs parallel verification channels at the session layer and performs a combined verification of the sub-results output by each channel. Only when the signature chain is complete, the evidence digest is consistent, and the time anchor mapping is completed, is the corresponding source included in the available source set. When a source is in a semi-available state, the system adds the source to the pending verification set and records the reason for the pending verification. After the processing link is completed, the system summarizes the available source set, the set to be reviewed, and the unavailable source set. At the same time, it organizes certificate verification snapshots, time alignment snapshots, and anomaly tags to form a source verification record. The source verification record includes fields such as set partitioning results, set count, source list, evidence citation, alignment offset distribution, review reason summary, and review suggestions.The above output field is named Source Verification Record. Its subsequent input position is the starting input item of the consistency verification process of S220, that is, the Source Verification Record of S220. At the same time, in the cross-main step scenario, it provides evidence to indirectly cite the source snapshot for the prediction record of the S400 stage.

[0068] Formula ① is used to calculate the verification score for source signature verification. This score is based on the consistency measure between the signature and the data elements, and uses a cosine similarity function to quantify the verification strength. Formula ① is defined as follows:

[0069]

[0070] in, This represents the verification score, with a value ranging from [0,1]. A larger value indicates a more valid signature. The k-th data element vector is composed of the batch number, fragment order index and registration verification value after encoding, and originates from the corresponding field of the original data batch package; The k-th signature vector originates from the signature data submitted by the acquisition agent; n is the vector dimension, determined by the encoding method of the data elements. For vector index; The batch number, fragment sequence index, and registration check value of the original data batch package are extracted and encoded into a vector index, denoted as . ; The signature data corresponding to the evidence digest verification value of the on-chain evidence pointer structure is extracted and encoded into a vector index, denoted as . Formula ① As a signature verification strength indicator in the source verification record, it is used for subsequent compliance calculations.

[0071] Data Source → Metrics → Variable Mapping: Extract batch number, fragment sequence index, and registration check value from the original data batch package and encode them as... The signature data is extracted from the on-chain evidence pointer structure and encoded as... Together, they form the input and output of formula ①. Boundary verification weights for S220.

[0072] Formula ② is used to calculate the quality score of time anchor alignment. This score is based on the absolute value of the time offset and uses an exponential decay function to evaluate the alignment quality. Formula ② is defined as follows:

[0073]

[0074] in, This represents the alignment quality score, with a value range of (0,1]. A larger value indicates better time alignment. This represents the time offset, calculated using the following formula: ,in The edge device time is derived from the timestamp field in the timing status summary of the original data batch packet; To unify the time anchor time, it is derived from the precise time protocol reference configured by the time anchor; is the attenuation coefficient, a positive constant, determined by system configuration. Formula ② As a time alignment quality indicator in the source verification record, it is used for subsequent boundary verification.

[0075] Data Source → Metrics → Variable Mapping: Extracting Edge Device Time from Raw Data Batch Packages Extract a unified time anchor from the time anchor configuration. ,calculate Substituting into formula ②, we get Formula ② directly uses the formula from formula ①. As a weighting factor, but formula ① and formula ② are not directly dependent, therefore formula ② is calculated independently; formula ②'s The time-series weights used as input to S220 for unit specification verification.

[0076] Further, in S220, the time window and boundary fields are extracted from the source verification record to perform unit specification verification and numerical boundary checks, generating a consistency verification record. The input source for this section is the source verification record produced by S210. The time window field comes from the time-aligned snapshot and session time window of the source verification record, and the boundary field covers unit and dimension definitions, measurement upper and lower limits, equipment status boundaries, measurement accuracy levels, anomaly label dictionary, and boundary rule versions. Specifically, each segment in the available source set is expanded one by one, and the segment's time window, unit and dimension, measured value, mass position, and equipment status label are read. Unit specification verification is completed according to the system's unified unit mapping dictionary. The verification process records the unit value before mapping, the value after mapping, and the mapping dictionary version. When a unit mapping is missing or a mapping conflict is found, the system generates a unit anomaly item and places the segment into the unit review queue. After completing the unit specification verification, the system calls the numerical boundary verification sub-process. This sub-process sets differentiated boundaries for different layered sources: the busbar outgoing end of the substation uses the busbar-level boundary, the secondary side of the transformer in the distribution area uses the distribution area-level boundary, and the metering point of key large users uses the user-level boundary. The verification process sequentially executes value range verification, jump verification, freeze verification, and arrangement stability verification. Value range verification is used to determine whether the segment value falls within the rated range and extended safety range of the equipment. Jump verification is used to identify non-physical sudden changes in a short period of time. Freeze verification is used to identify behaviors that remain constant for a long time and are not compatible with the equipment operating conditions. Arrangement stability verification is used to check the orderliness and gaps of the segments on the aligned time axis. Furthermore, for records marked as delayed segments, duplicate segments, missing segments, or imbalanced segments by S120, this section merges the abnormal label dictionary during verification, cross-compares the cleaned-side labels with the local verification conclusions, and outputs three results: consistent label, appended label, or conflict label. When a conflict label appears, the system adds the segment to the conflict review set and uses it as a separate factor in the subsequent gating strategy. Understandably, time window processing is present throughout this section. While verifying units and boundaries, the system compares the time window in which the segment is located with the holiday label. If there is a contradiction between the holiday label and the time window, a holiday conflict label is added to the segment and the source of the label is recorded. When performing numerical boundary verification on segments in adjacent transformer areas or on the same feeder, the system can call the neighborhood reference service to pull a statistical overview of the nearest segments and write this overview into the verification evidence summary. Through the above processing, the system summarizes the unit specification verification list, numerical boundary check list, anomaly review list and neighborhood reference summary at the batch scale. At the same time, it aggregates the fragment layer results into batch layer statistics to form a consistency verification record. The consistency verification record includes fields such as batch number, time window, boundary rule version, unit mapping overview, boundary check overview, anomaly label overview, review set and conflict set, evidence citation and reference summary.The above output field is named Consistency Verification Record. Its subsequent input position is the gating strategy judgment input item of S230, that is, the consistency verification record of S230. At the same time, this record can provide the basis for selecting compliant fragments for the construction of time series features of S300 in cross-main step scenarios, and prevent non-compliant fragments from entering the feature construction channel.

[0077] Formula ③ is used for unit conversion calculations in unit specification verification, converting the original measured values ​​into standard dimensions. Formula ③ is defined as follows:

[0078]

[0079] in, This represents the converted measurement value; This indicates the original measurement value, which is derived from the fragment measurement value field of the source verification record; This represents the transformation factor, derived from the transformation parameters of the identity mapping dictionary. (Formula ③) As a unit specification indicator in the consistency verification record, it is used for subsequent numerical boundary verification.

[0080] Data Source → Metrics → Variable Mapping: Extracting raw measurement values ​​from source verification records Transformation factors are extracted from the unit mapping dictionary. ,calculate This is then used for inputting formula ④; formula ③ directly uses formula ②. As weights, but formulas ② and ③ are not directly dependent, therefore formula ③ is calculated independently; formula ③'s It is used as an input item for S230 for compliance calculation.

[0081] Formula ④ is used to calculate the violation score for numerical boundary checks. This score is based on the distance between the measured value and the boundary, and uses an exponential function to quantify the degree of violation. Formula ④ is defined as follows:

[0082]

[0083] in, This represents the boundary violation score, with a value range of (0,1]. A larger value indicates a lower degree of violation. This indicates the converted measurement value, which comes from the output of formula ③; This represents the boundary threshold, derived from the upper and lower limits of measurement in the boundary rule version; This represents the scaling parameter, a positive constant determined by system configuration. (Formula ④) As a compliance indicator of numerical boundaries in the consistency verification record, it is used for subsequent gating strategy determination.

[0084] Data source → Indicator → Variable mapping: Output of formula ③ As input, the boundary threshold is extracted from the boundary rule version. ,calculate Formula ④ directly uses formula ③. As input items, forming a closed loop of the formula chain; Formula ④ It is used as an input item for S230 for compliance calculation.

[0085] Furthermore, in S230, gating policy determination and pass mark generation are performed on the consistency verification records, generating a gating pass credential structure. The input source for this section is the consistency verification records output by S220. The gating policy is hosted by an on-chain smart contract and specifies the policy version, judgment threshold, compliance factor, exception list, manual review interface, and multi-party approval process. The multi-party approval process records the approval role, approval order, and approval time limit. Specifically, the system uses the consistency verification records as the judgment input. First, it reads the batch number, time window, unit mapping overview, and boundary check overview, and calls the compliance factor calculation process defined in the gating policy to complete the batch compliance determination. The determination process does not introduce new quantitative formulas but completes the compliance comparison by enumerating and counting the compliance factor triggering situations. When a batch has a conflict set or review set, the system puts the batch into a suspended state according to the exception list and manual review interface in the policy and records the suspension reason and pending actions. Furthermore, the gating strategy provides fine-grained judgment paths at the fragment level. When batch-level compliance meets the strategy requirements but there are exceptions at the fragment level, the system generates a fragment exception list on the contract side and requires a resubmission of the gating judgment after the exception list is processed. Processing actions include fragment removal, fragment deweighting, fragment revision, and fragment review. The execution records and executing entities of these actions are written to the processing log, and the contract side records the processing receipt. Understandably, the gating strategy supports multi-party approval processes. When cross-domain or cross-layer sources are involved, the system initiates approval requests according to the strategy order, with the data owner, operations and maintenance party, and prediction-side responsible party completing approvals in sequence. If the approval chain is not completed within the allotted time, the system places the batch in an approval timeout queue and prioritizes processing this queue in subsequent judgments. Gating is performed after a batch completes compliance assessment, exception list processing, and multi-party approval. The system generates a unique pass token, gating strategy version fingerprint, approval chain summary, processing log index, and evidence reference for that batch. The contract returns a pass receipt and registers its location. Simultaneously with pass token generation, the system prepares a list of available fragments, a set of available sources, and a time window mapping table for the subsequent prediction phase. Evidence references are written for these three types of products at the batch level. These evidence references are linked to the aforementioned on-chain evidence pointer structure and consistency verification records, ensuring traceability of input evidence and verification basis during the prediction phase. Through this process, the system generates a gating pass credential structure. This structure includes fields such as pass token, batch number, strategy version fingerprint, approval chain summary, processing log index, available fragment list reference, available source set reference, time window mapping table reference, and evidence reference.The above output field is named Gated Pass Certificate Structure, and its subsequent input position is the Gated Pass Certificate Structure of S310, which is used to drive the construction of time series features and the binding of holiday labels. In terms of cross-main step connection, the Gated Pass Certificate Structure also provides a gated basis reference for the prediction record writing in the S400 stage, thereby associating input evidence, verification conclusions and gated receipts at the record layer.

[0086] Formula ⑤ is used to calculate the overall compliance score, which is based on prior verification, alignment, unit and boundary scores. A weighted summation function is used to fuse multi-source indicators. Formula ⑤ is defined as follows:

[0087]

[0088] in, This represents the overall compliance score, with a value range of [0,1]. A larger value indicates a higher batch compliance. The verification score for formula ① is derived from the source verification record. The alignment quality score in Formula ② is derived from the source verification record; The boundary violation score for formula ④ is derived from the consistency check record; , , This represents the weighting coefficient, derived from the compliance factor of the gating strategy, and satisfies... Formula ⑤ As a compliance indicator in the gate control credential structure, it is used for the determination in Formula ⑥.

[0089] Data source → Indicator → Variable mapping: Output of formula ① Output of Formula ② Output of Formula ④ As input, the weight coefficients are extracted by the gating strategy. , , ,calculate Formula ⑤ directly uses the results of Formulas ①, ②, and ④ as inputs, forming a closed loop of formulas; Formula ⑤ As an input to formula ⑥.

[0090] Formula ⑥ is used to generate the pass flag for gating policy determination. This flag is based on a comparison between the comprehensive compliance score and a threshold, and uses a unit step function to decide the pass status. Formula ⑥ is defined as follows:

[0091]

[0092] in, This is a pass / fail flag, with a value of 0 or 1, where 1 indicates pass and 0 indicates fail. This represents the overall compliance score according to formula ⑤. This represents the decision threshold, derived from the decision threshold of the gating strategy. (Formula ⑥) As the core output of the gating through the credential structure, it is used to drive the subsequent prediction process.

[0093] Data source → Indicator → Variable mapping: Output of formula ⑤ As input, the decision threshold is extracted by the gating strategy. ,calculate Formula 6 directly uses formula 5. As input items, forming a closed loop of formula chain; Formula ⑥ The S310 is used as an input for admission control in the construction of temporal features.

[0094] This section summarizes the technical effects: By verifying the source signature and aligning the time anchor, a reliable source foundation is established; by verifying the unit specifications and checking the numerical boundaries, a conclusion on the consistency between fragments and batches is formed; and by using a contract-hosted gating strategy and multi-party approval, an access certificate for the prediction pipeline is generated, enabling the inputs entering the prediction stage to undergo pre-verification in terms of source, time, and content consistency, and to form an interlocking reference relationship with on-chain evidence.

[0095] like Figure 4 As shown, Figure 4 This is a flowchart illustrating step S300 provided in an embodiment of this application. Step S300 includes at least steps S310-S330:

[0096] S310. Obtain the time series feature set by constructing time series features and binding holiday tags through the voucher structure and cleaned data batch package;

[0097] The input sources for this section are the gating pass credential structure output from the preceding step S230 and the cleaned data batch package output from S120. The gating pass credential structure includes pass flag, batch number, policy version fingerprint, approval chain summary, disposal log index, available fragment list reference, available source set reference, time window mapping table reference, and evidence reference. The cleaned data batch package includes batch number, source batch reference, normalized version, labeled version, segment count, abnormal fragment count, and abnormal category distribution. Specifically, gating uses the available fragment list reference in the credential structure and the time window mapping table reference as admission constraints to filter and align records in the cleaned data batch package. Filtering retains only records covered by batch-level pass tags, and alignment rearranges the hierarchical source records to a unified timeline according to the time window mapping table, retaining the position index before and after rearrangement and cross-source splicing points in the aligned records. After alignment, the data is grouped according to session number and unique collection point identifier, triggering the time series feature construction process. The time series feature construction process executes time window slicing, gap filling marking, and fragment splicing marking within each group. Time window slices are used to construct fixed-length analysis segments, gap filling markings are used to prompt downstream models to execute masking strategies at corresponding time points, and fragment splicing markings are used to record the connection positions across batches or across sources to avoid cross-boundary points being mistaken for mutations. Furthermore, in response to the rhythmic differences from different hierarchical sources, the system derives multi-layered time context features for each analysis segment, including intraday time index, intraweek time index, intramonth time index, peak-valley-normal time zone labels, weekday and non-working day labels, holiday labels, and local special event labels. The system also selects the corresponding holiday caliber based on the geographical scope indicated by the approval chain summary. During the derivation process, records that switch across time zones or daylight saving time are converted using a time window mapping table. Records with conflicting holiday labels are marked with holiday conflict flags, and the cleaned and labeled version and gating strategy version fingerprints are written to the label source field. Understandably, to reflect load changes across multiple time scales, the system calculates multi-scale statistics and rate of change statistics for each analysis segment according to the configured window length. These statistics include moving averages, moving quantiles, moving volatility, and duration counts, while the rate of change statistics include differences between adjacent segments and cross-day isotopic differences. These statistics are recorded as derived fields. For locations with gap-filling markers, no corresponding statistics are generated; instead, gap-derived placeholders are written, and the gap type and gap span are recorded in the placeholder field. After the statistics and label derivation are completed, the system reassembles the analysis segments within the group into time-series sample sequences. Each sample sequence includes original record fields, normalized fields, anomaly annotation summaries, time context labels, holiday labels, and multi-scale statistical derived fields. When packaging the sample sequences, a unique identifier for the collection point, source level, session number, and batch number are written for subsequent cross-source aggregation and traceability.Through the above processing, the system summarizes the sample sequences and labels of all groups to form the output product of this section, the temporal feature set. The temporal feature set includes sample index, time position index, derived field index, label dictionary, and retrospective reference. The above output field is named temporal feature set, and its subsequent input position is the temporal feature set of S320. It is used to construct the relationship and adjacency information of the collection points and trigger graph topology construction and message aggregation. At the same time, in the cross-main step connection scenario, the temporal feature set will be used as the temporal branch input of the dual-channel fusion and inference execution of S330, and will be referenced in the S400 stage to mark the model version and sample index relationship when generating prediction record pointers.

[0098] S320. Extract the relationship and adjacency information of the collection points from the temporal feature set, perform graph topology construction and message aggregation, and generate a graph feature set;

[0099] The input source for this section is the time-series feature set output by S310. The time-series feature set provides unique identifiers for collection points, source levels, sample indexes, time location indexes, tag dictionaries, and traceability references. Specifically, the system first extracts the relationships between collection points from electrical and geographical perspectives. The electrical perspective forms electrical adjacencies based on the feeder connections between the substation busbar outgoing terminals, the secondary side of the transformer substation, and the metering points of key large users, the transformer platform relationships, and the metering circuit and phase information. The geographical perspective forms spatial adjacencies based on the geographical coordinates of the collection points, the substation boundary, and the service radius. Both perspectives introduce time location indexes for segmentation during relationship extraction, enabling independent adjacency snapshots to be formed for connection changes caused by equipment maintenance, load transfer, or temporary reconnection within the corresponding time periods. Furthermore, the system generates graph nodes and edges based on the relationships between the collection points. Each graph node corresponds to a unique identifier for the collection point and is attached with node attributes. Node attributes include source level, installed capacity, metering loop characteristics, historical alarm counts, and a rhythm tag overview. Graph edges correspond to electrical or geographical adjacencies and are attached with edge attributes. Edge attributes include relationship type, connection direction, impedance segment identifier or geographical distance segment identifier, common event count, and valid time interval. When the same pair of collection points has both electrical and geographical adjacencies, the system establishes two edges in the graph model and writes the edge type distinction, or records the composite relationship on a single edge with a relationship type field. When the edges are aggregated, a weight mapping is performed on the composite relationship, and the weight mapping table loads the mapping value under the configuration version indicated by the disposal log index. Understandably, after the graph topology is constructed, the system initiates a message aggregation process. Message aggregation uses time points as outer indices and gathers neighborhood summaries for each node from its first-order and second-order adjacencies at each time point. The neighborhood summaries are derived from sample sequences and derived fields of the time-series feature set. Aggregation strategies include summation aggregation, mean aggregation, maximum value aggregation, and stability label aggregation. During aggregation, fields carrying gap-derived placeholders are masked to ensure that the aggregation result does not introduce fictitious values ​​at gap points. Message aggregation also calculates neighborhood consistency labels to express whether the direction of the same feeder or adjacent transformer areas is consistent at the current time point, whether the peak and valley labels are consistent, and whether the anomaly labels are synchronized. Neighborhood consistency labels serve as a reference for the fusion gate during subsequent dual-channel fusion. Furthermore, to adapt to the differences in data density across different source levels, the system performs hierarchical merging and expansion of nodes: at the substation busbar outgoing end level, complete nodes are retained; at the transformer substation secondary side level, when multiple metering points within the same substation are statistically similarly distributed, hierarchical merging can be performed and a merging mapping table is recorded; at the key large user metering point level, when the same user has multiple circuit metering points and there is a significant phase difference in the business process, hierarchical expansion is performed and an expansion mapping table is recorded; all of the above merging and expansion are versioned at the time point dimension to ensure that the corresponding mapping version is referenced at different time periods.Through message aggregation and hierarchical processing, the system generates node-level and edge-level aggregation vectors and tag sets for each time point, which are then combined with node attributes, edge attributes, and trace references to form a graph feature snapshot. At the batch level, the system aggregates graph feature snapshots from all time points to form a graph feature set. The graph feature set includes node feature indexes, edge feature indexes, neighborhood summary indexes, merging and expansion mapping indexes, and time point indexes, and trace references and evidence references are written at the set level. The above output field is named Graph Feature Set, and its subsequent input position is the graph feature set of S330, which is used together with the time-series feature set output from S310 to enter the dual-channel fusion and inference execution. At the same time, in cross-main-step connection scenarios, the graph feature set can be selectively written into the prediction record summary in stage S400 as the topological basis for model version and structure configuration in the current batch.

[0100] S330. Perform dual-channel fusion and inference on the time series feature set and graph feature set to generate the section load prediction results and confidence intervals;

[0101] The input sources for this section are the time-series feature set output by S310 and the graph feature set output by S320. The time-series feature set provides sample sequences organized by collection point and time position, time context labels, and multi-scale statistical derived fields. The graph feature set provides node features, edge features, neighborhood summaries, and contraction-expansion mapping indexes organized by time position. Specifically, the system establishes a dual-channel processing pipeline for each batch within the inference scheduler. The first channel of the pipeline is the time-series branch, which receives the sample sequences and labels from the time-series feature set and performs intra-segment encoding, cross-segment aggregation, and time context injection. Intra-segment encoding maps the analysis segment to a fixed-length representation and applies a masking strategy to the gap-derived placeholders. The masking strategy skips the corresponding positions during intra-segment encoding and records the masking bitmap. Cross-segment aggregation is used to transfer trend and periodic information between adjacent analysis segments. Time context injection is used to inject intraday indexes, intraweek indexes, peak-valley-flat labels, and holiday labels into the segment-level representation and records the injection position index. The second channel of the pipeline is a graph topology branch. The second channel receives node features, edge features, and neighborhood summaries from the graph feature set, and performs multi-hop message propagation, edge weight mapping, and hierarchical restoration. Multi-hop message propagation propagates node status and neighborhood summaries along the adjacency relationship on the graph in units of time points. The number of propagation steps is determined by the configuration parameters indicated by the disposal log index. Edge weight mapping loads weights according to the relationship type and impedance or geographical distance segment identifiers, and the contributions of different relationship types are weighted and superimposed during propagation. Hierarchical restoration is used to restore the condensed nodes at the substation level to the original metering point granularity, or to aggregate multi-loop extended nodes at the user level into a user-view representation. Both restoration and aggregation retain the mapping index and trace reference. Furthermore, the system constructs a fusion gate between the two channels. The fusion gate refers to the neighborhood consistency marker generated by S320 and the rhythm label overview generated by S310 to align the time series representation and graph representation of the same time point and generate fusion weights. The fusion weights and the fusion strategy indicated by the disposal log index jointly determine the channel contribution at different source levels, different rhythm segments, and different neighborhood consistency states. During fusion execution, the system outputs a fused time point representation for each acquisition point, and attaches a masking bitmap, neighborhood consistency reference, and rhythm injection index to the representation for subsequent confidence interval estimation and sample tracing. Understandably, inference execution loops with a time window as the outer layer in each batch, generating segment-level load predictions for each time point, and generating confidence intervals outside the predictions. The generation of confidence intervals depends on two types of information: one is the quality and uncertainty indications provided by the masking bitmap and anomaly annotation summary, which characterize the completeness and stability of the input at that time point; the other is the external consistency indications provided by the neighborhood consistency marker and rhythm label overview, which characterize the cooperative state of the same or neighboring domains at that time point. Under the constraints of these two types of information, the system gives the upper and lower bounds and records the source summary of the confidence interval.Through the aforementioned fusion and inference, the system organizes the prediction outputs for each collection point and each time point at the batch level, forming segment load prediction results and confidence intervals. The session number, batch number, model version fingerprint, channel configuration summary, fusion strategy identifier, and traceability reference are written into the output record. The output field is named "Segment Load Prediction Results and Confidential Intervals," and its subsequent input position is the segment load prediction results and confidence intervals of S410, used to generate prediction summaries and model version registrations and obtain prediction record pointers. Simultaneously, in cross-main step connection scenarios, the segment load prediction results and confidence intervals are referenced by S420 to complete on-chain writing and traceable index generation, and by S430 for bias calculation and parameter update proposal preparation. In summary, the technical effect of this step is: by establishing dual-channel processing of temporal branches and graph topology branches on the layered aligned samples, and generating fusion outputs by the fusion gate under the constraints of neighborhood consistency and rhythm labeling, segment load prediction results and confidence intervals consistent with the input evidence, topological context, and temporal context can be generated at the same time point, providing a stable index and traceability basis for subsequent recording and callback steps.

[0102] In one embodiment, the gating pass credential structure and the cleaned data batch package are obtained, and time-series features are constructed and holiday tag binding is performed to obtain a time-series feature set. The input sources of this section are the gating pass credential structure output by the preceding step S230 and the cleaned data batch package output by S120. The gating pass credential structure includes a pass tag, batch number, policy version fingerprint, approval chain summary, disposal log index, available fragment list reference, available source set reference, time window mapping table reference, and evidence reference. The cleaned data batch package includes a batch number, source batch reference, normalized version, labeled version, segment count, abnormal fragment count, and abnormal category distribution. Specifically, gating uses the available fragment list reference in the credential structure and the time window mapping table reference as admission constraints to filter and align records in the cleaned data batch package. Filtering retains only records covered by batch-level pass tags, and alignment rearranges the hierarchical source records to a unified timeline according to the time window mapping table, retaining the position index before and after rearrangement and cross-source splicing points in the aligned records. After alignment, the data is grouped according to session number and unique collection point identifier, triggering the time series feature construction process. The time series feature construction process executes time window slicing, gap filling marking, and fragment splicing marking within each group. Time window slices are used to construct fixed-length analysis segments, gap filling markings are used to prompt downstream models to execute masking strategies at corresponding time points, and fragment splicing markings are used to record the connection positions across batches or across sources to avoid cross-boundary points being mistaken for mutations. Furthermore, in response to the rhythmic differences from different hierarchical sources, the system derives multi-layered time context features for each analysis segment, including intraday time index, intraweek time index, intramonth time index, peak-valley-normal time zone labels, weekday and non-working day labels, holiday labels, and local special event labels. The system also selects the corresponding holiday caliber based on the geographical scope indicated by the approval chain summary. During the derivation process, records that switch across time zones or daylight saving time are converted using a time window mapping table. Records with conflicting holiday labels are marked with holiday conflict flags, and the cleaned and labeled version and gating strategy version fingerprints are written to the label source field.Understandably, to reflect load changes across multiple time scales, the system calculates multi-scale statistics and rate of change statistics for each analysis segment according to the configured window length. These statistics include moving averages, moving quantiles, moving volatility, and duration counts, while the rate of change statistics include differences between adjacent segments and cross-day isotopic differences. These statistics are recorded as derived fields. For locations with gap-filling markers, no corresponding statistics are generated; instead, gap-derived placeholders are written, and the gap type and gap span are recorded in the placeholder field. After the statistics and label derivation are completed, the system reassembles the analysis segments within the group into time-series sample sequences. Each sample sequence includes original record fields, normalized fields, anomaly annotation summaries, time context labels, holiday labels, and rapid statistical derivation fields. When packaging the sample sequences, a unique identifier for the collection point, source level, session number, and batch number are written for subsequent cross-source aggregation and traceability. Through the above processing, the system summarizes the sample sequences and speeds of all groups to form the output product of this section, the temporal feature set. The temporal feature set includes sample index, time position index, derived field index, label dictionary, and trace reference. The above output field is named temporal feature set, and its subsequent input position is the temporal feature set of S320. It is used to construct the relationship and adjacency information of the collection points and trigger graph topology construction and message aggregation. At the same time, in the cross-main step connection scenario, the temporal feature set will be used as the temporal branch input of the dual-channel fusion and inference execution of S330, and will be referenced in the S400 stage to annotate the relationship between the model speed and sample index when generating the prediction record pointer.

[0103] Formula ⑦ is used to calculate the moving average statistic for the analysis period. This statistic is based on the average of measurements within the time window and uses a mean function to smooth short-term fluctuations. Formula ⑦ is defined as follows:

[0104]

[0105] in, It represents the moving average, with the same dimensions as the original measurement, and its range is determined by the data. Indicates the window length, derived from the configuration parameters of the time window mapping table, and its value range is positive integer; This indicates that the i-th measurement value is derived from the measurement index in the normalized field of the cleaned data batch package; This represents the index within the window, with a value range of integers from 1 to W. (Formula ⑦) As a derived field of the moving average in the time series feature set, it is used for subsequent graph aggregation.

[0106] Data Source → Metrics → Variable Mapping: Extracting Measurement Values ​​from Cleaned Data Batch Packages Extract the window length from the time window mapping table. ,calculate This is then used for inputting formula ⑧; formula ⑦ It is used as an input item for S320 for node attribute construction.

[0107] Formula ⑧ is used to calculate the influence factor of holiday labels. This factor is based on the matching degree between time context features and holiday labels, and uses a dot product function to quantify the influence strength. Formula ⑧ is defined as follows:

[0108]

[0109] in, This represents the impact factor of holidays, with a value range of [0,1]. This represents the j-th temporal context feature vector, which originates from the temporal context label field of the temporal feature set; This represents the j-th holiday tag vector, which originates from the holiday tag field of the time-series feature set; This represents the feature dimension, which is determined by the label encoding method, and its value range is positive integers. This represents the feature index, with a value range of integers from 1 to M. (Formula ⑧) As a derived field of holiday influence in the time series feature set, it is used for subsequent fusion.

[0110] Data Source → High-Speed ​​→ Variable Mapping: Extracting Time Context Labels from Time Series Feature Sets and holiday tags ,calculate Formula ⑧ directly uses formula ⑦. As a weighting reference, but formulas ⑦ and ⑧ are not directly dependent, therefore formula ⑧ is calculated independently; formula ⑧'s It is used as an input to S330 for calculating fusion gate weights.

[0111] Furthermore, in S320, the relationship and adjacency information of the collection points are extracted from the time-series feature set, and graph topology construction and message aggregation are performed to generate a graph feature set. The input source for this section is the time-series feature set output by S310. The time-series feature set provides unique identifiers for collection points, source levels, sample indexes, time position indexes, tag dictionaries, and traceability references. Specifically, the system first rapidly collects point relationships from both electrical and geographical perspectives. From the electrical perspective, electrical adjacency is formed based on the feeder connections between the substation busbar outgoing terminals, the secondary side of the transformer substation, and the metering points of key large users, as well as the relationships between transformer platforms, metering circuits, and phase information. From the geographical perspective, spatial adjacency is formed based on the geographical coordinates of the collection points, the boundaries of the transformer substation, and the service radius. Both approaches introduce time position indexes for segmentation during relationship extraction, enabling independent adjacency snapshots to be formed within the corresponding time periods when connection changes occur due to equipment maintenance, load transfer, or temporary reconnection. Furthermore, the system generates graph nodes and edges based on the relationships between the collection points. Each graph node corresponds to a unique identifier for the collection point and is attached with node attributes. Node attributes include source level, installed capacity, metering loop characteristics, historical alarm counts, and a rhythm tag overview. Graph edges correspond to electrical or geographical adjacencies and are attached with edge attributes. Edge attributes include relationship type, connection direction, impedance segment identifier or geographical distance segment identifier, common event count, and valid time interval. When the same pair of collection points has both electrical and geographical adjacencies, the system establishes two edges in the graph model and writes the edge type distinction, or records the composite relationship on a single edge with a relationship type field. When the edges are aggregated, a weight mapping is performed on the composite relationship, and the weight mapping table loads the mapping value under the configuration version indicated by the disposal log index. Understandably, after the graph topology is constructed, the system initiates a message aggregation process. Message aggregation uses time points as outer indices and gathers neighborhood summaries for each node from its first-order and second-order adjacencies at each time point. The neighborhood summaries are derived from the sample sequences and derived fields of the time-series feature set. Aggregation strategies include summation aggregation, mean aggregation, maximum value aggregation, and stability label aggregation. During aggregation, fields carrying gap-derived placeholders are masked to ensure that the aggregation result does not introduce fictitious values ​​at gap points. Message aggregation also calculates neighborhood consistency labels to express whether the direction of the same feeder or high-speed transformer area is consistent at the current time point, whether the peak and valley labels are consistent, and whether the anomaly labels are synchronized. Neighborhood consistency labels serve as a reference for the fusion gate during subsequent dual-channel fusion.Furthermore, to adapt to the differences in data density across different source levels, the system performs hierarchical merging and expansion of nodes: at the substation busbar outgoing end level, complete nodes are retained; at the transformer substation secondary side level, when multiple metering points within the same substation are statistically similarly distributed, hierarchical merging can be performed and a merging mapping table is recorded; at the key large user metering point level, when the same user has multiple circuit metering speeds and significant phase differences in the business process, hierarchical expansion is performed and a speed mapping table is recorded; all the above merging and expansion are versioned at the time point dimension to ensure that the corresponding mapping version is referenced at different time periods. Through message aggregation and hierarchical processing, the system forms node-level and edge-level aggregation vectors and tag sets for each time point, and merges them with node attributes, edge attributes, and trace references into a graph feature snapshot; at the batch level, the system summarizes the graph feature snapshots of all time points to form a graph feature set, which includes node feature index, edge feature index, neighborhood summary index, merging and expansion mapping index, and time point index, and writes trace references and evidence references at the set level. The above output field is named Graph Feature Set. Its subsequent input position is the Graph Feature Set of S330, which is used together with the temporal feature set output by S310 to enter the dual-channel fusion and inference execution. At the same time, in the cross-main step connection scenario, the Graph Feature Set can be selectively written into the prediction record summary in the S400 stage as the topological basis for the model version and structure configuration in the current batch.

[0112] Formula 9 is used to calculate the message aggregation value of a node's neighborhood. This value is based on the summation of features of first-order neighboring nodes, using a summation function to aggregate neighborhood information. Formula 9 is defined as follows:

[0113]

[0114] in, This represents the aggregated neighborhood summary value, with the same units as the node features, and its value range is determined by the data. This represents the set of first-order adjacent nodes of node v, derived from the connection relationships in the edge attributes of the graph topology; The feature value of the adjacent node k is derived from the derived field or node attribute of the time series feature set; This represents the index of the adjacent node, and its value is an integer. (Formula 9) It serves as a neighborhood summary index in the graph feature set, used for subsequent consistency calculations.

[0115] Data source → Metrics → Variable mapping: Extracting adjacency relationships from graph feature sets Feature values ​​are extracted from time series feature sets. ,calculate Formula 9 directly uses formula 8. As a weighting reference, but formula ⑧ and formula ⑨ are not directly dependent, therefore formula ⑨ is calculated independently; formula ⑨'s It is used as an input to formula ⑩ for consistency mark calculation.

[0116] Formula 10 is used to calculate the neighborhood consistency label, which is based on the variance of the neighborhood summary values. A variance function is used to quantify the degree of consistency. Formula 10 is defined as follows:

[0117]

[0118] in, This represents the neighborhood consistency flag, with a value range of [0, ∞). The smaller the value, the higher the consistency. This represents the number of adjacent nodes, derived from the node degree of the graph topology; Represents the mean of neighborhood features; From the definition of formula ⑨. Formula ⑩ It serves as a consistency marker in the graph feature set and is used for subsequent fusion gates.

[0119] Data source → Metrics → Variable mapping: Extracting adjacency relationships from graph feature sets and degree The output of formula ⑨ As input, calculate Formula 10 directly uses formula 9. As input items, forming a closed loop of formula chain; Formula 10 It is used as an input to S330 for fusion weight calculation.

[0120] Furthermore, in S330, the temporal feature set and graph feature set are fused and inferred in a dual-channel manner to generate the segment load prediction results and confidence intervals. The input sources for this section are the temporal feature set output by S310 and the graph feature set output by S320. The temporal feature set provides sample sequences organized by collection point and time position, time context labels, and multi-scale statistical derived fields. The graph feature set provides node features, edge features, neighborhood summaries, and contraction-expansion mapping index organized by time position. Specifically, the system establishes a dual-channel processing pipeline for each batch within the inference scheduler. The first channel of the pipeline is the time-series branch, which receives the sample sequence and labels of the time-series feature set and performs intra-segment encoding, cross-segment aggregation, and time context injection. Intra-segment encoding maps the analysis segment to a fixed-length representation and applies a masking strategy to the gap-derived placeholders. The masking strategy skips the corresponding positions and records the masking bitmap during intra-segment encoding. Cross-segment aggregation is used to transfer trend and periodic information between adjacent analysis segments. Time context injection is used to inject intraday indexes, intraweek indexes, peak-valley-flat labels, and holiday labels into the segment-level representation and records the injection position index. The second channel of the pipeline is a graph topology branch. The second channel receives node features, edge features, and neighborhood summaries from the graph feature set, and performs multi-hop message propagation, edge weight mapping, and hierarchical restoration. Multi-hop message propagation propagates node status and neighborhood summaries along the adjacency relationship on the graph in units of time points. The number of propagation steps is determined by the configuration parameters indicated by the disposal log index. Edge rapid mapping loads weights according to the relationship type and impedance or geographical distance segment identifiers, and the contributions of different relationship types are weighted and superimposed during propagation. Hierarchical restoration is used to restore the condensed nodes at the substation level to the original metering point granularity, or to aggregate multi-loop extended nodes at the user level into a user-view representation. Both restoration and aggregation retain the mapping index and trace reference. Furthermore, the system constructs a fusion gate between the two channels. The fusion gate refers to the neighborhood consistency marker generated by S320 and the rhythm label overview generated by S310 to align the time series representation and graph representation of the same time point and generate fusion weights. The fusion weights and the fusion strategy indicated by the disposal log index jointly determine the channel contribution at different source levels, different rhythm segments, and different neighborhood consistency states. During fusion execution, the system outputs a fused time point representation for each acquisition point, and attaches a masking bitmap, neighborhood consistency reference, and rhythm injection index to the representation for subsequent confidence interval estimation and sample tracing.Understandably, inference execution loops with a time window as the outer layer in each batch, generating segment-level load predictions for each time point, and generating confidence intervals outside the predictions. The generation of confidence intervals depends on two types of information: one is the quality and uncertainty indications provided by the masking bitmap and anomaly annotation summary, which characterize the completeness and stability of the input at that time point; the other is the external consistency indications provided by the neighborhood consistency marker and rhythm label overview, which characterize the cooperative state of the same or neighboring domains at that time point. Under the constraints of these two types of information, the system gives the upper and lower bounds and records the source summary of the confidence interval. Through the above fusion and reasoning, the system organizes the prediction output of each collection point and each time point at the batch level to form the segment load prediction result and the confidence interval. The session number, batch number, model version fingerprint, channel configuration summary, fusion strategy identifier and traceability reference are written into the output record. The above output field is named Segment Load Prediction Result and Confidential Interval. Its subsequent input position is the Segment Load Prediction Result and Confidential Interval of S410, which is used to generate the prediction summary and model version registration and obtain the ultra-fast record pointer. At the same time, in the cross-main step connection scenario, the segment load prediction result and the confidence interval are referenced by S420 to complete the on-chain writing and the generation of the traceable index, and are referenced by S430 for deviation calculation and parameter update proposal preparation.

[0121] Formula 11 is used to calculate the attention weights of the fusion gate. These weights are based on the similarity between the temporal representation and the graph representation, and are normalized using the softmax function. Formula 11 is defined as follows:

[0122]

[0123] in, This represents the fusion weight, with a value range of [0,1]. This represents the temperature parameter, which is derived from the configuration of the fusion strategy identifier. The similarity function, using dot product similarity, is defined as follows: ; The segment-level representation of the temporal branch is derived from the segment-level encoded output of the temporal feature set; The node representation of the graph topological branch is derived from the hierarchical reconstruction output of the graph feature set, and the dimension is the dimension of the encoded representation. This indicates an index, with a value range of integers. (Formula 11) As a fusion weight, it is used for weighted fusion in formula ⑫.

[0124] Data source → Indicator → Variable mapping: Extracting segment-level representations from time series feature sets Node representation is extracted from graph feature set. Temperature parameters are extracted from the fusion strategy identifier. ,calculate Formula 11 directly uses formula 10. As a reference for similarity adjustment, but formula ⑩ and formula ⑪ are not directly dependent, therefore formula ⑪ is calculated independently; formula ⑪'s It is used as an input to formula ⑫ for the calculation of predicted values.

[0125] Formula 12 is used to generate the section load forecast value. This value is based on the weighted fusion of the dual-channel representation and uses a linear weighting function to calculate the forecast output. Formula 12 is defined as follows:

[0126]

[0127] in, This represents the predicted load for a given area; the range of values ​​is determined by the data. This represents the fusion weights in formula 11; Indicates sequential branching; This represents the topological branching of the graph. Formula 12... As the core output of the section load forecast results, it is used to generate the confidence interval.

[0128] Data source → Indicator → Variable mapping: Output of formula 11 As weights, they are calculated using the representations of the temporal feature set and the graph feature set as input. Formula 12 directly uses formula 11. As input items, forming a closed loop of formula chain; Formula ⑫ It is used as an input to S410 for predictive summary generation.

[0129] This section summarizes the technical effects: By establishing a dual-channel processing mechanism of temporal branch and graph topology branch on the samples after hierarchical alignment, and generating a fusion output by a fusion gate under the constraints of neighborhood consistency and rhythm label, it can generate segment load prediction results and confidence intervals that are consistent with the input evidence, topological context and temporal context at the same time point, and provide a stable index and traceability basis for subsequent recording and callback steps.

[0130] like Figure 5 As shown, Figure 5 This is a flowchart illustrating step S400 provided in an embodiment of this application. Step S400 includes at least steps S410-S430:

[0131] S410. Obtain the section load prediction results and confidence intervals, perform prediction summary generation and model version registration processing, and obtain the prediction record pointer;

[0132] The input sources for this section are the segment load prediction results and confidence intervals output from the preceding step S330. The segment load prediction results are organized using the unique identifier of the collection point, time point, batch number, and session number as organizational units, recording the segment-level load value generated at each time point and its corresponding explanatory summary. The confidence interval records the upper and lower bound descriptions related to the corresponding time point, the basis for interval generation, and retrospective references. Specifically, using the segment load prediction results and confidence intervals as input, a prediction summary generation process constructs summary records associated with the current batch. This process first groups the outputs from different source levels according to the time window and the unique identifier of the collection point, and within each group, reads contextual information such as the rhythm tag overview, neighborhood consistency markers, and masking bitmaps to form record-oriented summary fragments. In these summary fragments, the system combines the time point, unique identifier of the collection point, source level, numerical source, confidence interval source, and fusion strategy identifier, retaining retrospective references during the combination. These retrospective references are used to record evidence pointers, gating receipts, and graph topology configuration summaries related to the prediction. Furthermore, to complete the model version registration process, this section loads the version list of the current batch on the prediction side. This version list, provided by the model training and deployment system, includes the main model version, time-series branch configuration version, graph topology branch configuration version, fusion gate policy version, and inference scheduler configuration version. Upon its first appearance, the version list introduces descriptions of two fundamental components: the Software Configuration Management system and the Version Control System. The former provides a method for solidifying deployment parameters and environment signatures, while the latter provides an indexing method for version evolution. Specifically, the system binds the version list to prediction summary fragments, forming batch-oriented version registration records, and generates a version fingerprint and registration timestamp for each type of version. When the prediction output involves differentiated strategies by region or source level, the system appends a region policy tag or level policy tag to the version registration record and writes it to the configuration version and disposition log index. Understandably, during the process of predictive summary generation and model version registration, the system simultaneously constructs a record entry for subsequent writing chains. At the batch level, the record entry includes the session number, batch number, and registration time window; at the time point level, it includes the unique identifier of the collection point, the time point, and the source summary of the reliable interval. When multiple sources output in parallel at the same time point, the record entry merges them according to the gating through the list of available fragments contained in the voucher, and marks the differences in the merging process in parallel.Through digest generation and version registration processing, the system aggregates entry information from all time points and collection points within a batch, constructs an index pointing structure and a traceability structure for on-chain writing, and encapsulates them as a predicted record pointer. The predicted record pointer includes a record entry index, version registration record reference, traceability reference, session number and batch number mapping relationship, time window mapping relationship, and a minimum data exposure strategy description. The minimum data exposure strategy description clarifies that this batch only carries pointers and digests when writing to the chain, without carrying plaintext data. The above output field is named "Predicted Record Pointer," and its subsequent input position is the predicted record pointer of S420. In cross-main step connection scenarios, it serves as the connection point between the evidence chain of stages S400 and S200, linking the segment load prediction result and the trusted interval to the on-chain evidence pointer structure and the gating through credential structure via pointers.

[0133] S420. Extract the associated session number and evidence association item from the prediction record pointer, perform on-chain writing and traceable index generation, and generate on-chain prediction record.

[0134] The input source for this section is the predicted record pointer output by S410. The predicted record pointer includes the record entry index, version registration record reference, retrospective reference, session number and batch number mapping relationship, time window mapping relationship, and data minimum exposure strategy description. Specifically, using the predicted record pointer as the write entry, the system first extracts the session number and batch number from the record entry index, and performs segmented mapping on the time points involved in this batch according to the time window mapping relationship. The purpose of segmented mapping is to align the time points with the block time series when they are registered on the chain, forming an on-chain time cursor. At the same time, the system parses the evidence association items from the retrospective reference. The evidence association items include the on-chain evidence pointer structure, the gating access credential structure, and the graph feature set reference digest. After parsing, an evidence reference table is formed and bound to the record entry in a one-to-one or one-to-many manner. Furthermore, the system calls the on-chain write interface, which is provided by the blockchain infrastructure. Upon its initial appearance, this interface introduces two fundamental building blocks: Distributed Ledger Technology (DLT) and Smart Contracts. The former provides an immutable record-keeping medium, while the latter provides a rule-based writing and approval channel. During the write process, the system follows a minimal data exposure strategy, registering only the prediction summary, version registration fingerprint, traceability reference, and evidence reference table. It does not write the original data or the complete prediction sequence. Simultaneously, the time window mapping relationship and the on-chain time cursor are written together into the registration field used for alignment. Understandably, before and after the write operation, the system generates two types of traceable indexes: a record index for quickly locating the prediction summary and version registration record for this batch on the chain; and an evidence index for tracing back from the prediction summary to the evidence pointer, gating receipt, and graph topology configuration summary. The record index includes the on-chain transaction location, block location, number of record entries, and entry verification value during generation. The evidence index includes the evidence citation table verification value, number of citations, and citation category distribution during generation. After generation, both types of indexes are submitted to the index registration function of the on-chain smart contract, which generates a registration receipt and returns the receipt position. Furthermore, to adapt to parallel writes across regions and source levels, the system establishes a write partition for each session number. All record entries within the write partition share the same time window and the same version of the registered record references. After the write is completed, the smart contract generates a partition summary receipt. When the write interface experiences delays or temporary unavailability, the system places the corresponding record entry into the write waiting queue and writes the pointer of the waiting queue to the off-chain buffer log. The buffer log records the retry strategy and the next attempt time, which is automatically replayed after on-chain writes resume.Through the aforementioned on-chain writing and traceable index generation, the system forms a two-layer structured prediction registration on-chain at both the batch level and the time-based level, resulting in on-chain prediction records. These on-chain prediction records include fields such as record index, evidence index, session number, batch number, time window, version registration fingerprint, prediction summary reference, and traceability reference. The output field is named "On-chain Prediction Record," and its subsequent input location is the on-chain prediction record of S430. In cross-main step connection scenarios, it forms a complete input-verification-prediction-registration chain with the source verification record and consistency verification record of stage S200, providing a verifiable source for subsequent deviation calculation and parameter update proposal preparation.

[0135] S430, Calculate the deviation between the on-chain prediction records and the actual meter reading receipts and compile parameter update proposals to generate parameter update annotation structures;

[0136] The input sources for this section are the on-chain prediction records output by the S420 and the actual meter reading receipts returned by the business settlement side. The on-chain prediction records provide record index, evidence index, session number, batch number, time window, version registration fingerprint, prediction summary reference, and traceability reference. The actual meter reading receipts, upon their first appearance, introduce the interface instructions between the business settlement system and the metering management system, including metering point identifier, settlement cycle, reading time, meter reading, quality digit, and verification status. Specifically, the system first locates the prediction summary related to the current batch using the record index of the on-chain prediction records, and then extracts the corresponding meter reading and quality digit from the actual meter reading receipts according to the session number and time window. When there is a missing meter reading, a verification status of "failed," or an abnormal quality digit indication, the time point is marked as a point that cannot be used for deviation calculation and recorded in the deviation exclusion list. Furthermore, within the available site range, the system pairs the predicted summary and the measurement reading according to the unique identifier of the collection point and the time point. During the pairing process, the source summary of the confidence interval and the neighborhood consistency marker are read and used as context labels for bias interpretation. For each successfully paired sample, the system generates a sample-level bias record, which includes the time point, the unique identifier of the collection point, the predicted summary reference, the measurement reading reference, the source summary of the confidence interval, the neighborhood consistency marker, and the quality bit. At the batch level, the system summarizes the sample-level bias records into a batch-level bias overview, which includes the number of available sites, the number of unavailable sites, the site distribution, the context label distribution, and the evidence reference. Understandably, parameter update proposals are developed based on a batch-level deviation overview. The system categorizes deviations according to model version registration fingerprints and fusion strategy identifiers. Deviations related to time-series branches are assigned to time-series features to construct parameter configuration suggestions, deviations related to graph topology branches are assigned to graph topology configuration suggestions, and deviations related to fusion gate strategies are assigned to fusion gate strategy suggestions. Each type of suggestion uses a defined parameter whitelist as the boundary for development. When the parameter whitelist appears for the first time, it introduces two types of control instructions: an Access Control List (ACC) and the Least Privilege principle. The ACC records the names and ranges of parameters that are allowed to be updated, while the Least Privilege principle records the minimum approval chain required for parameter updates. When deviations involve cross-regional or cross-source levels, the system adds an impact surface label and approval order description to the proposal.Furthermore, the system categorizes various suggestions into parameter update proposals. Each parameter update proposal includes a suggestion category, suggested parameters, suggested scope, impact label, approval order description, evidence citations, and retrospective citations. To ensure the proposals can be judged within the on-chain contract, the system calls the proposal registration interface of the on-chain smart contract. The registration interface returns a proposal registration receipt, and the receipt location and proposal summary are written into the annotation. After registration, the system performs amplitude limiting processing according to the amplitude limiting strategy of the on-chain contract. The amplitude limiting strategy is managed by the contract side and introduces two types of restrictions upon its first appearance: change amplitude limiting and frequency amplitude limiting. Change amplitude limiting is used to limit the magnitude of a single parameter change, while frequency amplitude limiting is used to limit the number of changes per unit time. After amplitude limiting processing, the system submits the proposal to a multi-party approval process. The multi-party approval process consists of the data owner, the operation and maintenance party, and the prediction side responsible party. The approval process records the approval roles, approval order, and approval time limits in the contract. If the approval chain times out at any stage, the system places the proposal in the timeout queue and prompts for supplementary materials or a reduction in scope. Through the aforementioned deviation calculation and parameter update proposal preparation, the system forms a parameter update annotation structure. The parameter update annotation structure includes the proposal registration receipt location, amplitude limiting processing result, approval chain summary, suggested parameter whitelist mapping, and evidence citation. It also records the reference location for the next round of construction process within the structure. The reference location points to the temporal feature construction parameter configuration of S310 and the graph topology construction parameter configuration of S320, and through the fusion gate strategy, it points to the fusion strategy configuration of S330. The above output field is named Parameter Update Annotation Structure. Its subsequent input locations are the temporal feature construction parameter configuration of S310 and the graph topology construction parameter configuration of S320, and the fusion strategy configuration reference is read by S330, thus forming a closed loop before the next batch enters the prediction. In summary, the technical effects of this step are as follows: By pairing on-chain prediction records with actual meter reading receipts point by point to generate sample-level and batch-level deviation records, and then compiling update proposals under the constraints of parameter whitelists and amplitude limiting strategies, and completing contract registration and multi-party approval, traceable deviation interpretations can be transformed into a controlled parameter update annotation structure, which will be referenced in the next round of feature construction and fusion configuration, forming a closed loop of registration-comparison-proposal-reference.

Claims

1. An AI-based blockchain-based method for predicting electricity load, characterized in that, include: Based on the hierarchical collection point list and time anchor configuration, the original data batch package is obtained through collection session initialization and data retrieval processing. The format standardization, including the unified conversion of units and dimensions, the abnormal fragment labeling including the cross-boundary fragment labeling, the packet-level and segment-level hash calculation and evidence packaging processing are performed to generate the cleaned data batch package and the on-chain evidence pointer structure. Based on the on-chain evidence pointer structure and the original data batch package, source signature verification and time anchor alignment, unit specification verification and numerical boundary checks are performed, along with gating strategy determination operations hosted by the on-chain smart contract, to obtain the gating pass certificate structure; specifically including: The consistency verification record is used as the input for judgment. The batch number, time window, unit mapping overview and boundary verification overview are read, and the compliance factor calculation process defined in the gating strategy is called to complete the batch compliance judgment. The judgment process completes the compliance comparison by enumerating and counting the compliance factor triggering situations. When there is a conflict set or review set in the batch, the batch is put into a suspended state according to the exception list in the strategy and the manual review interface, and the reason for suspension and pending actions are recorded. The gating strategy provides a fine-grained judgment path at the fragment level. When the compliance at the batch level meets the strategy requirements but there are exceptions at the fragment level, a fragment exception list is generated on the contract side, and the gating judgment is required to be submitted again after the exception list is processed. The processing actions include fragment removal, fragment demotion, fragment revision and fragment review. The execution record and execution subject of the processing actions are written to the processing log, and the processing receipt is recorded by the contract side. The gating system obtains the voucher structure and cleaned data batch package, performs time-series feature construction and holiday label binding, graph topology construction and message aggregation, and dual-channel fusion inference processing through fusion gate reference neighborhood consistency label and rhythm label overview to generate segment load prediction results and reliable intervals; The dual-channel fusion inference process includes: constructing a fusion gate between the two channels; the fusion gate references the neighborhood consistency marker and rhythm label overview; aligning the temporal representation and graph representation at the same time point and generating a fusion weight; the fusion weight and the fusion strategy indicated by the disposal log index jointly determine the channel contribution at different source levels, different rhythm segments, and different neighborhood consistency states. During the fusion process, the system outputs a fused time location representation for each acquisition point, and attaches a masking bitmap, neighborhood consistency reference and rhythm injection index to the representation for subsequent reliable interval estimation and sample tracing. The system extracts correlation information from the segment load forecast results and the reliable interval, performs forecast summary generation and model version registration, on-chain writing and traceable index generation, deviation calculation based on real meter reading receipts and parameter update proposal preparation and processing including parameter whitelist, and constructs on-chain forecast record and parameter update annotation structure.

2. The method according to claim 1, characterized in that, The tiered data collection point list and time anchor configuration include: The hierarchical data collection point list includes the unique identifier of the data collection point, the geographical location information of the data collection point, the electrical topology location, the equipment category and maintenance status, the sampling channel and sampling frequency, the data access interface and authentication method, the data collection agent type and deployment medium; the time anchor configuration includes the subnet alignment parameters based on the precision time protocol, the timing channel and timing alarm threshold based on the global satellite navigation system, the sampling window and sliding step size, the time zone and daylight saving time policy, the cross-domain boundary buffer duration, the playback window and fault tolerance delay.

3. The method according to claim 1, characterized in that, The on-chain evidence pointer structure and the original data batch package include: The on-chain evidence pointer structure includes evidence pointer, block location, transaction index, batch number, session number, time window, hierarchical source range, evidence digest verification value, and trusted storage domain access guide; the raw data batch package includes batch number, collection range, number of fragments, fragment order index, hierarchical source details, time window, timing status digest, and health digest.

4. The method according to claim 1, characterized in that, Gating via the credential structure and cleaned data batch package includes: The gated pass credentials structure includes pass tags, batch numbers, policy version fingerprints, approval chain summaries, disposal log indexes, available fragment list references, available source set references, time window mapping table references, and evidence references; the cleaned data batch package includes batch numbers, source batch references, normalized versions, labeled versions, segment counts, abnormal fragment counts, and abnormal category distributions.

5. The method according to claim 1, characterized in that, The gating strategy process also includes: The gating policy supports a multi-party approval process. When cross-domain or cross-layer sources are involved, the system initiates approval requests in the order of the policy, and the data owner, the operation and maintenance party and the predictive side responsible party complete the approval in sequence. If the approval chain is not completed at any stage, the system will place the batch into the approval timeout queue and prioritize processing the queue in subsequent judgments. Gating is generated by marking after the batch completes compliance determination, exception list handling, and multi-party approval. The system generates a unique pass mark, gating strategy version fingerprint, approval chain summary, handling log index, and evidence reference for the batch. The contract side returns a pass receipt and registers the receipt location.

6. The method according to claim 1, characterized in that, The process of constructing time-series features and binding holiday labels, constructing graph topology and aggregating messages, and performing dual-channel fusion inference processing by referencing neighborhood consistency tags and rhythmic label overviews through fusion gates also includes: Within the inference scheduler, a dual-channel processing pipeline is established for each batch. The first channel of the pipeline is the temporal branch, which receives the sample sequence and label of the temporal feature set and performs intra-segment encoding, cross-segment aggregation, and temporal context injection. Intra-segment encoding maps the analyzed segment to a fixed-length representation and applies a masking strategy to the gap-derived placeholders. The masking strategy skips the corresponding positions and records the masking bitmap during intra-segment encoding. The second channel of the pipeline is the graph topology branch, which receives the node features, edge features and neighborhood summaries of the graph feature set, and performs multi-hop message propagation, edge weight mapping and hierarchical restoration. Multi-hop message propagation propagates node status and neighborhood summaries along adjacency relationships on the graph in units of time points. The number of propagation steps is determined by the configuration parameters indicated by the disposal log index.

7. The method according to claim 1, characterized in that, The process of generating section load forecast results and confidence intervals also includes: For each time point, segment-level load forecasts are generated, and a confidence interval is generated outside the forecast values; The generation of credible intervals relies on two types of information: one is the quality and uncertainty indication provided by the masking bitmap and the anomaly annotation summary, which is used to characterize the completeness and stability of the input at that time point; the other is the external consistency indication provided by the neighborhood consistency marker and the rhythm label overview, which is used to characterize the cooperative state of the same or neighboring domains at that time point. Given upper and lower bounds under the constraints of two types of information, and record the source summary of the credible interval.

8. The method according to claim 1, characterized in that, The process of extracting correlation information from the segment load forecast results and the reliable interval, generating forecast summaries and registering model versions, writing on-chain data and generating traceable indexes, and calculating deviations based on real meter reading receipts and compiling parameter update proposals including parameter whitelists also includes: Write the session number, batch number, model version fingerprint, channel configuration summary, fusion strategy identifier and trace reference into the output record to form the segment load prediction result and the confidence interval; The segment load prediction results and trusted intervals are referenced to complete on-chain writing and traceable index generation, and are also referenced for deviation calculation and parameter update proposal preparation, constructing an on-chain prediction record and parameter update annotation structure.

Images