Power source-load-price linkage prediction method and system based on spatiotemporal correlation modeling

Abstract

The application discloses a power source-load-price linkage prediction method and system based on space-time correlation modeling, and relates to the technical field of power prediction. The method acquires power operation and transaction multi-source data, aligns according to a preset transaction time granularity, and constructs a heterogeneous space-time sample containing source-side nodes, load-side nodes, price nodes, meteorological nodes and rule nodes; generates a causality availability label according to the availability of data at a prediction time and generates rule constraint information according to contract and rule data; determines a candidate correlation relationship with a direction identifier and a time lag amount based on physical correlation, meteorological propagation correlation, historical statistical correlation and constraint correlation, generates a dynamic sparse correlation graph in combination with the causality availability label and the rule constraint information; extracts node space-time hidden states and performs multi-task probability prediction to generate source-side, load-side, price and bias cost probability prediction results, and forms a machine-readable power transaction prediction result package. The application improves data consistency and result integrity in source-load-price linkage prediction.

Description

Technical Field

[0001] This invention relates to the field of power forecasting technology, and in particular to a method and system for forecasting power source-load price linkage based on spatiotemporal correlation modeling. Background Technology

[0002] With the increasing proportion of renewable energy integration and the continuous improvement of the electricity market trading mechanism, the coupling relationship between source-side output, load-side load, electricity price changes, meteorological conditions, and trading constraints in the power operation process is becoming increasingly stronger. Existing power forecasting systems typically model renewable energy output, load, or electricity prices separately, or make forecasts based solely on fixed time series, fixed topological relationships, or historical statistical relationships, which makes it difficult to fully reflect the linkage relationships between different power objects that arise from changes in time, space, and market operating conditions.

[0003] In real-world electricity trading forecasting scenarios, data from different sources exhibit inconsistencies in collection frequency, release time, applicable regions, metering standards, and version rules. Some data is only available after the forecast time, and some is also affected by daily updates, post-confirmation, or rule version changes. If the forecasting model fails to differentiate data availability at the forecast time or fails to standardize different data standards, inconsistencies in input conditions between the training and inference phases can easily arise, impacting the reliability and usability of the forecast results.

[0004] Electricity trading forecasts typically need to simultaneously serve multiple forecasting tasks, including available power output on the source side, load on the load side, electricity prices, and deviation costs. Existing methods often focus on forecasting a single indicator or treat trading constraints and deviation cost calculations as independent post-forecast processing steps, resulting in insufficient alignment between the forecasting model and actual trading standards, deviation calculation standards, and output requirements. Therefore, there is a need for a forecasting method and system that can uniformly process multi-source data, spatiotemporal correlations, data availability, and trading standards for electricity source-load-price linkage scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting power source-load price linkage based on spatiotemporal correlation modeling, in order to solve the problems that existing power prediction methods are difficult to uniformly model the available output on the source side, the load on the load side, the electricity price and related transaction constraints, and that the availability of prediction results is easily affected when the availability of multi-source data release time, data caliber, rule version and prediction time is inconsistent. This invention enables the prediction process to form continuous, unified and implementable prediction results for power source-load price linkage scenarios.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] On the one hand, this invention provides a method for predicting the linkage between power source and load prices based on spatiotemporal correlation modeling, including the following steps:

[0008] Step S1: Obtain multi-source data on power operation and trading, including available output data on the source side, load data on the load side, electricity price data, meteorological data, and contract and rule data;

[0009] Step S2: Time-align the power operation and transaction multi-source data according to the preset transaction time granularity to construct a heterogeneous spatiotemporal sample including source-side nodes, load-side nodes, electricity price nodes, meteorological nodes and rule nodes;

[0010] Step S3: Based on the data availability at the prediction time, generate causal availability tags for the node feature data in the heterogeneous spatiotemporal samples, and generate rule constraint information based on contract and rule data;

[0011] Step S4: Based on the physical association, meteorological propagation association, historical statistical association and constraint association between different nodes in the heterogeneous spatiotemporal sample, candidate association relationships with directional identifiers and time lags are determined according to the association type. The candidate association relationships are constrained and screened based on causal availability labels and machine-readable rule constraint information to generate a set of effective association edges. The set of effective association edges forms a dynamic sparse association graph that changes with the prediction time.

[0012] Step S5: Based on the dynamic sparse association graph, perform spatiotemporal feature extraction on the node feature data in the heterogeneous spatiotemporal samples to obtain the spatiotemporal hidden state of the nodes;

[0013] Step S6: Perform multi-task probability prediction based on the spatiotemporal hidden state of nodes, machine-readable rule constraint information, and causal availability markers to generate source-side available output probability prediction results, load-side load probability prediction results, and electricity price probability prediction results; after data caliber matching, generate deviation cost probability prediction results based on the load-side load probability prediction results, electricity price probability prediction results, predicted baseline electricity volume, and machine-readable rule constraint information; generate a machine-readable electricity trading prediction result package based on the source-side available output probability prediction results, load-side load probability prediction results, electricity price probability prediction results, and deviation cost probability prediction results.

[0014] Preferably, the acquisition of multi-source data on power operation and transactions is performed in the following manner:

[0015] Obtain available power output data from the source side, load data from the load side, electricity price data, meteorological data, and contract and rule data;

[0016] Configure data category identifiers, data source identifiers, timestamp identifiers, market zone identifiers, rule version identifiers, and data availability attributes for source-side available output data, load-side load data, electricity price data, meteorological data, contract and rule data;

[0017] Based on data category identifiers, data source identifiers, timestamp identifiers, market partition identifiers, rule version identifiers, and data availability attributes, multi-source power operation and transaction data with the required time alignment identifiers are generated.

[0018] Preferably, the construction of the heterogeneous spatiotemporal samples includes:

[0019] According to the preset transaction time granularity, the multi-source data of power operation and transaction are time-aligned to obtain node feature data with a unified time index;

[0020] Source-side nodes are generated based on available output data from the source side, load-side nodes are generated based on load data from the load side, electricity price nodes are generated based on electricity price data, meteorological nodes are generated based on meteorological data, and rule nodes are generated based on contract and rule data.

[0021] Configure node type identifiers, forecast time identifiers, market partition identifiers, rule version identifiers, and data availability attributes for source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, and rule nodes to form heterogeneous spatiotemporal samples.

[0022] Preferably, the generation of causal availability tags and the generation of rule constraint information based on contract and rule data specifically include:

[0023] Based on the prediction time, the node characteristic data are divided into data that can be obtained before the prediction time, data that can be obtained on the running day, and data that can be obtained after the settlement.

[0024] Based on the data available before the prediction time and the data available on the operating day, a causal availability label is generated, and the data available for post-settlement is marked as data disabled during the inference phase.

[0025] Extract contract mapping relationships, settlement aggregation criteria, deviation assessment criteria, data availability limitations, effective time and rule version number from contract and rule data to generate rule constraint information;

[0026] During the training phase of the prediction model used to perform multi-task probabilistic prediction, causal availability tags are generated for historical node feature data based on the data availability attributes corresponding to the historical prediction time, consistent with the inference phase. Data released after the historical prediction time and data available for post-settlement are then masked based on the causal availability tags.

[0027] Preferably, the determination of the candidate association includes:

[0028] Physical correlation, meteorological propagation correlation, historical statistical correlation, and rule-based correlation are respectively regarded as different correlation types;

[0029] Dimensionless candidate association factors are generated for different association types, and the credibility of the dimensionless candidate association factors is calibrated according to the data quality status, rule version validity status, and data availability status at the prediction time, forming the composite association strength under the corresponding association type.

[0030] Configure direction identifiers and time lag values ​​for different association types. The time lag values ​​are determined according to the physical propagation relationship, meteorological propagation relationship, historical statistical lag relationship, or rule time mapping relationship of the corresponding association type.

[0031] Preferably, the generation of the dynamic sparse association graph is performed by determining the set of valid association edges according to the following rules:

[0032] For any association type, based on the corresponding composite association strength, direction identifier, time lag, causal availability marker, and rule constraint information, a set of corresponding valid association edges is generated:

[0033] ;

[0034] In the formula, Indicates the predicted time The set of valid associated edges below; Represents the set of candidate associations; and Represents a node; Indicates the association type; Indicates the strength of the composite correlation; Represents a combination of node types In association type The dimensionless correlation strength threshold is below; Indicated by direction identifier Time lag and prediction time Defined association type constraint functions; This indicates the availability status of data at the predicted time, as determined by the causal availability label. This indicates a consistent state of transaction criteria determined by rule-based constraint information; Indicates the predicted time Corresponding rule constraint information A state where the rules are clearly defined as prohibited;

[0035] in, , , and Both are binary state functions;

[0036] When the rule constraint information changes, update the set of valid associated edges at the corresponding prediction time.

[0037] A dynamic sparse association graph is formed based on the set of valid associated edges.

[0038] Preferably, the spatiotemporal feature extraction process includes:

[0039] Based on node type, the node feature data in the heterogeneous spatiotemporal samples are subjected to type-adaptive projection to obtain source-side node representation, load-side node representation, electricity price node representation, meteorological node representation, and regular node representation;

[0040] Based on the effective association edges in the dynamic sparse association graph, graph neighborhood aggregation is performed on the source-side node representation, load-side node representation, electricity price node representation, meteorological node representation, and rule node representation to obtain the graph-constrained node representation;

[0041] The temporal feature encoding of the graph constraint node representation is performed according to the predicted time sequence to obtain the spatiotemporal hidden state of the node.

[0042] The spatiotemporal latent state of a node is generated from the available power output information on the source side, the load information on the load side, the electricity price information, the meteorological propagation information, and the rule constraint information.

[0043] Preferably, the multi-task probability prediction generates the probability prediction interval in the following manner:

[0044] The spatiotemporal hidden states of nodes obtained by the constraints of the dynamic sparse correlation graph are respectively input into the prediction branches corresponding to the source-side available output task, the load-side load task, and the electricity price task.

[0045] Electricity pricing targets include day-ahead electricity pricing targets and real-time electricity pricing targets;

[0046] Each prediction branch generates a probability prediction interval for the corresponding prediction task based on the spatiotemporal latent state of the node:

[0047] ;

[0048] In the formula, Represents a node At the predicted time Next The probability prediction interval for the prediction task; This indicates the type of forecast task, which includes source-side available output tasks, load-side load tasks, and electricity price tasks. Indicates the first The prediction branch corresponding to the prediction task; Represents nodes obtained from dynamic sparse association graph constraints. At the predicted time The spatiotemporal hidden state of the node; Indicates the predicted time Corresponding rule constraint information; This represents the set of data availability states at the predicted time, formed by causal availability labels;

[0049] The probability prediction interval includes the median prediction, upper quantile prediction, and lower quantile prediction for the prediction task. The upper and lower quantile predictions are generated by the non-negative quantile expansion in the target space of the prediction task relative to the median prediction.

[0050] Preferably, when generating the deviation cost probability prediction result and the machine-readable power trading prediction result package, the following is included:

[0051] The electricity price probability prediction results are divided into day-ahead electricity price probability prediction results and real-time electricity price probability prediction results. Based on the power mapping relationship, data aggregation caliber, deviation calculation caliber and constraint version number in the machine-readable rule constraint information, the data caliber matching of the load-side load probability prediction results, day-ahead electricity price probability prediction results, real-time electricity price probability prediction results and the predicted benchmark power is performed under the same preset transaction time granularity.

[0052] Based on the load-side load probability prediction results after caliber matching, the day-ahead electricity price probability prediction results, the real-time electricity price probability prediction results, and the predicted benchmark electricity volume, a source-load price quantile combination set is constructed. Each quantile combination in the source-load price quantile combination set includes a load-side load prediction quantile, a real-time electricity price prediction quantile, and a day-ahead electricity price prediction quantile.

[0053] For each quantile combination in the source-load price quantile combination set, the predicted power deviation is determined based on the difference between the predicted baseline power and the predicted power of the load side corresponding to the quantile combination. The price difference is determined based on the difference between the real-time electricity price prediction value corresponding to the quantile combination and the day-ahead electricity price prediction value. Finally, the candidate value of the deviation cost is determined based on the absolute value of the product of the predicted power deviation and the price difference.

[0054] Based on the candidate values ​​of deviation cost under each quantile combination and the corresponding source-load price linkage weights, the predicted value of deviation cost under the target quantile is generated according to the following formula:

[0055] ;

[0056] In the formula, Indicates market segmentation At the predicted time and deviation cost target quantile The predicted deviation cost value; Indicates market segmentation At the predicted time The following is a set of source-load price quantile combinations; This represents any quantile combination in the set of source-load quantile combinations; Representing quantile combination Candidate values ​​for deviation costs; Representing quantile combination The source-load price linkage weight; This represents a weighted quantile operator, which determines the target quantile of deviation cost based on candidate deviation cost values ​​and corresponding source-load price linkage weights. The predicted deviation cost value;

[0057] The source-load-price linkage weight is a non-negative normalized weight, which is determined by the effective association status between the source-side nodes, load-side nodes and price nodes in the dynamic sparse association graph, machine-readable rule constraint information, and the data availability status at the prediction time.

[0058] Based on the deviation cost prediction values ​​at multiple deviation cost target quantiles, a deviation cost probability prediction result is generated, and a machine-readable power trading prediction result package is generated based on the source-side available output probability prediction result, the load-side load probability prediction result, the electricity price probability prediction result, and the deviation cost probability prediction result.

[0059] On the other hand, the present invention also provides a power source-load price linkage prediction system based on spatiotemporal correlation modeling, which is applied to the power source-load price linkage prediction method based on spatiotemporal correlation modeling as described above, including:

[0060] The data acquisition module is used to acquire multi-source data on power operation and trading. This multi-source data includes available power output data on the source side, load data on the load side, electricity price data, meteorological data, and contract and rule data.

[0061] The heterogeneous spatiotemporal sample construction module is used to time-align multi-source data of power operation and trading according to the preset transaction time granularity, and construct heterogeneous spatiotemporal samples including source-side nodes, load-side nodes, electricity price nodes, meteorological nodes and rule nodes.

[0062] The constraint information generation module is used to generate causal availability tags for node feature data in heterogeneous spatiotemporal samples based on the data availability at the prediction time, and to generate machine-readable rule constraint information based on contract and rule data.

[0063] The dynamic sparse association graph generation module is used to determine candidate association relationships with directional identifiers and time lags based on physical associations, meteorological propagation associations, historical statistical associations and constraint associations between different nodes in heterogeneous spatiotemporal samples. It also performs constraint screening on the candidate association relationships based on causal availability labels and machine-readable rule constraint information, generates a set of effective association edges, and forms a dynamic sparse association graph that changes with the prediction time from the set of effective association edges.

[0064] The spatiotemporal feature extraction module is used to extract spatiotemporal features from node feature data in heterogeneous spatiotemporal samples according to the dynamic sparse association graph, and obtain the spatiotemporal hidden state of the node.

[0065] The prediction result generation module is used to perform multi-task probabilistic prediction based on the spatiotemporal latent state of nodes, machine-readable rule constraint information, and causal availability markers, generating source-side available output probability prediction results, load-side load probability prediction results, and electricity price probability prediction results. After data caliber matching, it generates deviation cost probability prediction results based on the load-side load probability prediction results, electricity price probability prediction results, predicted baseline electricity volume, and machine-readable rule constraint information. Finally, it generates a machine-readable electricity trading prediction result package based on the source-side available output probability prediction results, load-side load probability prediction results, electricity price probability prediction results, and deviation cost probability prediction results.

[0066] The beneficial effects of this invention are as follows: This invention unifies the available power output data on the source side, load data on the load side, electricity price data, meteorological data, and contract and rule data into a preset transaction time granularity for processing, and constructs a heterogeneous spatiotemporal sample for source-load-price linkage prediction. This can reduce the impact of inconsistent time granularity, market partitioning, and data caliber of multi-source data on prediction modeling, and enable source-side, load-side, electricity price, meteorological, and rule-related data to participate in subsequent prediction calculations on a unified data basis.

[0067] This invention generates causal availability markers based on the data availability at the prediction time, and combines rule constraint information to constrain and filter candidate associations, enabling the dynamic sparse association graph to be updated as the prediction time, data availability status, and rule conditions change. This reduces the risk of data after the prediction time, post-confirmation data, and inconsistent data participating in modeling, and improves data consistency between the training process and the actual prediction process.

[0068] This invention extracts spatiotemporal features based on dynamic sparse correlation graphs and combines node spatiotemporal hidden states, rule constraint information, and causal availability labels for multi-task probability prediction. It can generate prediction results related to source-side available output, load-side load, electricity price, and deviation cost within the same prediction framework. At the same time, it generates deviation cost probability prediction results through data caliber matching, enabling machine-readable power trading prediction result packages to reflect the source-load-price linkage and deviation cost changes, thereby improving the data consistency, output completeness, and engineering feasibility of the prediction results. Attached Figure Description

[0069] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the modular structure of the system of the present invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0071] like Figure 1 As shown, in one embodiment of the present invention, a method for predicting power source-load price linkage based on spatiotemporal correlation modeling is provided. This method can be executed by a power trading prediction system. The power trading prediction system includes a data acquisition module, a heterogeneous spatiotemporal sample construction module, a causal availability and rule constraint generation module, a candidate correlation determination module, a dynamic sparse correlation graph generation module, a spatiotemporal feature extraction module, a multi-task probability prediction module, and a prediction result generation module. These modules are connected sequentially, with the output of one module serving as the input of the next, thus forming a complete processing flow from multi-source data acquisition, spatiotemporal sample construction, dynamic correlation modeling, probability prediction to result package generation.

[0072] Step S1: Obtain multi-source data on power operation and trading.

[0073] In this embodiment, step S1 is executed by the data acquisition module of the power trading forecasting system. The data acquisition module acquires multi-source power operation and trading data from the source-side operation system, load metering system, power market data system, meteorological data system, and contract and rule management system. The multi-source power operation and trading data includes source-side available output data, load-side data, electricity price data, meteorological data, and contract and rule data. Among them, source-side available output data is used to describe the power generation capacity of the source-side object at the forecast time; load-side load data is used to describe the electricity demand of the load-side object at the forecast time; electricity price data is used to describe the price changes of the market price object at the trading time granularity; meteorological data is used to describe the environmental conditions affecting the changes in source-side output and load-side load; and contract and rule data is used to describe the power mapping relationship, data aggregation scope, deviation calculation scope, rule effective time, and rule version information.

[0074] After acquiring the aforementioned data, the data acquisition module configures data category identifiers, data source identifiers, timestamp identifiers, market zone identifiers, rule version identifiers, and data availability attributes for source-side available power output data, load-side load data, electricity price data, meteorological data, and contract and rule data. The data category identifier distinguishes data types; the data source identifier records the business system, interface, data table, or data file from which the data originates; the timestamp identifier records the data collection time, release time, and applicable time, where the collection time is used for data traceability, the release time is used to determine data availability, and the applicable time is used for subsequent time alignment; the market zone identifier records the applicable electricity market zone; the rule version identifier records the applicable contract and rule version; and the data availability attribute records the readable status of the data relative to the prediction time and serves as the basis for generating causal availability markers.

[0075] Through the above processing, step S1 generates multi-source power operation and transaction data with the identification information required for time alignment. This multi-source power operation and transaction data is a standardized data record that can be called upon in the subsequent heterogeneous spatiotemporal sample construction process, enabling data from different sources, with different time granularities, and different business perspectives to undergo time alignment, node construction, and rule constraint processing based on a unified identifier.

[0076] Step S2: Time-align the power operation and transaction multi-source data according to the preset transaction time granularity to construct a heterogeneous spatiotemporal sample including source-side nodes, load-side nodes, electricity price nodes, meteorological nodes and rule nodes;

[0077] In this embodiment, step S2 is executed by the heterogeneous spatiotemporal sample construction module. The heterogeneous spatiotemporal sample construction module receives the multi-source power operation and trading data output from step S1, and performs time alignment on data from different sources, with different sampling frequencies, and different applicable times according to a preset trading time granularity. The preset trading time granularity can be determined based on power trading rules, forecasting task requirements, or data release frequency, for example, by day, by hour, or by more granular trading periods.

[0078] Specifically, the heterogeneous spatiotemporal sample construction module maps available power output data from the source side, load data from the load side, electricity price data, meteorological data, and contract and rule data to the corresponding prediction time based on the timestamp identifiers in the multi-source power operation and transaction data, thus obtaining node feature data with a unified time index. For data with a sampling frequency higher than the preset transaction time granularity, aggregation can be performed according to the preset transaction time granularity; for data with a sampling frequency lower than the preset transaction time granularity, filling or mapping can be performed according to the applicable time; for data with differences between release time and applicable time, time alignment is prioritized based on the applicable time, while the release time is retained for subsequent data availability judgment.

[0079] After time alignment is completed, the heterogeneous spatiotemporal sample construction module generates corresponding nodes based on data categories. Source-side nodes are generated based on available output data from the source side, carrying node characteristics such as available output, available capacity, equipment status, or output forecasts for the source-side object. Load-side nodes are generated based on load data from the load side, carrying load values, load forecasts, or load aggregation information for the load-side object. Electricity price nodes are generated based on electricity price data, carrying day-ahead electricity prices, real-time electricity prices, price differences, or market zoning price information. Meteorological nodes are generated based on meteorological data, carrying meteorological grids, meteorological release times, and meteorological element information. Rule nodes are generated based on contract and rule data, carrying information on electricity mapping relationships, data aggregation standards, deviation calculation standards, and rule version information.

[0080] To enable different types of nodes to be identified and processed by subsequent models within the same heterogeneous spatiotemporal sample, the heterogeneous spatiotemporal sample construction module configures node type identifiers, forecast time identifiers, market zone identifiers, rule version identifiers, and data availability attributes for source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, and rule nodes. The node type identifier distinguishes node categories; the forecast time identifier indicates the forecast time location corresponding to the node characteristics; the market zone identifier indicates the market zone to which the node belongs or is applicable; the rule version identifier ensures that the node characteristics are consistent with the corresponding rule version; and the data availability attribute is obtained from the data availability attribute in step S1 and can be updated according to the node generation process.

[0081] Through the above processing, step S2 generates a heterogeneous spatiotemporal sample containing source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, and rule nodes. This heterogeneous spatiotemporal sample organizes node feature data with a unified time index and retains attributes such as node type, prediction time, market partition, rule version, and data availability. It can serve as input for subsequent causal availability label generation, rule constraint information generation, candidate correlation determination, and dynamic sparse correlation graph construction.

[0082] Step S3: Based on the data availability at the prediction time, generate causal availability tags for the node feature data in the heterogeneous spatiotemporal samples, and generate rule constraint information based on contract and rule data;

[0083] In this embodiment, step S3 is executed by the causal availability and rule constraint generation module. This module receives the heterogeneous spatiotemporal samples output from step S2 and the contract and rule data obtained from step S1. It is used to determine whether the node feature data can be used in the actual prediction process based on the prediction time, and to convert the contract and rule data into machine-readable rule constraint information that can be called by subsequent candidate association filtering and multi-task probability prediction.

[0084] Specifically, the causal availability and rule constraint generation module categorizes node characteristic data in heterogeneous spatiotemporal samples into data available before the prediction time, data available during the operating day, and data available after settlement, based on the data availability attributes, release time, applicable time, and prediction time of the node characteristic data. Data available before the prediction time refers to data that has been released and can be read by the power trading prediction system before the current prediction time; data available during the operating day refers to data that can be obtained during the operating day according to the updated rules of the dispatching, market, meteorological, or metering systems; and data available after settlement refers to data that is confirmed or released after the prediction time, including post-settlement confirmed data, deviation assessment result data, and post-correction data.

[0085] After data partitioning, the causal availability and rule constraint generation module generates causal availability tags based on data available before the prediction time and data available on the runtime day. These tags identify whether the corresponding node's feature data is allowed to participate in model input, candidate association determination, and dynamic sparse association graph construction at the prediction time. For data available after settlement, the module marks it as disabled data during the inference phase, preventing it from being used as model input in actual prediction and inference. This process ensures consistent data availability conditions between the training and inference phases, reducing the risk of data after the prediction time being misused for prediction calculations.

[0086] Simultaneously, the causal availability and rule constraint generation module extracts contract mapping relationships, settlement aggregation criteria, deviation assessment criteria, data availability restrictions, effective time, and rule version numbers from contract and rule data to generate machine-readable rule constraint information. Specifically, contract mapping relationships describe the mapping relationships between different market segments, trading objects, or electricity objects; settlement aggregation criteria describe the data aggregation methods between different time granularities, node objects, or segment objects; deviation assessment criteria describe the calculation criteria for deviation amount and deviation cost-related data; data availability restrictions limit the usability of specific data at different forecast times; and effective time and rule version numbers determine the applicability of rule constraint information at different forecast times.

[0087] During the predictive model training phase, the causal availability and rule constraint generation module generates causal availability labels for historical node feature data, consistent with those used in the inference phase, based on the data availability attributes corresponding to historical prediction times. For data released after the historical prediction time and data available only after post-calculation, even if such data already existed during the construction of training samples, it is masked according to the causal availability labels, preventing it from being included in the model input for the corresponding historical prediction time. This process ensures that the data input conditions during the training phase are consistent with those during the actual prediction phase, thereby avoiding the introduction of future information into the training samples.

[0088] Through the above processing, step S3 outputs causal availability labels and machine-readable rule constraint information. The causal availability labels are used for data availability control in subsequent candidate association filtering, dynamic sparse association graph generation, and multi-task probability prediction; the machine-readable rule constraint information is used for subsequent determination of constrained associations, filtering of valid association edge sets, data caliber matching, and generation of deviation cost probability prediction results. This step enables node feature data in heterogeneous spatiotemporal samples to enter the subsequent modeling process under the data conditions and contract rule constraints at the prediction time.

[0089] Step S4: Based on the physical association, meteorological propagation association, historical statistical association and constraint association between different nodes in the heterogeneous spatiotemporal sample, candidate association relationships with directional identifiers and time lags are determined according to the association type. The candidate association relationships are constrained and screened based on causal availability labels and machine-readable rule constraint information to generate a set of effective association edges. The set of effective association edges forms a dynamic sparse association graph that changes with the prediction time.

[0090] In this embodiment, step S4 is performed by the candidate association determination module and the dynamic sparse association graph generation module. The candidate association determination module receives the heterogeneous spatiotemporal samples formed in step S2 and, in conjunction with the causal availability markers and machine-readable rule constraint information generated in step S3, identifies the possible associations between different nodes in the heterogeneous spatiotemporal samples. The nodes in the heterogeneous spatiotemporal samples include source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, and rule nodes. The associations between different types of nodes can originate from the physical connection of the power grid, the influence of meteorological propagation, historical data statistical relationships, and contract rule constraint relationships.

[0091] The candidate correlation determination module classifies physical correlation, meteorological propagation correlation, historical statistical correlation, and constraint correlation as different correlation types. Physical correlation describes the connection relationships between source-side nodes, load-side nodes, and electricity price nodes in the power grid topology, market zoning, or power transfer relationships; meteorological propagation correlation describes the impact of meteorological nodes on the available output of the source side or the load changes of the load side; historical statistical correlation describes the lagged correlation relationships between different nodes in their historical time series; and constraint correlation describes the constraints imposed by contract mapping relationships, data aggregation caliber, deviation assessment caliber, and rule versions on data connections or data usage scope between nodes.

[0092] In one implementation, the candidate association set consists of node pairs that satisfy preset candidate conditions. These preset candidate conditions include the existence of power grid topology mapping relationships, market zoning mapping relationships, meteorological propagation coverage relationships, historical lag correlation relationships, or rule-based mapping relationships. For any node... and nodes When both satisfy any of the above preset candidate conditions, the system adds the corresponding node pair and its association type to the candidate association set. A candidate association in the candidate association set can be represented as... ,in and Represents a node. This indicates the association type. For the same pair of nodes, if multiple candidate conditions for association types are met simultaneously, multiple candidate association relationships are generated according to different association types.

[0093] For each candidate association, the candidate association determination module generates a corresponding dimensionless candidate association factor according to the association type, and performs credibility calibration based on data quality status, rule version validity status, and data availability status at the prediction time to obtain the composite association strength. The dimensionless candidate association factor is used to transform association information from different sources into a comparable numerical space, avoiding the direct mixing of data with different dimensions such as electricity volume, electricity price, distance, and time. Data quality status can be determined based on data missingness, anomalies, and timestamp consistency; rule version validity status can be determined based on rule version number, effective time, and expiration time; and data availability status at the prediction time can be determined based on causal availability markers. The composite association strength after credibility calibration is a dimensionless value and lies in the same numerical space as the dimensionless association strength threshold, allowing for direct comparison.

[0094] The candidate association determination module also configures direction identifiers and time lag values ​​for different association types. Direction identifiers indicate the direction of information transmission between nodes, such as the influence direction of a meteorological node on a source-side node, the influence direction of a meteorological node on a load-side node, or the influence direction of a source-side or load-side node on an electricity price node. Time lag values ​​represent the time delay in the influence relationship between nodes. For physical associations, the time lag value can be determined based on physical propagation relationships; for meteorological propagation associations, it can be determined based on meteorological propagation relationships; for historical statistical associations, it can be determined based on historical statistical lag relationships; and for constrained associations, it can be determined based on rule-based time mapping relationships. These time lag values ​​are used to subsequently determine whether the data corresponding to the candidate association at the prediction time can be used.

[0095] After obtaining candidate associations, the dynamic sparse association graph generation module performs constraint filtering on the candidate associations based on composite association strength, direction identifier, time lag, causal availability label, and machine-readable rule constraint information, generating a set of valid association edges. The set of valid association edges can be represented as:

[0096] ;

[0097] In the formula, Indicates the predicted time The set of valid associated edges below; Represents the set of candidate associations; and Represents a node; Indicates the association type; Represents a node With nodes At the predicted time and related types The strength of the composite correlation under the following conditions; Represents a combination of node types In association type The dimensionless correlation strength threshold is below; and Representing nodes respectively and nodes The node type; Represents a node With nodes At the predicted time and related types The time lag; Indicates the predicted time The corresponding machine-readable rule constraint information.

[0098] Indicated by direction identifier Time lag and prediction time Defined association type constraint functions; This indicates the availability status of data at the predicted time, as determined by the causal availability label. This indicates a consistent state of transaction criteria determined by rule-based constraint information; Indicates the predicted time Corresponding rule constraint information A state where the rules are clearly defined as prohibited;

[0099] in, , , and Both are binary state functions;

[0100] This represents the association type constraint function, used to determine whether the direction identifier, time lag, and prediction time under the corresponding association type satisfy the constraint conditions of that association type. When the direction identifier is consistent with the allowed propagation direction of the corresponding association type, and the time lag is within the allowed lag window of that association type, the association type constraint function takes a value of 1; otherwise, it takes a value of 0.

[0101] This indicates the availability status of data at the predicted time, as determined by the causal availability label. When the node... ,node And when the node feature data and edge feature data referenced by the candidate association according to the time lag are both allowed to be used by the causal availability tag at the prediction time, The value is 1; otherwise, the value is 0.

[0102] This indicates a consistent data caliber determined by machine-readable rule constraints. It occurs when the data at both ends of a candidate relationship can be mapped to the same preset transaction time granularity, the same market partition, or the same rule version. The value is 1; otherwise, the value is 0.

[0103] Indicates the predicted time The connection prohibition status is determined by the corresponding machine-readable rule constraint information. This applies when the machine-readable rule constraint information prohibits the connection to the corresponding node, prohibits the corresponding data combination, or the corresponding rule version is invalid. The value is 1; otherwise, the value is 0.

[0104] In the above screening process, only candidate associations that simultaneously meet the following conditions are retained as valid association edges: the composite association strength reaches the corresponding dimensionless association strength threshold, the association type constraint is valid, the data at the prediction time is available, the data caliber is consistent, and the association is not prohibited by the rules. Since the composite association strength and the dimensionless association strength threshold are in the same dimensionless numerical space, comparing the two will not cause inconsistency issues in dimensions; the remaining screening conditions are all binary state judgments and do not involve direct calculations between different physical quantities.

[0105] On the computational boundary, candidate associations lacking necessary node features, edge features, or machine-readable rule constraints are not included in the set of valid association edges. When the rule version changes, the set of valid association edges is recalculated according to the rule version applicable at the prediction time. When the composite association strength of a candidate association is missing, the corresponding composite association strength can be set to 0, or the candidate association can be marked as invalid. When there are no valid association edges that meet the conditions at a certain prediction time, the system retains the nodes and their node features in the heterogeneous spatiotemporal samples, and can form a dynamic graph input based on node self-connection to ensure that the subsequent spatiotemporal feature extraction process continues to be executed.

[0106] The dynamic sparse correlation graph generation module constructs a dynamic sparse correlation graph based on the set of valid correlation edges. Nodes in the dynamic sparse correlation graph originate from source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, and rule nodes in heterogeneous spatiotemporal samples, while edges come from the set of valid correlation edges. As the prediction time progresses, the data availability status of node feature data, meteorological propagation relationships, historical statistical relationships, and machine-readable rule constraints may change. The dynamic sparse correlation graph generation module redetermines the set of valid correlation edges based on the updated input, thus forming a dynamic sparse correlation graph that changes with the prediction time.

[0107] Through the above processing, step S4 outputs a dynamic sparse association graph and its corresponding set of valid association edges. This output serves as the input for spatiotemporal feature extraction in step S5, enabling the subsequent graph neighborhood aggregation process to proceed between node connections that satisfy causal availability, data consistency, and rule constraints.

[0108] Step S5: Based on the dynamic sparse association graph, perform spatiotemporal feature extraction on the node feature data in the heterogeneous spatiotemporal samples to obtain the spatiotemporal hidden state of the nodes;

[0109] The spatiotemporal feature extraction process includes:

[0110] Based on node type, the node feature data in the heterogeneous spatiotemporal samples are subjected to type-adaptive projection to obtain source-side node representation, load-side node representation, electricity price node representation, meteorological node representation, and regular node representation;

[0111] Based on the effective association edges in the dynamic sparse association graph, graph neighborhood aggregation is performed on the source-side node representation, load-side node representation, electricity price node representation, meteorological node representation, and rule node representation to obtain the graph-constrained node representation;

[0112] The temporal feature encoding of the graph constraint node representation is performed according to the predicted time sequence to obtain the spatiotemporal hidden state of the node.

[0113] The spatiotemporal latent state of a node is generated from the available power output information on the source side, the load information on the load side, the electricity price information, the meteorological propagation information, and the rule constraint information.

[0114] In this embodiment, step S5 is performed by the spatiotemporal feature extraction module. The spatiotemporal feature extraction module receives the heterogeneous spatiotemporal samples formed in step S2 and the dynamic sparse correlation graph generated in step S4, and extracts features from the node feature data in the heterogeneous spatiotemporal samples under the constraints of the dynamic sparse correlation graph to obtain the node spatiotemporal latent states used for subsequent multi-task probability prediction.

[0115] Because the heterogeneous spatiotemporal samples contain source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, and rule nodes, the original feature dimensions, numerical ranges, and business meanings of different types of nodes are not the same. Therefore, the spatiotemporal feature extraction module first performs type-adaptive projection on the node feature data according to the node type. For source-side nodes, the available output, available capacity, equipment status, and source-side forecast data are mapped to source-side node representations; for load-side nodes, load values, load forecast values, and load aggregation information are mapped to load-side node representations; for electricity price nodes, day-ahead electricity prices, real-time electricity prices, and price difference data are mapped to electricity price node representations; for meteorological nodes, meteorological grid data and meteorological element data are mapped to meteorological node representations; for rule nodes, contract mapping relationships, settlement aggregation standards, deviation assessment standards, and rule version information are mapped to rule node representations. The above type-adaptive projection can be implemented using linear mapping, embedded mapping, or multilayer perceptrons. Its function is to transform the features of different types of nodes to a unified latent space dimension, which facilitates subsequent graph neighborhood aggregation.

[0116] After completing the type-adaptive projection, the spatiotemporal feature extraction module performs graph neighborhood aggregation based on the effective association edges in the dynamic sparse association graph. The effective association edges in the dynamic sparse association graph are selected in step S4 based on composite association strength, direction identifier, time lag, causal availability marker, and machine-readable rule constraint information. Therefore, graph neighborhood aggregation is only performed between node connections that satisfy data availability, data caliber consistency, and rule constraint conditions. For any node, the spatiotemporal feature extraction module obtains the node representations of adjacent source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, or rule nodes based on the effective association edges connected to that node, and performs information aggregation by combining the association type and direction information corresponding to the effective association edges to obtain the graph-constrained node representation.

[0117] In one implementation, graph neighborhood aggregation can be achieved using weighted summation, attention aggregation, message passing, or other graph neural network aggregation methods. Aggregation weights can be determined based on the composite association strength, association type, direction identifier, or rule constraint state corresponding to the effective associated edges. For effective associated edges with time lags, the spatiotemporal feature extraction module references the node representations of neighboring nodes at the corresponding historical prediction times according to the corresponding time lag, ensuring that meteorological propagation, historical statistical lags, and rule-based time mapping relationships are reflected in the node representation update process. This aggregation process can integrate source-side available power output information, load-side load information, electricity price information, meteorological propagation information, and rule constraint information into the graph-constrained node representation.

[0118] After completing graph neighborhood aggregation, the spatiotemporal feature extraction module performs temporal feature encoding on the graph constraint node representations according to the prediction time sequence. Temporal feature encoding is used to learn the variation patterns of the same node at consecutive prediction times, as well as the temporal dependencies between different prediction times. Temporal feature encoding can be implemented using recurrent neural networks, gated recurrent units, long short-term memory networks, temporal convolutional networks, Transformer temporal encoders, or other temporal encoding structures. The input is a sequence of graph constraint node representations arranged according to prediction times, and the output is the spatiotemporal latent state of the node at the corresponding prediction time.

[0119] The spatiotemporal latent state of a node is used to represent the spatiotemporal characteristics of the node after fusion at the prediction time. For source-side nodes, the spatiotemporal latent state includes temporal variations, meteorological influences, and rule constraints related to the available output of the source side; for load-side nodes, the spatiotemporal latent state includes temporal patterns, meteorological influences, and market zoning correlation information related to load variations; for price nodes, the spatiotemporal latent state includes source-load linkage information and rule constraints related to day-ahead price, real-time price, and price difference changes; for meteorological nodes and rule nodes, the spatiotemporal latent state is used to transmit meteorological propagation information and rule constraint information to relevant source-side nodes, load-side nodes, and price nodes.

[0120] Through the above processing, step S5 outputs the spatiotemporal hidden state of the node. This spatiotemporal hidden state of the node is jointly generated by the source-side available power output information, load-side load information, electricity price information, meteorological propagation information, and rule constraint information after being constrained by the dynamic sparse correlation graph. It can be used as a unified feature input for the source-side available power output prediction, load-side load prediction, electricity price prediction, and deviation cost prediction in step S6.

[0121] Step S6: Perform multi-task probability prediction based on the spatiotemporal hidden state of nodes, machine-readable rule constraint information, and causal availability markers to generate source-side available output probability prediction results, load-side load probability prediction results, and electricity price probability prediction results; after data caliber matching, generate deviation cost probability prediction results based on the load-side load probability prediction results, electricity price probability prediction results, predicted baseline electricity volume, and machine-readable rule constraint information; generate a machine-readable electricity trading prediction result package based on the source-side available output probability prediction results, load-side load probability prediction results, electricity price probability prediction results, and deviation cost probability prediction results.

[0122] In this embodiment, step S6 is executed by the multi-task probability prediction module and the prediction result generation module. The multi-task probability prediction module receives the node spatiotemporal latent state output in step S5, and receives the machine-readable rule constraint information and causal availability tag generated in step S3, which are used to generate the source-side available output probability prediction result, the load-side load probability prediction result, and the electricity price probability prediction result. The prediction result generation module generates the deviation cost probability prediction result based on the load-side load probability prediction result, the electricity price probability prediction result, the predicted baseline electricity volume, and the machine-readable rule constraint information, and further generates a machine-readable electricity trading prediction result package.

[0123] The multi-task probabilistic prediction module includes a shared feature input layer, a rule-state embedding layer, a data availability state embedding layer, a task prediction branch, and a quantile output layer. The shared feature input layer receives the spatiotemporal latent states of nodes; the rule-state embedding layer converts machine-readable rule constraints corresponding to the prediction time into rule features that the model can process; and the data availability state embedding layer converts causal availability labels into a set of data availability states at the prediction time. The outputs of the shared feature input layer, rule-state embedding layer, and data availability state embedding layer are concatenated or fused and then input into the source-side available output task prediction branch, the load-side load task prediction branch, and the electricity price task prediction branch, respectively. The electricity price task prediction branch includes a day-ahead electricity price task prediction branch and a real-time electricity price task prediction branch, which output the day-ahead electricity price probabilistic prediction results and the real-time electricity price probabilistic prediction results, respectively.

[0124] Each prediction branch generates a probability prediction interval for the corresponding prediction task based on the spatiotemporal latent state of the node, expressed as:

[0125] ;

[0126] In the formula, Represents a node At the predicted time Next The probability prediction interval for the prediction task; This indicates the type of forecast task, which includes source-side available output tasks, load-side load tasks, and electricity price tasks. Indicates the first The prediction branch corresponding to the prediction task; Represents nodes obtained from dynamic sparse association graph constraints. At the predicted time The spatiotemporal hidden state of the node; Indicates the predicted time Corresponding rule constraint information; This represents the set of data availability states at the predicted time, formed by causal availability labels;

[0127] In one implementation, the probability prediction interval includes the median prediction, upper quantile prediction, and lower quantile prediction for the corresponding prediction task. The prediction branch first outputs the median prediction for the corresponding prediction task, and then outputs the non-negative quantile expansion in the target space of the corresponding prediction task. For the ... For prediction tasks, the upper quantile predicted value and the lower quantile predicted value can be represented as:

[0128] ;

[0129] In the formula, Indicates the first The median predicted value for a predictive task; This indicates upper quantiles greater than 0.5; This indicates the lower quantiles less than 0.5; and They represent the first The non-negative quantile extension in the target space of a prediction task. The non-negative quantile extension can be obtained from a non-negative activation function, absolute value mapping, or a non-negative constraint output layer, ensuring that the lower quantile, median, and upper quantile predictions for the same prediction task satisfy the quantile order. For source-side available output tasks and load-side load tasks, the predicted values ​​can be boundary-constrained according to the physical value range to ensure that the corresponding power or electricity is not lower than the physical allowable lower limit; for electricity price tasks, the predicted values ​​retain the corresponding price value range according to applicable market rules.

[0130] Further, in step S6, the prediction result generation module receives the source-side available output probability prediction results, load-side load probability prediction results, and electricity price probability prediction results output by the multi-task probability prediction module, and also receives machine-readable rule constraint information, causal availability tags, dynamic sparse correlation graphs, and predicted baseline electricity volume, which are used to generate deviation cost probability prediction results and machine-readable electricity trading prediction result packages. The electricity price probability prediction results include day-ahead electricity price probability prediction results and real-time electricity price probability prediction results.

[0131] The forecasting result generation module first performs data caliber matching on the load-side probability forecasting results, day-ahead electricity price probability forecasting results, real-time electricity price probability forecasting results, and the forecast benchmark electricity volume under the same preset trading time granularity. Data caliber matching is performed based on the electricity volume mapping relationship, data aggregation caliber, deviation calculation caliber, and constraint version number in the machine-readable rule constraint information. For the load-side probability forecasting results, the forecasting result generation module converts the load-side probability forecasting value into the corresponding market partition and forecasting time's predicted electricity volume based on the market partition identifier, preset trading time granularity, and data aggregation caliber. For the day-ahead and real-time electricity price probability forecasting results, the forecasting result generation module converts the day-ahead and real-time electricity price forecasting values ​​into electricity price forecasting values ​​under the same market partition, the same preset trading time granularity, and the same rule version based on the market partition identifier, price applicable period, and constraint version number. For the forecast benchmark electricity volume, the forecasting result generation module converts the forecast benchmark electricity volume into an electricity benchmark under the same time granularity and data caliber as the load-side probability forecasting electricity volume, based on the electricity volume mapping relationship, data aggregation caliber, and deviation calculation caliber. After data caliber matching, the electricity and price data used in deviation cost calculation have consistent time granularity, market segmentation, and rule version.

[0132] After completing the data caliber matching, the forecast result generation module constructs a source-load price quantile combination set based on the caliber-matched load-side load probability forecast results, day-ahead electricity price probability forecast results, real-time electricity price probability forecast results, and the predicted baseline electricity volume. The module reads the load-side load forecast quantile set from the load-side load probability forecast results, the real-time electricity price forecast quantile set from the real-time electricity price probability forecast results, and the day-ahead electricity price forecast quantile set from the day-ahead electricity price probability forecast results. It then combines the load-side load forecast quantiles, real-time electricity price forecast quantiles, and day-ahead electricity price forecast quantiles according to the forecast time and market segment to form market segments. At the predicted time The following is a set of source-load price quantile combinations Each quantile combination in the source-load price quantile combination set includes a load-side load forecast quantile, a real-time electricity price forecast quantile, and a day-ahead electricity price forecast quantile. (Source-load price quantile combination set) Any quantile combination in is denoted as .

[0133] The source-load price quantile combination set introduces source-side node constraints through source-load price linkage weights. The predicted probability of available output from the source side is determined by the source-side nodes in the dynamic sparse correlation graph participating in the source-load price linkage weights. The effective correlation states between source-side nodes and load-side nodes, as well as electricity price nodes, are used to constrain the validity of the quantile combination and weight allocation. The predicted probability of available output from the source side is not used as an input to the product of candidate deviation cost values; instead, it participates in the generation of the deviation cost probability prediction result through the source-load price linkage weights. This setup allows the deviation cost probability prediction result to simultaneously utilize load-side load prediction results, electricity price prediction results, and the dynamic correlation information between source-side nodes and load-side nodes and electricity price nodes.

[0134] For each quantile combination in the source-load price quantile combination set, the prediction result generation module determines the predicted load-side load based on the load-side load prediction quantile corresponding to the quantile combination, determines the predicted real-time electricity price based on the real-time electricity price prediction quantile corresponding to the quantile combination, and determines the predicted day-ahead electricity price based on the day-ahead electricity price prediction quantile corresponding to the quantile combination. The predicted load-side load, real-time electricity price, and day-ahead electricity price are all data after data caliber matching.

[0135] The prediction result generation module then determines the candidate values ​​for deviation costs under the corresponding quantile combination. First, the module calculates the difference between the predicted baseline electricity volume after data caliber matching and the predicted load electricity volume corresponding to the quantile combination, and uses this difference as the predicted electricity volume deviation. Next, it calculates the difference between the real-time electricity price prediction value and the day-ahead electricity price prediction value corresponding to the quantile combination, and uses this difference as the price difference. Finally, based on the deviation calculation caliber in the machine-readable rule constraint information, it determines the calculation method for the candidate values ​​for deviation costs by generating the predicted electricity volume deviation and price difference.

[0136] In the implementation where the deviation calculation uses an absolute deviation cost approach, the prediction result generation module determines the candidate deviation cost value for the quantile combination based on the absolute value of the product of the predicted electricity deviation and the price difference. In the implementation where the trading rules differentiate between positive and negative deviations, use segmented assessment coefficients, deviation penalty coefficients, or revenue deduction rules for deviation calculation, the prediction result generation module calculates the predicted electricity deviation and price difference according to the deviation calculation method recorded in the machine-readable rule constraint information, obtaining the candidate deviation cost value for the corresponding quantile combination. Therefore, the calculation process for the candidate deviation cost value is consistent with the trading rules that take effect at the prediction time.

[0137] The prediction result generation module also generates prediction results for each quantile combination. Determine the source-load price linkage weight The source-load-price linkage weight represents the contribution of the corresponding quantile combination to the deviation cost target quantile under the linkage relationship between the source side, load side, and electricity price. The source-load-price linkage weight is jointly determined by the effective association state between the source-side nodes, load-side nodes, and electricity price nodes in the dynamic sparse association graph, machine-readable rule constraint information, and the data availability state at the prediction time.

[0138] The prediction result generation module determines the linkage strength corresponding to the quantile combination based on the composite linkage strength of the effective correlation edges between the source-side nodes, load-side nodes, and price nodes. It then calibrates and non-normalizes the linkage strength by combining the data caliber consistency status determined by machine-readable rule constraints and the data availability status determined by causal availability markers, ensuring that the sum of the source-load-price linkage weights within the same source-load-price quantile combination set is 1. For quantile combinations that do not meet the data caliber consistency or data availability requirements, the prediction result generation module sets the corresponding source-load-price linkage weight to 0 or does not include the corresponding quantile combination in the source-load-price quantile combination set. When there are no normalizable effective source-load-price linkage weights within the same source-load-price quantile combination set, the prediction result generation module assigns equal weights to quantile combinations that meet the data caliber matching and data availability requirements.

[0139] After obtaining the candidate values ​​of deviation cost and the corresponding source-load price linkage weights for each quantile combination, the prediction result generation module generates the predicted value of deviation cost at the target quantile of deviation cost according to the weighted quantile operator.

[0140] ;

[0141] In the formula, Indicates market segmentation At the predicted time and deviation cost target quantile The predicted deviation cost value; Indicates market segmentation At the predicted time The following is a set of source-load price quantile combinations; This represents any quantile combination in the set of source-load quantile combinations; Representing quantile combination Candidate values ​​for deviation costs; Representing quantile combination The source-load price linkage weight; This represents a weighted quantile operator, which determines the target quantile of deviation cost based on candidate deviation cost values ​​and corresponding source-load price linkage weights. The predicted deviation cost value;

[0142] The weighted quantile operator is used to determine the target quantile of deviation cost based on the order of candidate deviation cost values ​​and their corresponding source-load price linkage weights. The predicted deviation cost is calculated based on the following: During execution, the prediction result generation module sorts the candidate deviation cost values ​​for the same market segment and the same prediction time from smallest to largest, and then accumulates them according to the source-load-price linkage weights after sorting; when the accumulated weights reach the target quantile for deviation cost... At that time, the prediction result generation module uses the candidate value of deviation cost corresponding to the current position as the predicted value of deviation cost. The weighted quantile operator acts on the set of candidate values ​​of deviation cost under the same cost dimension, and the output of the weighted quantile operator is the predicted value of deviation cost.

[0143] The prediction result generation module performs the above calculations for multiple target quantiles of deviation costs, generating deviation cost probability prediction results. These results include the median deviation cost prediction, the upper quantile deviation cost prediction, and the lower quantile deviation cost prediction. In other embodiments, the deviation cost probability prediction results also include deviation cost predictions at other target quantiles set according to the needs of power trading risk management.

[0144] At the calculation boundary, when a certain quantile combination lacks the predicted load power, real-time electricity price forecast, day-ahead electricity price forecast, or predicted baseline power, the prediction result generation module will not include the quantile combination in the source-load price quantile combination set; when a certain data is determined by the causal availability marker to be unavailable at the prediction time, the prediction result generation module will not use the above data to participate in the deviation cost calculation at the corresponding prediction time; when the rule version changes, the prediction result generation module will re-execute the data caliber matching and source-load price linkage weight calculation according to the new constraint version number.

[0145] The forecast result generation module generates a machine-readable electricity trading forecast result package based on the source-side available output probability forecast results, load-side load probability forecast results, electricity price probability forecast results, and deviation cost probability forecast results. The machine-readable electricity trading forecast result package includes the forecast date, market zone identifier, preset trading time granularity identifier, constraint version number, model version number, source-side available output probability forecast results, load-side load probability forecast results, electricity price probability forecast results, and deviation cost probability forecast results. The machine-readable electricity trading forecast result package is output in a structured data format, enabling electricity trading systems, risk assessment systems, or dispatch support decision-making systems to read and access it.

[0146] like Figure 2As shown, another embodiment of the present invention also provides a power source-load price linkage prediction system based on spatiotemporal correlation modeling. This system can be deployed on servers, cloud platforms, edge computing devices, or power trading prediction business systems, and is used to execute the aforementioned power source-load price linkage prediction method based on spatiotemporal correlation modeling. The system includes a data acquisition module, a heterogeneous spatiotemporal sample construction module, a constraint information generation module, a dynamic sparse correlation graph generation module, a spatiotemporal feature extraction module, and a prediction result generation module. Each module can be implemented through software programs, hardware circuits, or a combination of software and hardware.

[0147] The data acquisition module is used to acquire multi-source data on power operation and trading, and to standardize and encapsulate source-side available output data, load-side load data, electricity price data, meteorological data, and contract and rule data, so that data from different sources have data category identifiers, data source identifiers, timestamp identifiers, market partition identifiers, rule version identifiers, and data availability attributes. The heterogeneous spatiotemporal sample construction module is connected to the data acquisition module and is used to time-align the multi-source data on power operation and trading according to the preset trading time granularity, and generate heterogeneous spatiotemporal samples containing source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, and rule nodes.

[0148] The constraint information generation module is connected to the heterogeneous spatiotemporal sample construction module. It generates causal availability labels for node feature data in the heterogeneous spatiotemporal samples based on data availability at the prediction time, and generates machine-readable rule constraint information based on contract and rule data. The dynamic sparse correlation graph generation module is connected to both the constraint information generation module and the heterogeneous spatiotemporal sample construction module. It determines candidate correlation relationships based on physical correlations, meteorological propagation correlations, historical statistical correlations, and constraint correlations. It then filters candidate correlation relationships using causal availability labels and machine-readable rule constraint information, forming a set of effective correlation edges and a dynamic sparse correlation graph that changes with the prediction time.

[0149] The spatiotemporal feature extraction module is connected to the dynamic sparse correlation graph generation module. It performs type-adaptive projection, graph neighborhood aggregation, and temporal feature encoding on the node feature data in heterogeneous spatiotemporal samples according to the dynamic sparse correlation graph, obtaining the spatiotemporal latent states of the nodes. The prediction result generation module is connected to the spatiotemporal feature extraction module. It performs multi-task probabilistic prediction based on the node spatiotemporal latent states, machine-readable rule constraints, and causal availability labels. It generates source-side available output probability prediction results, load-side load probability prediction results, and electricity price probability prediction results. After data caliber matching, it generates deviation cost probability prediction results, and finally outputs a machine-readable power trading prediction result package.

[0150] In practical implementation, the above modules can be integrated into the same computing device or distributed across multiple computing nodes; data can be exchanged between modules through databases, message queues, application programming interfaces (APIs), or in-memory data structures. The functions of each module in the above system embodiments correspond to the corresponding steps in the foregoing method embodiments. Details not described in detail can be found in the relevant descriptions of the foregoing method embodiments and will not be repeated here.

[0151] This embodiment organizes multi-source power operation and trading data into heterogeneous spatiotemporal samples and generates a dynamic sparse correlation graph under the combined effect of data availability at the prediction time and rule constraint information. This allows the correlation modeling between source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, and rule nodes to adjust with changes in the prediction time and rule conditions. Furthermore, based on the dynamic sparse correlation graph, the spatiotemporal latent states of nodes are extracted, and based on data caliber matching, probabilistic prediction results related to available power output on the source side, load on the load side, electricity price, and deviation cost are generated. This forms a continuous technical solution from multi-source data processing and dynamic correlation modeling to machine-readable result output.

[0152] In summary, this invention addresses the challenges of heterogeneous multi-source data, limited data conditions at the prediction time, and complex trading criteria in power source-load price linkage forecasting scenarios. It constructs a complete processing flow from data organization and correlation modeling to probabilistic prediction and result output. This solution can be integrated with existing power trading forecasting systems, data management systems, and model service platforms. It is applicable to various scenarios such as renewable energy output forecasting, load forecasting, electricity price forecasting, deviation risk analysis, and trading decision support, demonstrating significant engineering deployment value and promising prospects for widespread application.

[0153] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. The aforementioned specific features, structures, materials, or characteristics may be combined in any appropriate manner in one or more embodiments or examples without contradicting each other. Those skilled in the art can combine or substitute technical features in different embodiments or examples based on the disclosure of this specification to form implementation methods capable of achieving the same or similar technical objectives.

[0154] The processes or methods described in conjunction with the flowcharts in this specification can be understood as consisting of one or more steps, modules, program fragments, or executable instructions used to implement corresponding logical functions or processing procedures. For each step in the method described in this application, the execution order, execution method, or module division of some steps can be adjusted according to the actual system deployment method without affecting data dependencies, processing logic, and technical effect implementation.

[0155] The above description is merely a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any equivalent changes, substitutions, or improvements that can be made by those skilled in the art based on the technical content disclosed in this application should fall within the scope of protection of this application. The scope of protection of this application shall be determined by the scope defined in the claims.

Claims

1. A method for predicting electricity source-load price linkage based on spatiotemporal correlation modeling, characterized in that, Includes the following steps: Acquire multi-source data on power operation and trading, including available power output data on the source side, load data on the load side, electricity price data, meteorological data, and contract and rule data; The power operation and transaction multi-source data are time-aligned according to the preset transaction time granularity to construct a heterogeneous spatiotemporal sample including source-side nodes, load-side nodes, electricity price nodes, meteorological nodes and rule nodes; Based on the availability of data at the prediction time, causal availability tags are generated for node feature data in heterogeneous spatiotemporal samples, and rule constraint information is generated based on contract and rule data. Based on the physical association, meteorological propagation association, historical statistical association and constraint association between different nodes in heterogeneous spatiotemporal samples, candidate association relationships with directional identifiers and time lags are determined according to the association type. The candidate association relationships are then constrained and screened based on causal availability labels and machine-readable rule constraint information to generate a set of effective association edges. The set of effective association edges forms a dynamic sparse association graph that changes with the prediction time. Based on the dynamic sparse correlation graph, spatiotemporal features are extracted from the node feature data in heterogeneous spatiotemporal samples to obtain the spatiotemporal hidden state of the node. Based on the spatiotemporal hidden state of nodes, machine-readable rule constraint information, and causal availability tags, multi-task probability prediction is performed to generate source-side available output probability prediction results, load-side load probability prediction results, and electricity price probability prediction results. After data caliber matching, deviation cost probability prediction results are generated based on load probability prediction results, electricity price probability prediction results, predicted baseline electricity volume, and machine-readable rule constraint information; a machine-readable power trading prediction result package is generated based on source-side available output probability prediction results, load probability prediction results, electricity price probability prediction results, and deviation cost probability prediction results.

2. The method for predicting power source-load price linkage based on spatiotemporal correlation modeling according to claim 1, characterized in that, The acquisition of multi-source data on power operation and transactions is carried out in the following ways: Obtain available power output data from the source side, load data from the load side, electricity price data, meteorological data, and contract and rule data; Configure data category identifiers, data source identifiers, timestamp identifiers, market zone identifiers, rule version identifiers, and data availability attributes for source-side available output data, load-side load data, electricity price data, meteorological data, contract and rule data; Based on data category identifiers, data source identifiers, timestamp identifiers, market partition identifiers, rule version identifiers, and data availability attributes, multi-source power operation and transaction data with the required time alignment identifiers are generated.

3. The method for predicting power source-load price linkage based on spatiotemporal correlation modeling according to claim 1, characterized in that, The construction of the heterogeneous spatiotemporal samples includes: According to the preset transaction time granularity, the multi-source data of power operation and transaction are time-aligned to obtain node feature data with a unified time index; Source-side nodes are generated based on available output data from the source side, load-side nodes are generated based on load data from the load side, electricity price nodes are generated based on electricity price data, meteorological nodes are generated based on meteorological data, and rule nodes are generated based on contract and rule data. Configure node type identifiers, forecast time identifiers, market partition identifiers, rule version identifiers, and data availability attributes for source-side nodes, load-side nodes, electricity price nodes, meteorological nodes, and rule nodes to form heterogeneous spatiotemporal samples.

4. The method for predicting power source-load price linkage based on spatiotemporal correlation modeling according to claim 1, characterized in that, The generation of causal availability markers and the generation of rule constraint information based on contract and rule data specifically include: Based on the prediction time, the node characteristic data are divided into data that can be obtained before the prediction time, data that can be obtained on the running day, and data that can be obtained after the settlement. Based on the data available before the prediction time and the data available on the operating day, a causal availability label is generated, and the data available for post-settlement is marked as data disabled during the inference phase. Extract contract mapping relationships, settlement aggregation criteria, deviation assessment criteria, data availability limitations, effective time and rule version number from contract and rule data to generate rule constraint information; During the training phase of the prediction model used to perform multi-task probabilistic prediction, causal availability tags are generated for historical node feature data based on the data availability attributes corresponding to the historical prediction time, consistent with the inference phase. Data released after the historical prediction time and data available for post-settlement are then masked based on the causal availability tags.

5. The method for predicting power source-load price linkage based on spatiotemporal correlation modeling according to claim 1, characterized in that, The determination of the candidate associations includes: Physical correlation, meteorological propagation correlation, historical statistical correlation, and rule-based correlation are respectively regarded as different correlation types; Dimensionless candidate association factors are generated for different association types, and the credibility of the dimensionless candidate association factors is calibrated according to the data quality status, rule version validity status, and data availability status at the prediction time, forming the composite association strength under the corresponding association type. Configure direction identifiers and time lag values ​​for different association types. The time lag values ​​are determined according to the physical propagation relationship, meteorological propagation relationship, historical statistical lag relationship, or rule time mapping relationship of the corresponding association type.

6. The method for predicting power source-load price linkage based on spatiotemporal correlation modeling according to claim 5, characterized in that, The set of valid association edges for generating the dynamic sparse association graph is determined according to the following rules: For any association type, based on the corresponding composite association strength, direction identifier, time lag, causal availability marker, and rule constraint information, a set of corresponding valid association edges is generated: ； In the formula, Indicates the predicted time The set of valid associated edges below; Represents the set of candidate associations; and Represents a node; Indicates the association type; Indicates the strength of the composite correlation; Represents a combination of node types In association type The dimensionless correlation strength threshold is below; Indicated by direction identifier Time lag and prediction time Defined association type constraint functions; This indicates the availability status of data at the predicted time, as determined by the causal availability label. This indicates a consistent state of transaction criteria determined by rule-based constraint information; Indicates the predicted time Corresponding rule constraint information A state where the rules are clearly defined as prohibited; , , and Both are binary state functions; When the rule constraint information changes, update the set of valid associated edges at the corresponding prediction time. A dynamic sparse association graph is formed based on the set of valid associated edges.

7. The method for predicting power source-load price linkage based on spatiotemporal correlation modeling according to claim 1, characterized in that, The spatiotemporal feature extraction process includes: Based on node type, the node feature data in the heterogeneous spatiotemporal samples are subjected to type-adaptive projection to obtain source-side node representation, load-side node representation, electricity price node representation, meteorological node representation, and regular node representation; Based on the effective association edges in the dynamic sparse association graph, graph neighborhood aggregation is performed on the source-side node representation, load-side node representation, electricity price node representation, meteorological node representation, and rule node representation to obtain the graph-constrained node representation; The temporal feature encoding of the graph constraint node representation is performed according to the predicted time sequence to obtain the spatiotemporal hidden state of the node. The spatiotemporal latent state of a node is generated from the available power output information on the source side, the load information on the load side, the electricity price information, the meteorological propagation information, and the rule constraint information.

8. The method for predicting power source-load price linkage based on spatiotemporal correlation modeling according to claim 1, characterized in that, The multi-task probability prediction generates probability prediction intervals as follows: The spatiotemporal hidden states of nodes obtained by the constraints of the dynamic sparse correlation graph are respectively input into the prediction branches corresponding to the source-side available output task, the load-side load task, and the electricity price task. Electricity pricing targets include day-ahead electricity pricing targets and real-time electricity pricing targets; Each prediction branch generates a probability prediction interval for the corresponding prediction task based on the spatiotemporal latent state of the node: ； In the formula, Represents a node At the predicted time Next The probability prediction interval for the prediction task; This indicates the type of forecast task, which includes source-side available output tasks, load-side load tasks, and electricity price tasks. Indicates the first The prediction branch corresponding to the prediction task; Represents nodes obtained from dynamic sparse association graph constraints. At the predicted time The spatiotemporal hidden state of the node; Indicates the predicted time Corresponding rule constraint information; This represents the set of data availability states at the predicted time, formed by causal availability labels; The probability prediction interval includes the median prediction, upper quantile prediction, and lower quantile prediction for the prediction task. The upper and lower quantile predictions are generated by the non-negative quantile expansion in the target space of the prediction task relative to the median prediction.

9. The method for predicting power source-load price linkage based on spatiotemporal correlation modeling according to claim 1, characterized in that, When generating the deviation cost probability prediction results and the machine-readable electricity trading prediction results package, the following are included: The electricity price probability prediction results are divided into day-ahead electricity price probability prediction results and real-time electricity price probability prediction results. Based on the power mapping relationship, data aggregation caliber, deviation calculation caliber and constraint version number in the machine-readable rule constraint information, the data caliber matching of the load-side load probability prediction results, day-ahead electricity price probability prediction results, real-time electricity price probability prediction results and the predicted benchmark power is performed under the same preset transaction time granularity. Based on the load-side load probability prediction results after caliber matching, the day-ahead electricity price probability prediction results, the real-time electricity price probability prediction results, and the predicted benchmark electricity volume, a source-load price quantile combination set is constructed. Each quantile combination in the source-load price quantile combination set includes a load-side load prediction quantile, a real-time electricity price prediction quantile, and a day-ahead electricity price prediction quantile. For each quantile combination in the source-load price quantile combination set, the predicted power deviation is determined based on the difference between the predicted baseline power and the predicted power of the load side corresponding to the quantile combination. The price difference is determined based on the difference between the real-time electricity price prediction value corresponding to the quantile combination and the day-ahead electricity price prediction value. Finally, the candidate value of the deviation cost is determined based on the absolute value of the product of the predicted power deviation and the price difference. Based on the candidate values ​​of deviation cost under each quantile combination and the corresponding source-load price linkage weights, the predicted value of deviation cost under the target quantile is generated according to the following formula: ； In the formula, Indicates market segmentation At the predicted time and deviation cost target quantile The predicted deviation cost value; Indicates market segmentation At the predicted time The following is a set of source-load price quantile combinations; This represents any quantile combination in the set of source-load quantile combinations; Representing quantile combination Candidate values ​​for deviation costs; Representing quantile combination The source-load price linkage weight; This represents a weighted quantile operator, which determines the target quantile of deviation cost based on candidate deviation cost values ​​and corresponding source-load price linkage weights. The predicted deviation cost value; The source-load-price linkage weight is a non-negative normalized weight, which is determined by the effective association status between the source-side nodes, load-side nodes and price nodes in the dynamic sparse association graph, machine-readable rule constraint information, and the data availability status at the prediction time. Based on the deviation cost prediction values ​​at multiple deviation cost target quantiles, a deviation cost probability prediction result is generated, and a machine-readable power trading prediction result package is generated based on the source-side available output probability prediction result, the load-side load probability prediction result, the electricity price probability prediction result, and the deviation cost probability prediction result.

10. A power source-load price linkage prediction system based on spatiotemporal correlation modeling, applied to the power source-load price linkage prediction method based on spatiotemporal correlation modeling as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-source data on power operation and trading. This multi-source data includes available power output data on the source side, load data on the load side, electricity price data, meteorological data, and contract and rule data. The heterogeneous spatiotemporal sample construction module is used to time-align multi-source data of power operation and trading according to the preset transaction time granularity, and construct heterogeneous spatiotemporal samples including source-side nodes, load-side nodes, electricity price nodes, meteorological nodes and rule nodes. The constraint information generation module is used to generate causal availability tags for node feature data in heterogeneous spatiotemporal samples based on the data availability at the prediction time, and to generate machine-readable rule constraint information based on contract and rule data. The dynamic sparse association graph generation module is used to determine candidate association relationships with directional identifiers and time lags based on physical associations, meteorological propagation associations, historical statistical associations and constraint associations between different nodes in heterogeneous spatiotemporal samples. It also performs constraint screening on the candidate association relationships based on causal availability labels and machine-readable rule constraint information, generates a set of effective association edges, and forms a dynamic sparse association graph that changes with the prediction time from the set of effective association edges. The spatiotemporal feature extraction module is used to extract spatiotemporal features from node feature data in heterogeneous spatiotemporal samples according to the dynamic sparse association graph, and obtain the spatiotemporal hidden state of the node. The prediction result generation module is used to perform multi-task probability prediction based on the spatiotemporal hidden state of nodes, machine-readable rule constraint information, and causal availability tags, and generate source-side available output probability prediction results, load-side load probability prediction results, and electricity price probability prediction results. After data caliber matching, deviation cost probability prediction results are generated based on load probability prediction results, electricity price probability prediction results, predicted baseline electricity volume, and machine-readable rule constraint information; a machine-readable power trading prediction result package is generated based on source-side available output probability prediction results, load probability prediction results, electricity price probability prediction results, and deviation cost probability prediction results.

Images