Green electricity transaction electricity price dynamic prediction system and method based on AI

Abstract

The invention discloses a green electricity transaction electricity price dynamic prediction system and method based on AI, and relates to the technical field of big data analysis. The method comprises the following steps: S1, collecting a green electricity price data set, and carrying out normalization and unified time base alignment processing; s2, constructing a time delay estimation model, a time deviation evaluation model and a unified time sequence view; s3, performing data splicing and time sequence sample construction to generate a time sequence sample view; s4, constructing a feature data set, performing sample quality evaluation, and constructing an electricity price dynamic prediction model; and S5, performing storage, data blood relationship management and prediction effect evaluation on the prediction result to form closed-loop feedback. According to the method, the time sequence consistency and the prediction credibility of the multi-source data in the dynamic prediction process of the green electricity transaction electricity price are effectively improved, and the problems of dynamic process reconstruction disorder and characteristic and label cross-sequence dislocation caused by non-uniform time reference and lack of multi-stage timestamp management are solved.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, specifically to an AI-based dynamic prediction system and method for green electricity trading prices. Background Technology

[0002] With the continuous expansion of green electricity trading, the increasing proportion of renewable energy installations, and the increasingly refined trading varieties and settlement cycles, more and more operational, metering, and market data are being introduced into the dynamic prediction and analysis of electricity prices. This places higher demands on the precise alignment and high-frequency reconstruction of multi-source data on a unified time axis. In existing technologies, data such as the output plans and actual output curves generated by the power dispatch system, the time-of-use electricity and metering price records generated by the metering system, the equipment status and operational event data uploaded by the field acquisition gateway, the declared and traded electricity price sequences recorded by the market platform, and the settlement price and fee details generated by the settlement platform are all continuously written into the big data platform with different time granularities and sampling methods. If a time-series correspondence refined to the second or even millisecond level cannot be established under a unified time benchmark, it will be difficult to accurately reconstruct the dynamic causal relationship between electricity price fluctuations and output changes, load fluctuations, and equipment status switching. How to construct a unified time benchmark and a multi-level timestamp management mechanism under complex multi-system, multi-link, and multi-time-granularity conditions has become a key prerequisite for supporting the improvement of the accuracy of big data analysis and dynamic electricity price prediction.

[0003] For example, invention patent CN119991351B discloses a method and system for optimal power supply strategy of new energy based on artificial intelligence, including: constructing an electricity price prediction model, a weather prediction model, and a population flow-load prediction model; constructing a dynamic scheduling optimization model based on power supply cost, curtailment rate, and load deviation, whereby the power supply cost is calculated based on the prediction results of the electricity price prediction model; the curtailment rate is determined based on the prediction results of the weather prediction model; the load deviation is calculated based on the prediction results of the population flow-load prediction model; and the system dynamically adjusts the economic, environmental, and stability strategies of the power grid based on the dynamic scheduling optimization model. The system includes a data collection module, an electricity price prediction module, a weather prediction module, a population flow-load prediction module, a dynamic scheduling optimization module, and a data output module. It can provide an efficient and intelligent solution for configuring optimal power supply strategy of new energy.

[0004] In existing technologies, existing systems are often loosely pieced together from multiple heterogeneous systems, such as dispatch automation systems, metering and data acquisition systems, field data acquisition gateways, market trading platforms, and settlement platforms. Each system maintains its own local clock and generates timestamps independently. Data acquisition messages typically only carry the time of a single device or business. When the data enters the big data platform, the platform's entry time is also added. There is a lack of unified time synchronization services and a lack of multi-level management and correlation constraints on device acquisition time, gateway reception time, and platform entry time. This leads to situations where the time sequence of the same physical process is easily reversed, the interval is abnormal, and there are cross-table misalignments. It is difficult to accurately reconstruct the dynamic correlation between electricity price changes and driving factors such as output, load, and operating status on a unified time axis.

[0005] Therefore, in order to address the above issues, there is an urgent need for an AI-based dynamic prediction system and method for green electricity trading prices. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an AI-based dynamic prediction system and method for green electricity trading prices, which solves the problems of inconsistent time bases and lack of multi-level timestamp management that lead to disordered dynamic process reconstruction and misalignment of features and labels across orders.

[0007] To achieve the above objectives, this invention provides the following technical solution: an AI-based dynamic prediction system and method for green electricity trading prices, comprising: S1, collecting a green electricity price dataset, normalizing and unifying the time base of the dataset to form a standardized data frame; S2, constructing a link delay estimation model and a time deviation evaluation model based on the standardized data frame to form a data source-level delay estimation parameter set and constructing a unified time series view for dynamic price prediction; S3, performing data splicing and time series sample construction on the standardized data frame to generate a time series sample view for dynamic price prediction; S4, constructing a feature dataset based on the data source-level delay estimation parameter set and evaluating sample quality, constructing a dynamic price prediction model to train, validate, and manage inference results on the feature dataset; S5, storing and managing the prediction results of the dynamic price prediction model, and evaluating the prediction effect on errors and stability to form a closed-loop feedback.

[0008] Furthermore, the specific process of collecting green electricity price datasets through multi-source system collaboration, and normalizing and unifying the time base alignment of the green electricity price datasets to form standardized data frames is as follows: Collecting green electricity price datasets, which include: transaction price and business result datasets, actual electricity price result data tables, dispatch and output datasets, load and monitoring measurement datasets, metering and settlement basic datasets, operation status and alarm datasets, time service and clock status datasets, and timestamp and channel identifier datasets; historical multi-source data sequences already stored under the same green electricity business scenario are used as historical data. The process involves running the data, marking the data source type and classifying the time attributes of the collected green electricity price dataset, and performing preliminary time alignment based on the data collection node reception time and the platform's data entry time to obtain a processed green electricity price dataset with multi-level timestamps. The processed green electricity price dataset is then encapsulated with a unified data structure to construct green electricity price collection data frames. The processed green electricity price dataset is then organized and cached through standardized collection data frame format, a global queue sorted by time, and a local queue sorted by data source type. Finally, the time-related fields in the collection data frames are preprocessed using z-score dimensionless normalization.

[0009] Furthermore, the specific process of constructing a link delay estimation model and a time deviation assessment model based on standardized data collection frames to form a data source-level delay estimation parameter set is as follows: Multi-level timestamp fields and time service and clock status datasets from the green electricity price collection data frames are input for joint modeling sample extraction to obtain a basic dataset for delay modeling. This basic dataset is then grouped, cleaned, and anomaly identified according to data source type, node combination, transmission path, and time period to construct the link delay estimation model and the time deviation assessment model. During the time deviation assessment process, the sign of the time deviation value directly observed by the node in the current time synchronization session is taken as the deviation sign, thus preserving the positive and negative directions of the deviation. The absolute value of the observed time deviation value is compared with the deviation baseline obtained in the previous time window. The absolute value of the difference is used to obtain the jump magnitude of the current observation deviation relative to the deviation baseline. The deviation jump sensitivity coefficient is multiplied by the jump magnitude of the current observation deviation relative to the deviation baseline to obtain the deviation jump term. The larger of the absolute value of the observation time deviation and the deviation jump term is selected as the effective deviation scale. The deviation sign is multiplied by the effective deviation scale to obtain the estimated time deviation of the node at the current time of synchronization. The obtained estimated time deviation is compared with the time deviation threshold in real time. When the estimated time deviation is greater than the time deviation threshold, the current node time state is marked as deviation exceeding the limit. When the estimated time deviation is less than or equal to the time deviation threshold, the current node time state is marked as deviation normal. The data source level time delay estimation parameter set and the time deviation estimation result dataset are output.

[0010] Furthermore, the specific process of constructing a unified time-series view for dynamic electricity price forecasting is as follows: Input green electricity price data frames and data source-level time delay estimation parameter sets; perform multi-level timestamp correction and unified analysis time; for data records with incomplete time fields or abnormal clock states, use the data source-level time delay estimation parameter set and residual time relationships to perform analysis time estimation and reliability marking; write the obtained unified analysis time field, along with the original multi-level timestamps, time precision markers, time source markers, and clock state marker time metadata, into a unified time-series index table; provide incremental update and batch recalculation mechanisms, and output the unified time-series index table, the dataset of analysis time and original time relationship, and the dataset of analysis time reliability statistics.

[0011] Furthermore, the specific process of stitching together standardized collected data frames and constructing time-series samples to generate time-series sample views for dynamic electricity price forecasting is as follows: Input a unified time-series index table and business data tables containing electricity price results, output data, load data, operating status data, and metering settlement data from the green electricity price collected data frames; perform unified analysis time alignment and sample reorganization; sort and group the green electricity price collected data frames on the unified analysis time axis to construct basic time-series sequence views by data source and by region; perform time window segmentation and sample organization on the basic time-series sequence views to construct a time-series sample dataset; utilize the analysis time reliability level and analysis time residual index in the unified time-series index table to perform quality labeling and level classification on the constructed time-series samples, forming a sample quality labeling system for forecasting; based on the time-series sample dataset, construct time-series sample views adapted to different forecasting tasks; output multi-source basic time-series sequence views, time-series sample datasets, and various types of time-series sample views.

[0012] Furthermore, the specific process of constructing a feature dataset and evaluating sample quality based on the data source-level time delay estimation parameter set is as follows: Input the time series sample dataset, extract feature fields and target fields, and construct a feature dataset view library for dynamic electricity price prediction; construct a sample quality evaluation mechanism, and for each time series sample, evaluate the sample quality based on the analysis time residual index, time reliability level, feature missing ratio, and number of outlier markers; multiply the time consistency sensitivity coefficient by the sample time consistency deviation to obtain the consistency term, and perform exponential operation after taking the negative value of the consistency term to obtain the time alignment error term; the missing rate sensitivity coefficient is multiplied by the number of missing effective fields in the sample. The missing term is obtained by multiplying the rates. The negative value of the missing term is then used for exponential operation to obtain the feature missing status term. The feature missing status term is multiplied by the time alignment error term to obtain the sample quality assessment value. When the sample quality assessment value is lower than the sample quality threshold, the sample is marked as a low-quality sample for rejection or only for robustness testing scenarios. When the sample quality assessment value is greater than or equal to the sample quality threshold, the sample is marked as a usable sample and enters the feature selection and dataset partitioning process. Feature datasets for different prediction tasks are constructed. The output features view library, sample quality assessment value dataset, feature dataset views corresponding to different modeling scenarios, feature dataset, and target dataset are output.

[0013] Furthermore, the specific process of constructing a dynamic electricity price prediction model for training, validating, and managing inference results on the feature dataset is as follows: The dynamic electricity price prediction model is constructed by inputting the feature dataset and the target dataset, and time-ordered partitioning and version management are performed; a time-series deep network is invoked to complete the training and performance evaluation of the dynamic electricity price prediction model; a dynamic electricity price prediction inference pipeline is constructed, performing streaming inference on the feature dataset and managing the prediction result data to obtain the prediction result dataset; error evaluation and operation monitoring are performed on the prediction result dataset to form the prediction error time series and distribution characteristics; the training record table of the dynamic electricity price prediction model, the model version library, the prediction result dataset, and the prediction error evaluation dataset are output.

[0014] Furthermore, the specific process for storing and managing the data lineage of the electricity price dynamic forecasting model is as follows: input the forecast result dataset and a unified time-series index table; perform unified analysis, time alignment, and metadata completion on each forecast result to construct a forecast result record for electricity price dynamic forecasting; associate the forecast result records with identifiers to construct a forecast result data table and a data lineage view; construct a versioned management and fast retrieval mechanism for the forecast results, supporting multi-dimensional queries by time, model version, and data caliber; output the forecast result data table, the forecast result data lineage view, and the forecast version management table.

[0015] Furthermore, the specific process for evaluating the prediction effect and forming a closed-loop feedback of error and stability is as follows: Input the prediction result data table and the actual electricity price result data table; perform unified analysis, time alignment, and error calculation on the prediction results and actual results to construct the error time series and error distribution dataset; construct a prediction stability evaluation mechanism to perform fine-grained evaluation on the performance on the time axis; exponentiate the ratio of the absolute value of the prediction error at the current time to the error amplitude reference value to obtain the error amplitude term; exponentiate the ratio of the absolute value of the difference between prediction errors at adjacent times to the error jump reference value to obtain the error jump term; multiply the error amplitude term and the error jump term to obtain the error term; multiply the error term by the negative value of the stability sensitivity coefficient to obtain the error stability term; perform exponential operation on the error stability term to obtain the prediction stability index at the k-th unified analysis time point; correlate the prediction error, prediction stability index, data quality index, and sample quality evaluation value to construct a prediction effect evaluation view; output the prediction error time series dataset and the prediction stability index dataset.

[0016] Furthermore, the second aspect of this invention provides an AI-based dynamic prediction system for green electricity trading prices, applied to an AI-based dynamic prediction method for green electricity trading prices, comprising: a multi-source data acquisition module, used to collaboratively acquire green electricity price datasets through multi-source systems, normalize and unify the time base alignment of the green electricity price datasets to form standardized acquisition data frames; a latency estimation module, used to construct a link latency estimation model and a time deviation evaluation model based on the standardized acquisition data frames, forming a data source-level latency estimation parameter set, and constructing a unified time series view for dynamic price prediction; a time series construction module, used to perform data splicing and time series sample construction on the standardized acquisition data frames, generating a time series sample view for dynamic price prediction; a price prediction model inference module, used to construct a feature dataset based on the data source-level latency estimation parameter set and perform sample quality evaluation, construct a dynamic price prediction model, and perform training, verification, and inference result management on the feature dataset; and a model evaluation module, used to store the prediction results of the dynamic price prediction model and manage data lineage, and evaluate the prediction effect on error and stability to form a closed-loop feedback.

[0017] The present invention has the following beneficial effects: (1) This invention introduces a unified time service and performs multi-level timestamp management and unified analysis time calculation for the equipment collection time, the collection node receiving time, and the platform entry time. This eliminates the time sequence disorder caused by the inconsistency of time bases of multiple systems and the uncertainty of link delay from the source of data, and ensures the fine alignment of multi-source data related to green electricity transactions on a unified time axis.

[0018] (2) This invention maps transaction results, output curves, load curves, operating status, and metering and settlement data onto a unified analysis time axis, constructs a basic time series view and time window samples, and realizes the organization of time series samples of multiple scales such as short cycle, medium cycle and long cycle, providing a structured data foundation covering different time scales for dynamic prediction of electricity prices.

[0019] (3) This invention constructs a prediction result data table and a data lineage view under a unified analysis time for the prediction results, and associates each prediction value with the corresponding sample version, feature set, data time window and model version one by one, so as to realize the full-link traceability management of the prediction results from input data to the configuration of the dynamic electricity price prediction model, and provides a refined basis for error analysis and optimization of the dynamic electricity price prediction model.

[0020] (4) This invention, by integrating the prediction error time series, prediction stability index, time deviation estimation parameter, time delay structure deviation index and data source quality index on a unified analysis time axis, constructs a prediction effect evaluation view and closed-loop feedback mechanism, which can promptly identify the correlation between prediction performance degradation and data quality decline, guide the adjustment of time correction strategy, sample selection conditions and electricity price dynamic prediction model training strategy, thereby continuously improving the accuracy and stability of green electricity trading electricity price dynamic prediction.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] Figure 1 This is a flowchart of the AI-based dynamic prediction method for green electricity trading prices according to the present invention. Figure 2 This is a framework diagram of the AI-based green electricity trading price dynamic prediction system of the present invention; Figure 3 This is a graph showing the changes in the estimated time deviation values ​​at each node of the present invention. Figure 4 This is a unified time-series view for dynamic electricity price forecasting in this invention; Figure 5 This is a flowchart illustrating the management and data lineage of electricity price forecast results based on a unified analysis time, as presented in this invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figures 1-5 This invention provides a technical solution: an AI-based dynamic prediction system and method for green electricity trading prices, comprising: S1, collecting a green electricity price dataset, normalizing and unifying the time base of the dataset to form a standardized data frame; S2, constructing a link delay estimation model and a time deviation evaluation model based on the standardized data frame to form a data source-level delay estimation parameter set and constructing a unified time series view for dynamic price prediction; S3, performing data splicing and time series sample construction on the standardized data frame to generate a time series sample view for dynamic price prediction; S4, constructing a feature dataset based on the data source-level delay estimation parameter set and evaluating sample quality, constructing a dynamic price prediction model to train, validate, and manage inference results on the feature dataset; S5, storing and managing the prediction results of the dynamic price prediction model, and evaluating the prediction effect on errors and stability to form a closed-loop feedback.

[0025] Specifically, the process of collecting green electricity price datasets through multi-source system collaboration, and then normalizing and unifying the time base of these datasets to form standardized data frames is as follows: The green electricity price dataset includes: transaction price and business result datasets, actual price result data tables, dispatch and output datasets, load and monitoring measurement datasets, metering and settlement basic datasets, operation status and alarm datasets, time service and clock status datasets, and timestamp and channel identifier datasets. The multi-source systems include: dispatch automation systems, metering acquisition systems, field acquisition gateways, market trading platforms, and settlement management platforms. The transaction price and business result datasets include: data collected by time granularity. The recorded electricity price values, corresponding electricity consumption values, business classification tags, settlement cycle tags, and data source tags are obtained by structured extraction from the electricity price-related result tables in the business platform. The price fields, electricity consumption fields, and associated time index fields in the records are then rewritten into a unified format for electricity price result records, serving as one of the target data sources for dynamic prediction. The actual electricity price result data table includes: a unified analysis time field, an actual electricity value field, a metering point identifier field, a logical channel number field, a metering cycle identifier field, a settlement cycle identifier field, a metered electricity consumption field, a currency identifier field, a data source type tag field, a data quality tag field, and a sample version tag field. The scheduling and output dataset includes: The time-period output values, estimated renewable energy output ratios, estimated reserve capacity, and operation mode labels are generated by cleaning and mapping the time-series operational measurement point records from the dispatch data source. This unifies the units and time granularities of various power and output physical quantities into continuous time series, used to characterize the operational status background related to electricity prices. The load and monitoring measurement dataset includes: regional load measurements, monitoring point power curves, and voltage and current measurements of important nodes. This is achieved by aggregating measurement point data from the load monitoring system and the distribution and utilization monitoring system, unifying the original measurement records organized by measurement point codes in different systems to load and measurement records with unified measurement point identifiers and unified time granularities. Metering The settlement foundation dataset includes: basic metering point files, metering cycle quantities, settlement cycle quantities, and verification tags. It is obtained by extracting the basic tables and cycle quantity tables of the metering system and the settlement system, and uniformly encoding and aligning the metering point codes, metering cycle start and end times, and measurement results with time. This is used for data-level verification of the relationship between electricity prices and actual measurements. The operation status and alarm dataset includes: equipment status tags, important measurement point over-limit tags, communication link interruption tags, and data quality tags. It is obtained by converting the status change logs recorded by events in the operation monitoring system and alarm system, and reorganizing discrete events into a state timeline in chronological order. This is used to add operation environment tags to time series data during the sample construction stage.The time service and clock status dataset includes: time synchronization session time, reference time, node local time, time deviation value, time synchronization round-trip time, and clock health level for each node. This is achieved by periodically synchronizing time with big data platform nodes, acquisition gateway nodes, and key system nodes through a unified time synchronization component, writing the measured deviations and round-trip delays into a dedicated data table in structured record form. The timestamp and channel identifier dataset includes: device acquisition time, acquisition node reception time, platform entry time, time precision marker, time source marker, data source type marker, logical channel number, and acquisition batch number. This is obtained by adding time and channel attribute fields to each data record encapsulated according to the unified access specification at the acquisition gateway and big data access layer, used to distinguish data records from different sources and channels on a unified timeline. Historical multi-source data sequences already stored under the same green electricity business scenario serve as historical operating data. Historical operating data includes: historical electricity price result sequences within different time periods, historical output and load curves, historical operating status and alarm event sequences, historical metering and settlement cycle quantities, and corresponding clock status data and timestamp distribution, used to statistically analyze link latency characteristics, clock deviation patterns, and the relationships between multi-source time fields.

[0026] The collected green electricity price dataset is categorized and managed according to data source type and time attribute. It is classified into transaction result, operational measurement, metering and settlement, operational status, and time quality categories. Time quality and timestamp records are prioritized for writing to the time metadata table in the access cache queue. The device acquisition time, acquisition node reception time, and platform entry time of each record are checked for time field integrity. Records with missing time fields are marked with a time missing flag and placed in a separate completion queue. A unified time synchronization component is used to distribute time bases to the big data platform nodes and acquisition gateway nodes. Clock correction is performed periodically on nodes supporting local time calibration, while nodes not supporting local calibration are... The green electricity price dataset is managed by recording clock deviations. Initial time alignment is performed based on the data acquisition node's receiving time and the platform's data entry time to obtain a processed dataset with multi-level timestamps. The device time field, data acquisition node's receiving time field, and platform's data entry time field in the green electricity price dataset are uniformly encoded, and millisecond-level time precision information is written into the time precision flag field. Based on the deviation range between the time service and clock status dataset records, time records exceeding the deviation range are marked as abnormal. By maintaining a global receiving queue ordered by time and a local receiving queue divided by data source type at the data acquisition node, a stable receiving order for data records from different systems is ensured.

[0027] The processed green electricity price dataset is encapsulated with a unified data structure to construct green electricity price collection data frames. These frames include: data source type marker, logical channel number, collection batch number, business data content field, measurement unit marker, equipment collection time, collection node reception time, platform data entry time, time precision marker, time source marker, clock status marker, and data quality marker. Analysis time and time alignment marker fields are reserved. The processed green electricity price dataset is then organized and cached using a standardized collection data frame format, a global queue sorted by time, and a local queue sorted by data source type. Based on the timestamp and the logical channel number and collection batch number in the dataset, the collection data frames are archived by channel and batch. The caching strategy prioritizes data frames with complete timestamps and normal clock status, while data frames with missing time fields or severe clock deviations are retained in limited quantities and marked separately to avoid interference with subsequent large-scale statistics. Finally, the time-related fields in the collection frames undergo z-score dimensionless normalization preprocessing. Timestamps recorded in different systems with different time precisions are mapped to a unified time representation through a unified time format and precision mark; physical quantity fields are linearly scaled and normalized according to a unified unit system to eliminate the dimensional differences caused by different systems and configurations; a global acquisition queue sorted by the reserved field of analysis time and a local acquisition queue grouped by data source type are maintained on the acquisition node.

[0028] This implementation plan collects green electricity price data and introduces unified field specifications, a unified unit system, and z-score dimensionless normalization, uniformly encapsulating it into a consistent green electricity price data frame. Simultaneously, based on time service and clock status data, multi-level timestamp verification, anomaly marking, and unified precision encoding are performed. A queue and archiving mechanism is built at the access end, managed in both chronological order and data source type. This ensures that the data entering the latency estimation and time series construction modules possesses unified, traceable, and filterable characteristics in terms of source labeling, physical quantity units, and time reference, fundamentally improving the temporal consistency, data quality controllability, and reliability of batch statistical analysis of green electricity price data.

[0029] Specifically, the process of constructing a link latency estimation model and a time deviation assessment model based on standardized data collection frames to form a data source-level latency estimation parameter set is as follows: Inputting multi-level timestamp fields from the green electricity price collection data frame and time service and clock status datasets for joint modeling sample extraction; reading device acquisition time, acquisition node reception time, platform entry time, time precision marker, time source marker, data source type marker, logical channel number, and acquisition batch number from the green electricity price collection data frame, and using records with complete time fields as candidate samples; reading time synchronization session time, unified time reference, node local time, time deviation value, time synchronization round-trip time, and clock health level from the time service and clock status datasets, and associating them with candidate samples according to node identifier and time interval to obtain the clock deviation and clock health status corresponding to each candidate sample at the acquisition time; and filtering out records with severely missing time fields or clock health levels below the health threshold to obtain the latency modeling basic dataset.

[0030] The latency modeling dataset is grouped, cleaned, and anomaly identified based on data source type, node combination, transmission path, and time period. Records with missing device acquisition time, missing acquisition node reception time, or missing platform data entry time are marked as time-missing samples and removed or archived separately. The total link time difference is calculated, including the difference between device acquisition time and acquisition node reception time, the difference between acquisition node reception time and platform data entry time, and the sum of the two. Based on the time difference sequence of similar links, samples falling outside the upper and lower bounds of the time difference are considered candidate anomalies violating deviation constraints. Simultaneously, based on the time synchronization round-trip time, network link bandwidth parameters, and node queue latency limits recorded in the time service and clock status datasets, the latency from device to acquisition node, acquisition time, and data entry time are calculated. The theoretical minimum and maximum latency of the node-to-platform and end-to-end total links are calculated. The actual time difference is compared with the theoretical minimum and maximum latency. When the actual time difference is less than the theoretical minimum latency or greater than the theoretical maximum latency of the corresponding link, the record is considered to violate the physical rationality of the link. For samples exceeding the deviation range threshold, cross-validation is performed using the time deviation value and time synchronization round-trip time in the time service and clock status datasets. Records that simultaneously violate deviation constraints and physical rationality of the link are marked as time anomaly samples and removed from the modeling samples. The remaining samples are bucketed according to data source type, node combination, transmission path, and date range, so that each bucket of samples has relatively stable link characteristics and clock status characteristics in a statistical sense.

[0031] Statistical analysis and fitting are performed on the time difference and time deviation in each sample bucket to construct a link delay estimation model and a time deviation evaluation model. For each type of data source and each typical transmission path, the first segment delay distribution from the device acquisition time to the acquisition node reception time, the second segment delay distribution from the acquisition node reception time to the platform storage time, and the total link delay distribution of the sum of the two segments are calculated to obtain the statistical indicators of mean, standard deviation, quantiles, and extreme value range. Piecewise linear models and distribution fitting models are used to describe and store various delay distributions. For each node, the time deviation value centrally recorded in the time service and clock status data is used to calculate the deviation baseline, long-term drift trend, and short-term fluctuation range between the node's local clock and the unified time reference in chronological order to obtain a time deviation evaluation model, which is used to estimate the expected deviation between the node's local time and the unified time reference at a certain point in time.

[0032] In the time deviation assessment process, the sign of the time deviation value directly observed by the node in the current time synchronization session is taken as the deviation sign, thus preserving the positive and negative directions of the deviation. The absolute value of the observed time deviation value, and the absolute value of the difference between the current observed deviation and the deviation baseline obtained in the previous time window, are used to obtain the jump magnitude of the current observed deviation relative to the deviation baseline. The deviation jump sensitivity coefficient is multiplied by the jump magnitude of the current observed deviation relative to the deviation baseline to obtain the deviation jump term. The larger of the absolute value of the observed time deviation value and the deviation jump term is selected as the effective deviation scale. The deviation sign is multiplied by the effective deviation scale to obtain the estimated time deviation value of the node at the current time synchronization moment. The specific calculation formula for the estimated time deviation value is as follows: ; In the formula, This represents the estimated time deviation of the node at the current time synchronization moment, used to correct the unified time deviation output of the device acquisition time, the acquisition node reception time, and the platform data entry time. This represents the time deviation value directly observed by the node in the current time synchronization session. It is calculated by the node interacting with the time synchronization source, recording the timestamps of the request and response, and subtracting the known interference term of propagation delay. It is used to provide direct observation data of the node's time deviation at the current moment. The deviation baseline obtained within the previous time window is obtained by collecting multiple sets of time deviations within the previous time window and performing statistical processing using the median method. It is used to characterize the normal deviation level of the node in the short term. It represents the jump magnitude of the current observation deviation relative to the deviation baseline. It is obtained by directly subtracting the current observed time deviation value from the deviation baseline obtained in the previous time window. It is used to quantify the fluctuation magnitude of the current observation deviation relative to the normal level. This represents the bias jump sensitivity coefficient, which is obtained by testing the bias correction effect under different scenarios. The value range is a positive number greater than zero, and it is used to adjust the amplification of the jump term on the final bias estimate. This is a sign function, obtained by directly calling the mathematical definition (outputting 1 when the input value is positive, -1 when it is negative, and 0 when it is 0), used to preserve the positive and negative directions of the deviation; The operation takes the larger of the current absolute value of the deviation and the jump amplitude after sensitivity amplification as the effective deviation scale, realizing an adaptive time deviation estimation mechanism that takes the observation deviation as the main factor when the deviation is stable and the jump amplitude as the main factor when there is a sharp shift in a short period of time.

[0033] The obtained time deviation estimate is compared with the time deviation threshold in real time. When the time deviation estimate is greater than the time deviation threshold, the current node's time status is marked as exceeding the deviation limit, the exceeding limit mark is written into the node's time status record, and a high deviation mark is added to the data records generated by the node in the corresponding time period. When the time deviation estimate is less than or equal to the time deviation threshold, the current node's time status is marked as normal, the normal mark is written into the node's time status record, a normal deviation mark is added to the data records generated by the node in the corresponding time period, and the statistics in the node's deviation observation sliding window are updated. The current time deviation estimate is included in the calculation range of the deviation median and deviation variance. The link delay estimation model and the time deviation evaluation model are organized into a data source-level delay estimation parameter set in parameterized form, recording the data source type, node combination, transmission path, effective time interval, and corresponding delay statistics and deviation evaluation parameters.

[0034] Table 1 shows the node time deviation estimation data, covering the calculation process and results of the time deviation for two nodes at different time synchronization times: NODE_01 Time Deviation Details: At the time synchronization time of 2024-05-20-08:00:00, the direct observation time deviation value is 28.3. After the node interacts with the time synchronization source, the current direct observation deviation obtained by deducting the propagation delay is positive, indicating that the node's local time is slightly ahead of the unified time reference; the baseline deviation of the previous time window is 25.1, which is calculated by the median of multiple sets of historical observation deviations within the previous time window, reflecting the short-term normal deviation level of the node; the jump amplitude is 3.2, and the current observation deviation is relative to the baseline. The positive fluctuation amplitude of the line quantifies short-term deviation changes; the deviation jump sensitivity coefficient is 1.2, a positive coefficient set through scenario testing, used to adjust the impact of jump terms on the estimated value; the effective deviation scale is 28.3, because the absolute value of the observed deviation is greater than the amplified jump amplitude, the observed deviation is taken as the effective scale; the estimated time deviation value is 28.3, retaining the positive sign of the observed deviation, and multiplying it by the effective deviation scale to obtain the final estimated value; the time deviation threshold is 50, the deviation critical value set by the system, used to determine the node time status; when the node time status is normal deviation, the data record is marked as normal deviation, and the node deviation sliding window statistics are updated simultaneously. At time synchronization on 2024-05-20-08:05:00, the directly observed time deviation was -15.7, a negative value indicating that the node's local time was slightly slower than the unified time reference; the baseline deviation of the previous time window was -18.2, the median result of the historical deviation of the previous window; the jump magnitude was 2.5, a positive jump in the current deviation relative to the baseline (the absolute value of the deviation decreased); the effective deviation scale was 15.7, the estimated time deviation was -15.7, retaining the negative sign; the node status was normal deviation, and the marker deviation was normal. At time synchronization on 2024-05-20-08:10:00, the directly observed time deviation was 42.6, the baseline deviation of the previous time window was 38.9, the jump magnitude was 3.7, the effective deviation scale was 42.6, the estimated time deviation was 42.6, the node status was normal, and the marker deviation was normal. Overall situation of NODE_01: The estimated time deviation values ​​of the three time synchronization observations are 28.3, -15.7 and 42.6 respectively. The deviation fluctuation is small, indicating that the local clock of the node is relatively stable and the time deviation is within the normal range.

[0035] NODE_02 Time Deviation Details: At time synchronization of 2024-05-20-08:00:00, the directly observed time deviation value is 62.5, the previous time window deviation baseline is 22.3, the jump amplitude is 40.2, the effective deviation scale is 62.5, the estimated time deviation value is 62.5, the node time status is deviation exceeding the limit, and the data record is marked as high deviation. An exceeding-limit mark is written to the node time status record. At time synchronization of 2024-05-20-08:15:00, the directly observed time deviation value is -58.7, the previous time window deviation baseline is -25.4, the jump amplitude is -33.3, the effective deviation scale is 58.7, the estimated time deviation value is -58.7, the node time status is deviation exceeding the limit, and the data record is marked as high deviation. Overall, the estimated time deviations of the two timing observations for NODE_02 were 62.5 and -58.7, respectively, with a large jump, indicating that there may be abnormal fluctuations in the local clock of the node (such as network interference or hardware failure). It is necessary to focus on monitoring and investigating the cause of the deviation exceeding the limit.

[0036] Table 1. Node Time Deviation Estimation Data

[0037] like Figure 3 The graph shows the time deviation estimates for each node. The horizontal axis represents the time synchronization time under a unified time reference, and the vertical axis represents the estimated time deviation obtained from the corresponding time synchronization session. The graph shows the time deviation estimation curves for nodes one through five. It can be seen that the curves for nodes one, two, and three are generally distributed in the lower range with small fluctuations, indicating that the deviation of the local clock relative to the unified time reference is small and stable, making them suitable as highly reliable time sources. The curves for nodes four and five are generally higher, with obvious peaks at certain times, reflecting that these two nodes have unstable time synchronization or increased local clock drift during certain periods. They are objects that need to be focused on and marked as having deviation risks or exceeding deviation limits. By comparing the time deviation estimation curves of different nodes in the same coordinate system, nodes and time periods with poor time synchronization quality can be intuitively identified, providing a basis for reducing the weight or eliminating data from nodes with high deviations.

[0038] A parameter update and version management mechanism is constructed to periodically re-evaluate and refit parameters for link characteristics and clock deviations that change with data accumulation. Incremental samples are extracted from the latest green electricity price data frames and time service and clock status datasets, and the same cleaning and bucketing strategies as in the initial modeling phase are used to update the latency modeling base dataset. Latency and deviation statistics are recalculated for each parameter bucket and compared with existing parameters. When the change in statistical indicators exceeds the statistical deviation threshold, a new version of latency estimation parameters and deviation evaluation parameters is generated for the bucket, and the version number, effective time, and expiration time are recorded in the parameter library. The effective parameter range for each time period is maintained through a parameter version management table. The data source-level latency estimation parameter set and time deviation estimation result dataset are output.

[0039] In this implementation plan, by jointly modeling the multi-level timestamps and time synchronization status of multi-source green electricity price collection data, a link delay estimation model and a time deviation evaluation model are constructed to ensure that the samples participating in the statistics have stable link characteristics and reliable clock status. This enables the unified analysis time calculation to call matching delay and deviation parameters for different links, achieving fine correction and dynamic updates of device collection time, collection node reception time, and platform data entry time, thereby significantly improving the alignment accuracy of multi-source data on a unified time axis and the reliability of long-term time series reconstruction.

[0040] Specifically, the process of constructing a unified time-series view for dynamic electricity price forecasting is as follows: Input green electricity price data frames and a set of data source-level latency estimation parameters; perform multi-level timestamp correction and unified analysis time; for records that simultaneously possess device acquisition time, acquisition node reception time, and platform entry time, based on the time deviation evaluation model of the corresponding node within the record's time period, subtract the respective time deviation estimates from the device acquisition time, acquisition node reception time, and platform entry time, converting the three to a corrected time under a unified time benchmark; then, based on the data source type and transmission path, read the corresponding first and second segment latency estimation parameters from the latency estimation parameter set; using the device acquisition time correction value as a reference, comprehensively consider the link forwarding characteristics from the device side to the acquisition node and from the acquisition node to the platform through superposition to obtain the unified analysis time of the record; calculate the residual between the unified analysis time and each original time field, and write the magnitude and direction of the residual into the analysis time residual field using numerical values ​​or markers.

[0041] For data records with incomplete time fields or abnormal clock states, the analysis time estimation and reliability labeling are performed using the data source level delay estimation parameter set and residual time relationship. By binding the analysis time reliability label to the unified analysis time field, the same unified analysis time slice carries time reliability information on the global time axis at the same time, automatically filtering or downweighting low reliability records, thereby significantly reducing the risk of cross-source misalignment caused by time deviations from different data sources.

[0042] The residual time relationship refers to the difference between the unified analysis time and the original timestamps at each level, and its classification results. This includes the difference between the unified analysis time and the device acquisition time, the difference between the unified analysis time and the acquisition node reception time, and the residual level label of the sum of the unified analysis time and the platform entry time. The residual level label's value range is {alignment normal, forward deviation, lag deviation, suspected anomaly}, used to indicate the type of residual deviation in time alignment for this record. For records lacking device acquisition time but possessing acquisition node reception time and platform entry time, the second-segment delay statistical parameters in the link delay estimation model and the node time deviation evaluation model are used, based on the acquisition node reception time... The system works by inversely calculating a reasonable time interval under a unified time base, extracting a representative time point from this interval as the unified analysis time, and marking it as an estimate based on the receiving time in the analysis time reliability field. For records with platform entry times, a more conservative latency estimation method is used, obtaining an approximate range of the unified analysis time based on the maximum and minimum total link latency. When selecting a representative time point as the analysis time, the reliability level is marked as low. For data records marked as having abnormal clock states, even if the time field is complete, a deviation range and conservatism are introduced during the analysis time calculation process to increase the expected range of the analysis time residual, which is then marked in the reliability field. The time precision marker field characterizes the minimum time resolution of the current record's timestamp, with values ​​in the range of {seconds, milliseconds, microseconds, and others}. The time source marker field identifies the source of the timestamp, with values ​​in the range of {device local clock, local time after time synchronization by the acquisition gateway, big data platform entry time, and external time source alignment time}, used to distinguish the credibility and weight of different time sources in time deviation assessment and reliability analysis.

[0043] The obtained unified analysis time field, along with the original multi-level timestamps, time precision markers, time source markers, and clock status markers, is written into a unified time-series index table. The unified time-series index table includes: a unified analysis time field, device acquisition time, acquisition node reception time, platform data entry time, time precision marker, time source marker, clock status marker, data source type marker, logical channel number, acquisition batch number, analysis time calculation method marker, analysis time reliability level, analysis time residual index, and time alignment status marker. This is achieved by generating a corresponding index record for each green electricity price acquisition data frame, with one unified time-series index record corresponding to each acquisition data frame. The unified time-series index table establishes a composite index based on the unified analysis time and data source type, and supports efficient querying and sorting by time range, data source type, and reliability level.

[0044] It provides incremental update and batch recalculation mechanisms. When the latency estimation parameter set is updated based on the latest clock state data and historical data to generate a new version, the affected data time period is located through the time validity interval recorded in the parameter version management table. Batch recalculation or incremental correction is performed on the analysis time field in the unified time series index table within the time period, associating the new latency estimation parameters with the original records. During the recalculation process, the difference index between the old version analysis time and the new version analysis time is retained for each record. For newly written incremental data records, the latest version of the latency estimation parameters is directly applied to calculate the analysis time, ensuring that all records after a certain time point use a unified parameter system. Output includes a unified time series index table, a dataset of the relationship between analysis time and original time, and a dataset of analysis time reliability statistics.

[0045] In this implementation scheme, a unified analysis time calculation mechanism based on a data source-level latency estimation parameter set is introduced. This mechanism not only performs fine correction on data with complete time fields, but also provides analysis time estimates with reliability levels for data with missing time or clock anomalies. Through a unified time series index table, metadata is managed in an integrated manner, and with the help of incremental update and batch recalculation mechanisms, the consistency and traceability of the unified time series view can be maintained even when latency parameters evolve. This provides an accurate, unified, and queryable time basis for time series sample construction, dynamic electricity price forecasting, and effect evaluation.

[0046] Specifically, the process of stitching together standardized data collection frames and constructing time-series samples to generate a time-series sample view for dynamic electricity price forecasting is as follows: Input a unified time-series index table and business data tables containing electricity price results, output data, load data, operating status data, and metering settlement data from the green electricity price collection data frames; perform unified analysis, time alignment, and sample reorganization; using the unified time-series index table as a unified entry point for time and identifiers, associate each index record with the corresponding electricity price result record, output record, load record, operating status record, and metering settlement record through record identifiers or logical channel numbers; and link the electricity price value, output measurement point value, load measurement point value, operating status marker, and metering settlement quantity. The segments are reorganized into multi-source aligned time series according to the unified analysis time. During the association process, the unified analysis time field, analysis time calculation method marker, analysis time reliability level, and analysis time residual index of the unified time series index table are referenced. The device acquisition time, acquisition node reception time, platform entry time and time accuracy, time source, and clock status time metadata are appended to the reorganized record. The time correction link and time quality information are preserved, without recalculating the unified analysis time itself. For index records where the corresponding details cannot be found in the business data table, missing markers or placeholder values ​​are filled according to the data source type and time granularity, and the missing type and missing ratio are recorded in the sample status field.

[0047] The green electricity price data frames are sorted and grouped on the unified analysis timeline to construct basic time series views by data source and by region. Using the unified analysis time as the sorting field, records under the same data source type and logical channel number are sorted by time to form basic time series for electricity price, output, load, operation status, and metering settlement. For records with spatial division or business partition attributes, the basic time series are further grouped according to the region identifier or business partition identifier to construct partitioned time series views by region or business unit, and the region number, channel set, and time coverage are recorded in the view metadata. The intervals between adjacent unified analysis times are detected in the basic time series views. When an interval exceeds the time granularity multiple threshold, the interval is marked as a sparse or missing interval, and the time range and gap length of the interval are recorded in the time series view metadata.

[0048] The system performs time window segmentation and sample organization on the basic time series view to construct a time series sample dataset. Based on the needs of dynamic electricity price forecasting, it configures various time window types: short-cycle, medium-cycle, and long-cycle windows. Short-cycle windows are used to depict rapid fluctuations within the nearest time period, medium-cycle windows are used to depict intraday or inter-day trends, and long-cycle windows are used to depict seasonal and periodic characteristics. Using a unified analysis time as a benchmark, the basic time series is segmented on the time axis using sliding or rolling windows. Electricity price, output, load, operating status, and metering settlement sequences within each window are extracted to form a time series sample. Each sample is assigned a unique sample identifier, and the system records the unified analysis time at the start and end of the window, the window length, sliding step size, window type, and the distribution of analysis time reliability levels at each time point within the window. For cases where there are unified analysis time gaps within a window, the system selects interpolation, retaining missing data, or truncation strategies according to the configuration, and records the processing method and gap ratio in the sample metadata.

[0049] Using the analysis time reliability level and analysis time residual index in the unified time series index table, the constructed time series samples are quality-labeled and classified, forming a sample quality labeling system oriented towards prediction. For each time series sample, the distribution of analysis time reliability level and the mean, maximum, and quantile of analysis time residuals recorded within its window are statistically analyzed. Samples with high analysis time reliability and small overall residuals are labeled as Level 1 quality time series samples; samples with unreliable time points but controllable overall residuals are labeled as Level 2 equal quality time series samples; and samples with a high proportion of unreliable time points or large residuals are labeled as Level 3 quality time series samples. The quality level, the proportion of unreliable time points within the window, and the residual statistical index are recorded in the sample quality field. During the quality labeling process, the calculation relationship between the unified analysis time and the original timestamp is not changed; only the existing statistical results in the analysis time residual dataset are referenced.

[0050] Based on the time-series sample dataset, time-series sample views adapted to different prediction tasks are constructed. For short-term electricity price dynamic prediction tasks, short-period window samples are selected from the time-series sample dataset as the main sample set, while medium- and long-period window samples are used as auxiliary reference sets to extract background trend features. For medium- and long-term electricity price trend judgment tasks, medium- and long-period window samples are used as the main samples, while short-period window samples are used to characterize local disturbances. A time-series sample view definition is established for each prediction task through configuration, specifying the window type, window length, sliding step size, quality level lower limit, and missing data handling strategy. A subset of samples meeting the conditions is extracted from the time-series sample dataset to form the corresponding time-series sample view. The view identifier, window type combination used, time range, sample quantity, and quality level distribution are recorded in the time-series sample view metadata table. Time-series sample inputs with consistent structure, unified caliber, and controllable time quality can be directly obtained by view identifier, realizing a smooth transition from a unified time-series index view to a unified time-series sample view. Outputs multi-source basic time-series sequence views, time-series sample datasets, and various types of time-series sample views.

[0051] like Figure 4 The unified time-series view for dynamic electricity price forecasting shown uses the unified analysis time as the horizontal axis and the electricity price value as the vertical axis to align and display the electricity price sequences of different metering channels under a unified time benchmark. The figure shows the electricity price change curves of metering channel 1, metering channel 2, and metering channel 3 over multiple consecutive natural days. It can be seen that the electricity prices of each channel exhibit similar periodic fluctuations within the day, and the peak and valley times basically overlap on the unified analysis time axis, with differences only in local amplitude and detailed fluctuations. This indicates that after unified time-series indexing and unified analysis time correction, the electricity price records of different data sources have been accurately mapped to the same time coordinate system, which not only preserves the cross-channel comparison relationship, but also provides an intuitive time-series view basis for constructing multi-channel joint features and carrying out dynamic electricity price forecasting on the unified time axis.

[0052] In this implementation plan, by aligning and splicing multi-source business data, constructing basic time-series sequences by source and partition, and segmenting and classifying time-series samples based on a unified analysis time, the data is organized into a time-series sample view with a unified structure, consistent time caliber, and quality labels. This achieves a smooth transition from a unified time-series index to a unified time-series sample input, providing a high-quality data foundation for the electricity price dynamic forecasting model to stably learn multi-source joint time-series features on a unified time axis and distinguish different forecast time domain requirements.

[0053] Specifically, the process of constructing a feature dataset based on the data source-level time delay estimation parameter set and evaluating sample quality is as follows: Input the time-series sample dataset, extract feature fields and target fields, and construct a feature dataset view library for dynamic electricity price prediction; read the unified analysis time field, electricity price target field, output time series field, load time series field, operating status marker field, metering and settlement quantity field, analysis time residual index related to time quality, and reliability level marker from the unified time-series index table and the time-series sample dataset; perform field splitting on each sample record, using electricity value and translation version as target fields, and output, load, operating status, metering quantity, analysis time residual, and time reliability level as candidate feature fields; according to the construction method of features in the time dimension, divide the candidate feature fields into basic time-series features, statistical aggregation features, rate of change features, and cross-time comparison features. Basic time-series features include: features within a single unified analysis time slice or a fixed time slice... The output, load, and historical electricity price values ​​within the long time window are directly input using the original measurements or smoothed time-series values. Statistical aggregation features include: mean, median, extreme values, quantiles, and fluctuation ranges of output, load, and electricity price sequences calculated within preceding and following time windows, centered on or bounded by the unified analysis time, used to characterize short-term and medium-term statistical states. Change rate features include: differences, relative change rates, and multi-step change rates between adjacent unified analysis time slices, used to characterize the speed and direction of change in time-series trends. Cross-time comparison features include: the ratio or difference between the current window statistics and the statistics of a longer-period reference window, used to characterize the degree of deviation of the current state from the historical baseline. The above features are divided and saved into multiple feature dataset views according to feature domains, and the definition, time window length, sliding step size, and construction rule version of each type of feature are recorded in the feature metadata table, forming a reusable feature view library.

[0054] A sample quality assessment mechanism was constructed. For each time-series sample, the sample quality was assessed based on the analysis time residual index, time reliability level, feature missing ratio, and number of outlier markers. The time consistency sensitivity coefficient was multiplied by the sample time consistency deviation to obtain the consistency term. The negative value of the consistency term was then exponentially calculated to obtain the time alignment error term. The missing rate sensitivity coefficient was multiplied by the effective field missing rate of the sample to obtain the missing term. The negative value of the missing term was then exponentially calculated to obtain the feature missing status term. The feature missing status term and the time alignment error term were multiplied to obtain the sample quality assessment value. The specific calculation formula for the sample quality assessment value is as follows: ; In the formula, Indicates the first The sample quality assessment value of each time series sample is used to measure the usability of the sample; The time consistency deviation, which is the normalized maximum residual between the unified analysis time of the sample and the equipment acquisition time, the acquisition node reception time, and the platform storage time, is calculated by taking the absolute value of the residual between the unified analysis time of the sample and the equipment acquisition time, the acquisition node reception time, and the platform storage time, respectively, and then mapping the maximum value to the 0-1 interval through a linear scaling method. The value range is 0-1, which is used to quantify the magnitude of the time alignment error. This represents the effective field missing rate of the sample during feature construction. It is calculated by dividing the number of feature fields that were not successfully extracted from the sample by the total number of feature fields in the current scenario. The value ranges from 0 to 1 and is used to characterize feature completeness. The time consistency sensitivity coefficient is obtained by statistically analyzing the changes in electricity price prediction error under different time residual levels on historical operating data. Its value ranges from 0.5 to 5 and is used to adjust the decay rate of the quality factor due to time deviation. The missing rate sensitivity coefficient is obtained by analyzing the prediction error distribution of samples with different missing rate intervals. Its value ranges from 0.5 to 10 and is used to control the decay rate of the quality factor due to feature missingness. The two exponential terms work together in a multiplicative manner to jointly suppress time alignment error and feature missing rate without using weighted summation, so that the quality factor decays rapidly when either quality factor deteriorates significantly.

[0055] When the sample quality assessment value is lower than the sample quality threshold, the sample is marked as a low-quality sample to be removed or used only in robustness testing scenarios. The specific causes of the low quality of the sample are recorded (such as excessive time consistency deviation or excessively high feature missing rate), and the sample is classified and archived in the low-quality sample library. At the same time, it is marked that it cannot be used in the training and prediction tasks of the regular electricity price dynamic prediction model. When the sample quality assessment value is greater than or equal to the sample quality threshold, the sample is marked as a usable sample and enters the feature selection and dataset partitioning process. Core features are selected, redundant and low-correlation features are removed, and the usable samples are divided into training set, validation set and test set according to the proportion, which are used for electricity price dynamic prediction model training, parameter tuning and performance evaluation, respectively.

[0056] Feature datasets for different prediction tasks are constructed. For short-cycle electricity price dynamic prediction scenarios, short-window basic time-series features, short-term statistical aggregation features, and short-term rate of change features are prioritized, while long-cycle cross-time comparison features are used as enhancement features. For medium- and long-term trend prediction scenarios, long-window statistical aggregation features and cross-time comparison features are added, while high-frequency rate of change features, which are more sensitive to short-term noise, are weakened. During feature selection, the feature set is pruned and grouped based on the feature's sample quality assessment value, correlation statistics with the target variable, and coupling degree with time quality indicators, generating multiple feature subset definitions. Corresponding feature columns are extracted from the feature view library according to the feature subset definitions to construct a structured feature dataset, and the feature set identifier, number of features, feature domain distribution, and construction rule version are recorded in the feature metadata table. The output includes the feature view library, sample quality assessment value dataset, feature dataset views corresponding to different modeling scenarios, the feature dataset, and the target dataset.

[0057] In this implementation plan, by constructing a reusable feature view library based on unified time-series samples and introducing an explicit sample quality assessment mechanism, on the one hand, low-quality samples with large time deviations and heavy feature missingness are eliminated in advance before entering the dynamic electricity price prediction model, so as to avoid interfering with the learning of the dynamic electricity price prediction model. On the other hand, different prediction tasks can be trained and evaluated on a feature dataset with unified caliber, controllable quality and consistent structure, thereby improving the robustness and interpretability of dynamic green electricity price prediction from the feature level.

[0058] Specifically, the process of constructing a dynamic electricity price prediction model and managing the training, validation, and inference results of the feature dataset is as follows: The dynamic electricity price prediction model is constructed by inputting the feature dataset and the target dataset, and time-ordered partitioning and version management are performed; using a unified analysis time field as the time axis, the first-level quality samples are sorted by time and divided into training, validation, and test sets to avoid future time information being leaked into past samples; when constructing the training set, the first-level quality samples within the training period are randomly shuffled and batched to ensure that each training batch has sufficient time coverage and diversity; the time order is preserved when constructing the validation and test sets to evaluate the prediction performance of the dynamic electricity price prediction model under real-time conditions; the dataset version management table records the training set time range, validation set time range, test set time range, feature set identifier, number of samples, and average sample quality evaluation value for each round of model training, achieving a one-to-one mapping between the training data version and the dynamic electricity price prediction model version.

[0059] A time-series deep network was used to train and evaluate the performance of a dynamic electricity price prediction model. During the training phase, the feature dataset was used as input, and the target electricity price sequence was used as the supervision signal. Multiple rounds of iterative training were conducted using a network structure and loss function adapted to the time-series data. During training, error and stability metrics were periodically calculated on the validation set. The error metrics included mean absolute error (MAE), mean squared error (MSE), and quantile error. MAE was calculated by averaging the absolute values ​​of the differences between the predicted and actual electricity prices for all samples within the same evaluation batch, used to measure the overall deviation level. MSE was calculated by averaging the squares of the differences between the predicted and actual electricity prices, amplifying the impact of large deviation samples on the results and identifying severe prediction distortions. Quantile error was calculated by sorting the absolute error values ​​within the evaluation batch from smallest to largest and reading the error values ​​at the 90th or 95th percentile, used to characterize the extreme risk at the tail of the error distribution. The stability metric was calculated based on the sequence of error changes over training rounds. Within a fixed-length training epoch sliding window, the standard deviation of the mean absolute error of the validation set and the average absolute value of the difference between two adjacent evaluation mean absolute errors are statistically analyzed. The smaller the standard deviation and the closer the difference is to zero, the smoother the convergence process of the dynamic electricity price prediction model, and the less likely the training will oscillate or degenerate. When the stability index remains within a stable threshold and the error index no longer decreases within several consecutive windows, the dynamic electricity price prediction model is considered to be close to convergence, which helps avoid overfitting and excessively long training. When the performance of the validation set meets the preset convergence conditions, a one-time evaluation is performed on the test set to obtain the generalization performance on data that did not participate in training and parameter tuning. The structural configuration, loss convergence curve, validation set error trajectory, stability index time series, and test set evaluation results of each training epoch of the dynamic electricity price prediction model are written into the dynamic electricity price prediction model training record table. A unique dynamic electricity price prediction model version number is generated for each model that passes the test and stored in the dynamic electricity price prediction model version repository to support version calls and rollbacks.

[0060] A dynamic electricity price prediction inference pipeline is constructed. This pipeline performs streaming inference on the feature dataset and manages the prediction results data to obtain the prediction result dataset. Data from the latest time period is acquired, and a feature dataset for inference is generated from the feature view library using the same feature set identifier as the training phase. A real-time feature sequence is constructed using the unified analysis time as an index. Based on the prediction frequency and prediction window length, a specified version of the dynamic electricity price prediction model is invoked at each prediction trigger time point to infer the electricity value for several future time slices, outputting the predicted electricity price time series. The prediction results, along with the corresponding unified analysis time, prediction time range, dynamic electricity price prediction model version number, feature set identifier, and sample quality statistics, are written into the prediction result dataset. An index is created in the prediction result index table with the unified analysis time and dynamic electricity price prediction model version number as the primary key, facilitating retrieval and comparative analysis.

[0061] Error assessment and operational monitoring are performed on the prediction result dataset. After the actual settlement results or posterior electricity price results arrive, the predicted electricity price time series is matched with the actual electricity price time series using a unified analysis time as the alignment benchmark. The error distribution under different prediction lead times is statistically analyzed to form a prediction error time series and distribution characteristics. The mean error, error fluctuation, and frequency of extreme errors for different versions of the dynamic electricity price prediction model are statistically analyzed for each time window. Time periods with increased errors are marked, and the reasons for the increased errors are determined by combining the statistical values ​​of the sample quality assessment values ​​for the corresponding time periods. When the error index of a certain version of the dynamic electricity price prediction model deteriorates beyond the deterioration threshold in multiple consecutive time windows, a performance degradation flag is added to the dynamic electricity price prediction model version management table, triggering a retraining or dynamic electricity price prediction model version switching process. The prediction results of different versions of the dynamic electricity price prediction model are stored and compared in parallel. Multiple versions of predicted values ​​and the final actual values ​​at the same time point can be queried according to a unified analysis time, providing data support for the version iteration and strategy selection of the dynamic electricity price prediction system. Outputs include a dynamic electricity price prediction model training record table, a model version library, a prediction result dataset, and a prediction error assessment dataset.

[0062] This implementation plan establishes a training and inference management mechanism for the dynamic electricity price prediction model based on a unified analysis timeline and sample quality labels. This avoids oscillations and overfitting in the time-series data, ensuring the model is traceable and rollbackable. It achieves a complete closed loop from training data management, model training and evaluation, online inference to error monitoring and version iteration, significantly improving the stability, controllability, and evolution capability of the green electricity price dynamic prediction model in engineering applications.

[0063] Specifically, the process of storing and managing the data lineage of the electricity price dynamic forecasting model's prediction results is as follows: Figure 5The flowchart shown illustrates the management and data lineage of electricity price forecast results based on unified analysis time. It inputs the forecast result dataset and a unified time series index table, aligns each forecast result with the unified analysis time, and completes metadata to construct a forecast result record for dynamic electricity price forecasting. From the forecast result dataset, it reads the forecast time range, forecast electricity value sequence, dynamic electricity price forecasting model version identifier, feature set identifier, and forecast generation time. It aligns each future time slice within the forecast time range with the unified analysis time axis to obtain the corresponding unified analysis time. Using the unified time series index table, it queries the data time window definition, participating sample versions, and data source type distribution near that time slice using the unified analysis time as the key. It writes the used data time window, sample version identifier, and data source type into the forecast result metadata. During the result writing process, a unique forecast record identifier is generated for each forecast result, along with the forecast generation time, unified analysis time, forecast step size, dynamic electricity price forecasting model version identifier, feature set identifier, and sample quality statistics, ensuring that each forecast value can be accurately traced back to the dataset and dynamic electricity price forecasting model configuration used.

[0064] The prediction results are linked by identifiers to construct a prediction results data table and a data lineage view. In the prediction results data table, the prediction record identifier is used as the primary key to uniformly store the unified analysis time, predicted electricity value, prediction step size, electricity price dynamic prediction model version identifier, feature set identifier, description of the data time window used, sample version identifier, data source type composition, average sample quality assessment value, and prediction generation time fields, achieving a one-to-one structural association between the prediction results and the input samples and the electricity price dynamic prediction model configuration. In the data lineage view, the records in the prediction results data table are aggregated using the electricity price dynamic prediction model version identifier, feature set identifier, and sample version identifier as dimensions to generate a multi-dimensional mapping relationship. Within the lineage view, a unified analysis time range, data source type composition ratio, and time quality index distribution are provided for each combination of records participating in the prediction, used for cross-analysis of the relationship between error changes and data quality changes during the effect evaluation phase.

[0065] A versioned management and rapid retrieval mechanism for prediction results is constructed, supporting multi-dimensional queries by time, model version, and data caliber. A composite index with the unified analysis time and the electricity price dynamic prediction model version identifier as the primary key, and an auxiliary index with the unified analysis time and prediction step size as secondary keys are established on the prediction result data table. A large number of prediction result records are stored in segments using physical partitioning by date or time window. In the version management table, the effective time, expiration time, corresponding feature set identifier, and training data time range are recorded for each electricity price dynamic prediction model version used online. The system can automatically determine which electricity price dynamic prediction model version a prediction result belongs to at a given moment based on the unified analysis time. By jointly querying the prediction result data table, version management table, and data lineage view, it is possible to query all prediction results of a specific version of the electricity price dynamic prediction model within a specified time period by time interval, or compare the prediction error distribution of different time periods by electricity price dynamic prediction model version. The prediction result data table, prediction result data lineage view, and prediction version management table are output.

[0066] This implementation plan, by structurally storing and managing the data lineage of the dynamic electricity price forecasting model output on a unified analysis timeline, not only constructs a forecast result data table and a forecast version management table that support multi-dimensional retrieval by time, model version, and data caliber, but also fully depicts the mapping link through a data lineage view. This enables rapid tracing of the data caliber and model configuration upon which any forecast result depends during error assessment, version comparison, strategy optimization, and compliance auditing. This significantly improves the traceability, interpretability, and controllability of the dynamic electricity price forecasting results, providing unified, standardized, and fine-grained data support for the online operation monitoring, version iteration, and effect review of the green electricity trading dynamic electricity price forecasting system.

[0067] Specifically, the process of evaluating the prediction effect and forming a closed-loop feedback for error and stability is as follows: Input the prediction result data table and the actual electricity price result data table; perform unified analysis time alignment and error calculation on the prediction results and actual results to construct an error time series and error distribution dataset; read the unified analysis time, predicted electricity value, and prediction step size from the prediction result data table; calculate the target actual time point corresponding to the predicted value using the unified analysis time and prediction step size; read the unified analysis time and actual electricity value from the actual electricity price result data table, match and align them with the target actual time point, and calculate the prediction for records that can be matched one-to-one. Error values ​​are used to form an error time series. The error time series includes point errors, absolute errors, and relative errors at each unified analysis time. The power price dynamic forecasting model version identifier, feature set identifier, sample version identifier, and sample quality statistics are recorded in the error metadata. For forecast records that cannot be matched with actual values, they are marked as pending completion. Error calculation and status updates are performed after the actual values ​​arrive. At the same time, the error time series are grouped according to the forecast step size and time window. The mean, median, quantile, and frequency of extreme errors at different forecast lead times and different time periods are statistically analyzed to form an error distribution dataset.

[0068] A predictive stability assessment mechanism is constructed to provide a fine-grained evaluation of performance over time. The ratio of the absolute value of the prediction error at the current time to the reference value of the error magnitude is raised to a power to obtain the error magnitude term. The ratio of the absolute value of the difference in prediction errors between adjacent time points to the reference value of the error jump is raised to a power to obtain the error jump term. The error magnitude term and the error jump term are multiplied together to obtain the error term. The error term is then multiplied by the negative value of the stability sensitivity coefficient to obtain the error stability term. An exponential operation is performed on the error stability term to obtain the predictive stability index at the k-th unified analysis time point. The specific calculation formula for the predictive stability index is as follows: ; In the formula, Indicates the first A unified analysis time point prediction stability index is used to simultaneously measure the magnitude of the prediction error at a given time and the stability of the error relative to the previous time. This represents the difference between the predicted electricity value and the actual electricity value at that moment. It is obtained by subtracting the actual electricity value under the same unified analysis time from the predicted electricity value at that moment, and reflects the error magnitude. This represents the prediction error at the previous unified analysis time point, calculated by analyzing the k-th time point. What is the result of subtracting the actual electricity value from the predicted electricity value at a single, unified analysis point in time used for? This indicates the magnitude of the error jump between adjacent time slices, calculated... and The absolute value of the difference is used to measure the severity of prediction error fluctuations over a short period of time. The reference value for the error magnitude is obtained by summarizing the absolute values ​​of the errors at all unified analysis time points in historical operational data, sorting them from smallest to largest, and then taking the quantile (e.g., the 95th percentile). It is used to normalize the error magnitude under different scale scenarios to a dimensionless space. The reference value for error jump is obtained by summarizing the error jump amplitudes at all unified analysis time points in historical operating data, sorting them from smallest to largest, and taking the percentile. It is used to normalize the error jump amplitudes under different fluctuation levels. The power exponent of the error magnitude normalization term is used to compare the prediction stability index under different combinations by sweeping different exponent combinations over historical operating data. The value range is a real number greater than zero, and it is used to adjust the relative sensitivity of the error magnitude in the geometric mean. The power exponent representing the error jump normalization term is determined by sweeping different exponent combinations across historical operating data and the consistency with the actual business stability evaluation results. Its value range is a real number greater than zero, and it is used to adjust the relative sensitivity of error jumps in the geometric mean. The stability suppression coefficient is obtained by statistically analyzing the acceptable predictable levels corresponding to different stability index value ranges in historical operating data and fitting the relationship between the stability index distribution and the business acceptable threshold. The value range is a real number greater than zero, which is used to control the rate at which the stability index decays when the geometric mean normalization error is larger.

[0069] A prediction performance evaluation view is constructed by linking prediction error, prediction stability index, data quality index, and sample quality assessment value. By jointly querying the error distribution dataset, unified time-series index table, and sample quality assessment value dataset, using a unified analysis time as the key, the error value and prediction stability index at each time point are linked to the corresponding data quality index for that time period. Data quality indicators include time residual distribution, time reliability level distribution, data source type composition, and sample quality assessment value. This information is then summarized in the prediction performance evaluation view. Within this view, errors and prediction stability indices are grouped and statistically analyzed according to the electricity price dynamic prediction model version, prediction step size, and time interval, revealing differences in performance across different electricity price dynamic prediction model versions with varying prediction lead times. Furthermore, for time intervals where errors increase or stability indices are low, changes in time deviation estimation parameters, time delay structure consistency deviation index, and data source quality index are extracted to determine whether error degradation is highly correlated with decreased time correction accuracy, anomalies in a particular data source, or a decline in sample quality. The change in time deviation estimation parameters refers to the time series of estimated time deviation values ​​for device nodes, acquisition nodes, and platform nodes extracted from the time deviation estimation result dataset. It calculates the deviation of each node's estimated deviation value relative to its historical baseline deviation and long-term mean, the deviation variance, and the deviation jump magnitude between adjacent timing sessions. This is used to characterize whether the time correction intensity of each node increases and whether abnormal time deviation drift occurs within the time interval. The delay structure consistency deviation index refers to comparing the current link delay estimation parameters (including the mean delay of the first link segment, the mean delay of the second link segment, and the corresponding delay standard deviation) for the same data source type and the same transmission path. The dimensionless deviation value, calculated using parameter difference normalization, parameter vector distance, and time delay distribution shape difference measurement, is used to measure whether the link delay structure has changed significantly within the current time interval. The data source quality index refers to the statistical distribution of the time reliability level of data records participating in the prediction within the selected time interval, the unified analysis of the mean and extreme values ​​of time residuals, the proportion of missing records, the proportion of low-quality samples, and the proportion of abnormal marked channels in the data source type composition, based on the unified time series index table and sample quality assessment value dataset. It is used to quantitatively describe the data integrity, time reliability, and sample quality level within the time interval, and to support targeted attribution analysis of the causes of prediction error degradation.

[0070] A closed-loop feedback mechanism is constructed to write the evaluation conclusions back to the electricity price dynamic prediction model version management and data processing chain. When the error index and stability index of a certain electricity price dynamic prediction model version consistently fail to meet the standards within a specific time interval, a performance degradation mark is added to the version in the electricity price dynamic prediction model version management table, and the degradation occurrence time, degradation duration, and related indicators with data quality are recorded to trigger retraining, feature set adjustment, or electricity price dynamic prediction model version switching processes. When the evaluation results indicate that the error degradation is mainly concentrated in a specific data source or a specific time deviation exceeding the limit, corresponding diagnostic information is written to the configuration table of the time correction and sample construction module, prompting adjustments to the acquisition parameters, time delay estimation parameters, or sample selection conditions of the corresponding data source. For time periods marked in the evaluation view as having prediction stability indices below the stability threshold, the unified analysis time range of the time period is recorded in the key inspection time window table. The output includes a prediction error time series dataset and a prediction stability index dataset, a prediction performance evaluation view, and a key inspection time window table.

[0071] This implementation plan constructs a closed-loop feedback mechanism by performing error calculation, stability quantification, and data quality linkage analysis on the prediction results and actual electricity price results on a unified analysis timeline, and writing the evaluation conclusions back to the model version management and data processing link. This mechanism can not only accurately locate the correlation between error degradation and time correction deviation, abnormal link delay structure, and data source quality degradation, and promptly trigger model retraining, feature and sample screening strategy optimization, and key review of high-risk time periods, but also significantly improve the accuracy, stability, interpretability, and operational controllability of green electricity trading price dynamic prediction during long-term online operation.

[0072] Specifically, the second aspect of this invention provides an AI-based dynamic prediction system for green electricity trading prices, applied to an AI-based dynamic prediction method for green electricity trading prices. It includes: a multi-source data acquisition module, used to collaboratively acquire green electricity price datasets through a dispatch automation system, a metering acquisition system, a field acquisition gateway, a market trading platform, and a settlement management platform; performing field normalization and dimensionless normalization on data from different sources; and completing multi-level timestamp correction based on a unified time synchronization service to generate standardized acquisition data frames carrying a unified analysis time; a latency estimation module, used to statistically analyze the link latency distribution and time deviation characteristics of each data source and transmission path based on the standardized acquisition data frames; constructing a link latency estimation model and a time deviation evaluation model; forming a data source-level latency estimation parameter set; calculating a unified analysis time; and generating a unified time series view for dynamic electricity price prediction; and a time series construction module, used to, based on the unified time series view, analyze the standardized acquisition data frames and associated electricity prices. Price result data, output data, and load data are sorted and stitched according to a unified analysis time. The time axis is segmented according to the configured window length and sliding step size to construct time series samples and generate a time series sample view for dynamic electricity price prediction. The electricity price prediction model inference module is used to extract feature datasets based on the time series sample view and the data source-level time delay estimation parameter set, to perform quality assessment and hierarchical screening of samples, to call the artificial intelligence training engine to complete the training and verification of the dynamic electricity price prediction model, and to perform batch or online inference on the newly input feature datasets to output prediction results. The model evaluation module is used to perform unified analysis time alignment and error calculation between the prediction results and the actual electricity price results. The prediction results, error indicators, prediction stability index, time correction, and sample quality information are written into the prediction effect evaluation view to realize the storage of prediction results and data lineage management. The closed-loop evaluation feedback of error and stability guides the update of time delay parameters, sample screening, and model retraining.

[0073] In this implementation plan, based on unified time synchronization and multi-level timestamp correction, multi-source data acquisition, link delay and time deviation modeling, unified time series sample construction, electricity price prediction model inference, and error and stability closed-loop evaluation are organically integrated. This achieves high-quality data reconstruction and traceable modeling of green electricity trading prices under a unified time benchmark, effectively improving the accuracy, stability, interpretability, and controllability of dynamic electricity price prediction and online operation.

[0074] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0075] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An AI-based dynamic prediction method for green electricity transaction price, characterized in that, The method comprises the following steps: S1, collecting green electricity price data set, normalizing and aligning the green electricity price data set to uniform time base to form a standardized collection data frame; S2, constructing a link delay estimation model and a time deviation evaluation model based on the standardized collection data frame to form a data source level delay estimation parameter set, and constructing a unified time sequence view for electricity price dynamic prediction; S3, performing data splicing and time sequence sample construction on the standardized collection data frame to generate a time sequence sample view for electricity price dynamic prediction; S4, constructing a feature data set based on the data source level delay estimation parameter set and performing sample quality evaluation, constructing an electricity price dynamic prediction model, training and verifying the feature data set, and managing the inference results; S5, storing and managing the prediction results of the electricity price dynamic prediction model, and evaluating the prediction effect of the error and stability to form a closed loop feedback.

2. The AI-based green electricity transaction electricity price dynamic prediction method according to claim 1, characterized in that: The specific process of collecting the green electricity price data set, normalizing and aligning the green electricity price data set to uniform time base to form a standardized collection data frame is as follows: The green electricity price data set includes: transaction price and business result data set, actual price result data table, dispatching and output data set, load and monitoring measurement data set, metering and settlement basis data set, operation state and alarm data set, time service and clock state data set, and time stamp and channel identification data set; the historical multi-source data sequence stored in the same green electricity business scenario as the historical operation data, the data source type label and time attribute classification management of the collected green electricity price data set are performed based on the collection node receiving time and platform storage time to obtain the processed green electricity price data set with multi-level time stamp after preliminary time alignment; The processed green electricity price data set is encapsulated in a unified data structure to construct a green electricity price collection data frame: the processed green electricity price data set is arranged and cached through the standardized collection data frame format, the global queue sorted by time and the local queue according to the data source type, and the time class field in the collection data frame is subjected to z-score non-dimensional normalization preprocessing.

3. The AI-based green electricity transaction electricity price dynamic prediction method of claim 1, wherein: The specific process of constructing a link delay estimation model and a time deviation evaluation model based on the standardized collection data frame to form a data source level delay estimation parameter set is as follows: The multi-level time stamp field in the green electricity price collection data frame and the time service and clock state data set are input to perform joint modeling sample extraction to obtain a delay modeling basis data set, the delay modeling basis data set is grouped, cleaned and abnormally identified according to data source type, node combination, transmission path and time period to construct a link delay estimation model and a time deviation evaluation model; In the time bias evaluation process, the sign of the time bias value directly observed by the node in the current time service session is taken as the bias sign, thereby preserving the positive and negative directions of the bias; the absolute value of the observed time bias value is obtained, the absolute value of the difference between the current observed bias and the bias baseline obtained in the previous time window is the jump amplitude of the current observed bias relative to the bias baseline, and the bias jump sensitivity coefficient multiplied by the jump amplitude of the current observed bias relative to the bias baseline is the bias jump term; the larger one between the absolute value of the observed time bias value and the bias jump term is selected as the effective bias scale; the bias sign is multiplied by the effective bias scale to obtain the time bias estimation value of the node at the current time service moment; The obtained time bias estimation value is compared with the time bias threshold in real time, when the time bias estimation value is greater than the time bias threshold, the current node time state is marked as a bias overrun state, when the time bias estimation value is less than or equal to the time bias threshold, the current node time state is marked as a bias normal state; Output the data source level time delay estimation parameter set and the time bias estimation result data set.

4. The AI-based green electricity transaction electricity price dynamic prediction method of claim 1, characterized in that: The specific process of constructing the unified time sequence view for the dynamic prediction of electricity price is: Input the green electricity price collection data frame and the data source level time delay estimation parameter set, perform multi-level timestamp correction and unified analysis time; for data records with incomplete time fields or abnormal clock states, use the data source level time delay estimation parameter set and the residual time relationship to analyze the time estimation and reliability marking; The obtained unified analysis time field is written into the unified time sequence index table together with the original multi-level timestamp, time precision mark, time source mark and clock state mark time metadata; an incremental update and batch recalculation mechanism is provided, and the unified time sequence index table, the analysis time and the original time relationship data set and the analysis time reliability statistical data set are output.

5. The AI-based green electricity transaction electricity price dynamic prediction method according to claim 1, characterized in that: The specific process of performing data splicing and time sequence sample construction on the standardized collection data frame to generate the time sequence sample view for the dynamic prediction of electricity price is: Input the unified time sequence index table and the business data table of the electricity price result data, output data, load data, running state data and metering settlement data in the green electricity price collection data frame, perform unified analysis time alignment and sample reorganization; Sort and group the green electricity price collection data frame on the unified analysis time axis to construct the basic time sequence sequence view according to the data source and the region; Perform time window division and sample organization on the basic time sequence sequence view to construct the time sequence sample data set; Use the analysis time reliability level and the analysis time residual index in the unified time sequence index table to mark the quality and divide the level of the constructed time sequence sample, form the sample quality label system for prediction; based on the time sequence sample data set, construct the time sequence sample view adapted to different prediction tasks; Output the multi-source basic time sequence sequence view, the time sequence sample data set and various time sequence sample views.

6. The AI-based green electricity transaction electricity price dynamic prediction method according to claim 1, characterized in that: The specific process of constructing the feature data set based on the data source level time delay estimation parameter set and performing sample quality evaluation is: The input time sequence sample data set is subjected to feature field and target field extraction to construct a feature data set view library for electricity price dynamic prediction, and a sample quality evaluation mechanism is constructed to evaluate the quality of each time sequence sample according to the analysis time residual index, time reliability level, feature missing rate and number of abnormal markers; the time consistency sensitive coefficient is multiplied by the sample time consistency deviation to obtain a consistency term, and the negative value of the consistency term is subjected to exponential operation to obtain a time alignment error term; the missing rate sensitive coefficient is multiplied by the sample effective field missing rate to obtain a missing term, and the negative value of the missing term is subjected to exponential operation to obtain a feature missing condition term; the feature missing condition term and the time alignment error term are multiplied to obtain a sample quality evaluation value; when the sample quality evaluation value is lower than a sample quality threshold, the sample is marked as a low-quality sample for elimination or only used in a robustness test scenario, and when the sample quality evaluation value is greater than or equal to the sample quality threshold, the sample is marked as an available sample and enters a feature screening and data set division process; a feature data set for different prediction tasks is constructed; a feature view library, a sample quality evaluation value data set, a feature data set view corresponding to different modeling scenarios, a feature data set and a target data set are output.

7. The AI-based green electricity transaction electricity price dynamic prediction method according to claim 1, characterized in that: The specific process of the electricity price dynamic prediction model for training, verification and inference result management is as follows: The feature data set and the target data set are input to construct an electricity price dynamic prediction model, time-ordered division and version management are performed, a time sequence deep network is called to complete training and performance evaluation of the electricity price dynamic prediction model, an electricity price dynamic prediction inference pipeline is constructed to perform streaming inference on the feature data set and manage prediction result data to obtain a prediction result data set, and error evaluation and operation monitoring are performed on the prediction result data set to form a prediction error time sequence and distribution characteristics; An electricity price dynamic prediction model training record table, a model version library, a prediction result data set and a prediction error evaluation data set are output.

8. The AI-based green electricity transaction electricity price dynamic prediction method according to claim 1, characterized in that: The specific process of storing and data blood relationship management of the prediction result of the electricity price dynamic prediction model is as follows: The prediction result data set and the unified time sequence index table are input, the unified analysis time alignment and metadata completion are performed on each prediction result to construct a prediction result record for electricity price dynamic prediction, the prediction result records are associated by identification to construct a prediction result data table and a data blood relationship view, a version management and rapid retrieval mechanism of the prediction result is constructed to support multi-dimensional queries according to time, model version and data caliber, and a prediction result data table, a prediction result data blood relationship view and a prediction version management table are output. 9.The AI-based green electricity transaction electricity price dynamic prediction method of claim 1, wherein: The specific process of forming a closed-loop feedback by predicting the error and stability for effect evaluation is as follows: The input prediction result data table and the actual electricity price result data table are aligned in time and error calculation is performed on the prediction result and the actual result to construct an error time series and an error distribution data set; a prediction stability evaluation mechanism is constructed to perform fine-grained evaluation on the performance on the time axis; the absolute value of the prediction error at the current time and the error amplitude reference value are subjected to power operation to obtain an error amplitude term; the absolute value of the prediction error difference between adjacent time points and the error jump reference value are subjected to power operation to obtain an error jump term; the error amplitude term and the error jump term are multiplied to obtain an error term, and the error term is multiplied by the negative value of the stability sensitivity coefficient to obtain an error stability term; The error stability term is subjected to exponential operation to obtain a prediction stability index at the kth unified analysis time point; the prediction error, the prediction stability index, the data quality index, and the sample quality evaluation value are associated to construct a prediction effect evaluation view; and a prediction error time series data set and a prediction stability index data set are output.

10. An AI-based dynamic green electricity trading price forecasting system, applying the AI-based dynamic green electricity trading price forecasting method according to any one of claims 1-9, characterized in that, Comprise: A multi-source data acquisition module for cooperatively collecting green electricity price data sets through a multi-source system, normalizing the green electricity price data sets, and performing unified time base alignment processing to form a standardized acquisition data frame; A time delay estimation module for constructing a link time delay estimation model and a time deviation evaluation model based on the standardized acquisition data frame, forming a data source level time delay estimation parameter set, and constructing a unified time sequence view for electricity price dynamic prediction; A time sequence construction module for performing data splicing and time sequence sample construction on the standardized acquisition data frame to generate a time sequence sample view for electricity price dynamic prediction; An electricity price prediction model inference module for constructing a feature data set based on the data source level time delay estimation parameter set and performing sample quality evaluation, constructing an electricity price dynamic prediction model, training and verifying the feature data set, and managing inference results; A model evaluation module for storing the prediction results of the electricity price dynamic prediction model and managing data blood relationship, evaluating the prediction effect of the error and stability, and forming a closed loop feedback.

Images