A source-load regulation potential evaluation method and system based on time-varying data management

Abstract

This application discloses a method and system for assessing source-load regulation potential based on time-varying data management, belonging to the field of power grid control technology. The method involves: verifying and weighting the multi-dimensional time-varying data to obtain a fused time-varying sequence and a collection level related to the data acquisition frequency; capturing the spatiotemporal correlation between the fused time-varying sequence and the power resources using a preset source-load potential assessment model to obtain the regulation potential boundary and constraints of the power resources within a prediction time window; and analyzing the adjustable parameters corresponding to each time segment within the prediction time window based on the regulation potential boundary and the constraints to obtain the source-load regulation potential assessment result of the power resources. Therefore, by implementing this application, the problems of existing technologies—such as the inability to effectively fuse power grid resource data from multiple data sources and the inability to achieve a balance between data processing accuracy and efficiency—can be solved.

Description

Technical Field

[0001] This application belongs to the field of power grid control technology, specifically relating to a method and system for evaluating source-load regulation potential based on time-varying data management. Background Technology

[0002] In new power systems, by regulating flexible resources such as distributed power sources, energy storage, and adjustable loads, problems such as short-term power supply and demand shortages and difficulties in integrating new energy sources can be addressed. These adjustable resources are called power resources, and the corresponding regulation range is called source-load regulation potential. Accurately assessing the range of source-load regulation potential is crucial for supporting real-time grid balance and optimized dispatch.

[0003] Currently, commonly used methods for assessing source-load regulation potential involve collecting operational data from various resources and calculating their available power regulation range based on fixed physical models or traditional machine learning models. Other methods attempt to integrate static nameplate parameters of resources, real-time power information, and simple environmental factors to construct an assessment model, outputting a numerical value or curve characterizing the resource's regulation capability to quantitatively assess the source-load regulation potential. However, all of these methods require data from various systems, and this data is heterogeneous; even simple weighted averaging cannot provide reliable input data. Furthermore, to ensure assessment accuracy, the assessment models employed have complex structures, resulting in high computational loads and large inference delays, making it difficult to meet the second-level real-time requirements of online regulation. This leads to low assessment efficiency and an inability to guarantee the real-time nature of the assessment results. Summary of the Invention

[0004] This application proposes a method and system for evaluating source-load regulation potential based on time-varying data management, which can effectively integrate multi-source heterogeneous power grid resource data and improve the evaluation efficiency of various source-load regulation potential while ensuring evaluation accuracy.

[0005] The first aspect of this application provides a method for assessing source-load regulation potential based on time-varying data management, the method comprising: Acquire multi-dimensional time-varying data for each power resource; wherein, the multi-dimensional time-varying data includes static attribute data, operating status data, environmental data, and power market data for each power resource; The multi-dimensional time-varying data is validated and weighted to obtain a fused time-varying sequence and a collection level related to the data collection frequency. By using a pre-defined source-load potential assessment model, the spatiotemporal correlation between the fused time-varying sequence and the power resources is captured, and the adjustment potential boundary and constraints of the power resources within the prediction time window are obtained; wherein, the source-load potential assessment model is constructed using the acquisition level; Based on the adjustment potential boundary and the constraints, the adjustable parameters corresponding to each time segment within the prediction time window are analyzed to obtain the source-load adjustment potential assessment results of the power resources.

[0006] To address the heterogeneity conflicts of multi-dimensional time-varying data from different data sources, the above scheme verifies the data sources before weighted fusion. It hierarchically manages the time-varying data based on data characteristics and collection urgency, verifies the reliability of the data sources, and fuses the data to generate a single, continuous, and reliable time-series data that best reflects the true state of resources, along with the corresponding collection levels. Adaptive hierarchical collection is achieved through these collection levels. Then, based on a pre-defined source-load potential assessment model, the adjustment capabilities of each resource in the present and future are accurately evaluated, yielding potential values ​​that more closely reflect the actual dispatching needs of the power grid and meet the real-time requirements of online control. Furthermore, constructing the source-load potential assessment model through the collection levels significantly reduces model parameters, lowers computational complexity, and improves the efficiency of adjustment potential assessment. Finally, the output source-load adjustment potential envelope unifies multiple physical constraints such as power, energy, and rate into a time-series representation, providing a high-precision and refined decision-making basis for power grid dispatch.

[0007] In one possible implementation of the first aspect, the multi-dimensional time-varying data is subjected to data source verification and weighted fusion to obtain a fused time-varying sequence and a collection level related to the data collection frequency, specifically: Based on preset data change rates and collection urgency, the multi-dimensional time-varying data is hierarchically divided to obtain the collection levels of the multi-dimensional time-varying data; wherein, the collection levels include a millisecond-level urgent critical layer, a minute-level dynamic change layer, and an hour-level slowly changing background layer; Based on the current control mode of the power system and the historical trend of the multi-dimensional time-varying data, adjust the acquisition frequency and sampling density of each acquisition level; By using preset dynamic credibility weights, the multi-dimensional time-varying data verified by the data source is weighted and fused to generate the fused time-varying sequence for each power resource.

[0008] The above scheme divides the time-varying data collected from various data sources into three acquisition levels based on the data change rate and the urgency of acquisition. By adjusting the acquisition frequency of each acquisition level, precise management of multi-dimensional data is achieved. Furthermore, by adjusting the levels according to the current control mode and historical change trends, the data acquisition strategy can flexibly adapt to the actual operating state of the power grid. Additionally, corresponding weights are assigned based on the reliability of the data sources, giving different data sources different influences, resulting in a fused time-varying sequence that better represents the true state of each resource.

[0009] In one possible implementation of the first aspect, the multi-dimensional time-varying data after data source verification specifically includes: Obtain similar time-varying data from at least a first threshold number of independent data sources for each power resource, and determine the dynamic reliability weight of the independent data sources based on the data type, real-time performance, and historical accuracy records of the independent data sources. The data types include direct measurement type and indirect estimation type, with the weighting factor of direct measurement type being greater than that of indirect estimation type; the data real-time performance is related to the data reporting delay of the same type of time-varying data, and the smaller the data reporting delay, the higher the weighting factor of the data real-time performance. Compare the same data item of each of the same time-varying data. If the difference between a value and other values ​​in the same data item exceeds a preset reasonable error range, then the value is defined as an outlier. The outliers are repaired, and the multi-dimensional time-varying data after data source verification is obtained based on the repair results.

[0010] The above scheme identifies outliers by comparing multi-source data for the same data item, effectively discovering and eliminating abnormal data caused by single sensor malfunctions or communication errors, thus improving data quality. Furthermore, the dynamically set weights evaluate different types of data sources, giving higher weights to more accurate data sources and enhancing the reliability of data fusion.

[0011] In one possible implementation of the first aspect, a pre-defined source-load potential assessment model is used to capture the spatiotemporal correlation between the fused time-varying sequence and the power resources, thereby obtaining the adjustment potential boundary and constraints of the power resources within the prediction time window. Specifically: The source load potential assessment model is constructed based on the preset grouped query attention mechanism and loss function; Based on the fused time-varying sequence, the constraints of each power resource at each time segment within the prediction time window are analyzed by the source-load potential assessment model; wherein, the constraints include real-time operating status constraints, physical regulation capability constraints, and external environmental incentive constraints. Based on the real-time operating status constraints and physical adjustment capability constraints, the maximum adjustable power range of each power resource at each time segment is calculated. Based on the external environmental incentive constraints, and by taking into account historical feedback information from the electricity market, the maximum adjustable power range is adjusted to obtain the effective adjustable power range for each time segment. The upper and lower limits of the effective adjustable power range are used to determine the adjustment potential boundary for each time segment.

[0012] The above scheme provides three different constraints to evaluate the regulation potential of each power resource from the perspectives of physical constraints of electrical equipment, safe operation requirements, and the incentive / inhibitory effects of the power market environment. The result is a regulation potential boundary that not only ensures stable operation of the power grid but also closely approximates the power market environment, making the prediction results more consistent with reality.

[0013] In one possible implementation of the first aspect, the source load potential evaluation model is constructed based on a preset grouped query attention mechanism and loss function, specifically as follows: A multi-head attention module of a preset teacher model is obtained. The attention heads of the multi-head attention module are divided according to the acquisition level to obtain several groups. Each group contains the same number of attention heads, and each group is only used to process data from one acquisition level. The model weights of the attention heads of each group are averaged and aggregated to generate the first model weight of the attention head of each group. Based on the attention heads of all groups, the group query attention mechanism is constructed. Obtain the model weights of all modules in the teacher model except for the first model weights to obtain the second model weights; Based on the second model weights, the grouped query attention mechanism is trained using the loss function to obtain the trained source load potential evaluation model.

[0014] The above scheme divides the attention heads of the multi-head attention module according to the collection level and averages the weights within each group, thereby constructing a more simplified grouped query attention mechanism. This significantly reduces the number of model parameters, lowers the model's computational complexity and data volume, improves the model's data processing efficiency, and further enhances the evaluation efficiency of adjustment potential. Moreover, the grouping corresponds to the collection level, ensuring that the evaluation model can specifically process data from different levels (different time scales), preserving the model's core spatiotemporal feature extraction capabilities, and achieving a balance between evaluation accuracy and evaluation efficiency.

[0015] In one possible implementation of the first aspect, the grouped query attention mechanism is trained using the loss function based on the second model weights to obtain the trained source load potential evaluation model, specifically: The second model weights are assigned to the corresponding parts of the source load potential assessment model. Then, using a preset knowledge distillation strategy, the weight coefficients of the group query attention mechanism are trained by minimizing the attention weight matrix of the teacher model and the source load potential assessment model through the loss function. The loss function includes a first loss term and a second loss term. The first loss term is used to calculate the difference in output distribution between the teacher model and the source load potential assessment model for the same training sample. The second loss term is used to calculate the difference in output distribution between the multi-head attention module and the grouped query attention mechanism.

[0016] The above scheme improves the performance of the assessment model by transferring the "knowledge" of the complex teacher model to a lightweight source-load potential assessment model through a knowledge distillation strategy. Moreover, by training only a portion of the weights of the assessment model, the parameter search space is greatly reduced, making the training process more stable.

[0017] In one possible implementation of the first aspect, the knowledge distillation strategy includes a first training phase and a second training phase. In the first training phase, the source load potential assessment model is trained using training samples with a completeness higher than a first threshold and a prediction time window length lower than a second threshold, and the adjustment potential boundary of all time sections is output. In the second training phase, the source load potential assessment model that has passed the first training phase is trained using training samples with a completeness lower than the first threshold or a prediction time window length higher than the second threshold, and the adjustment potential boundary of the next time segment is calculated based on the output value of the previous time segment.

[0018] The aforementioned approach divides the knowledge distillation strategy into two progressively more challenging stages to comprehensively enhance the model's robustness and generalization ability. This enables the model to learn to handle long-range dependencies, compensate for uncertainties in input data, and ensure the temporal consistency and rationality of multi-step predictions. Such training effectively enhances the model's robustness and accuracy in handling complex, long-term evaluation tasks.

[0019] In one possible implementation of the first aspect, based on the adjustment potential boundary and the constraints, the adjustable parameters corresponding to each time segment within the prediction time window are analyzed to obtain the source-load adjustment potential assessment result of the power resources, specifically: Based on the adjustment potential boundary and the constraints, the adjustable parameters for each time segment are constructed; the adjustable parameters include sub-vectors of four key indicators, namely, maximum adjustable upward power, maximum adjustable downward power, total sustainable adjustment energy, and maximum allowable adjustment rate; By integrating the vector elements of all the time sections, a time-varying source load regulation potential envelope is constructed to obtain the source load regulation potential assessment result.

[0020] The above scheme integrates the potential boundaries and key constraint parameters scattered across various future time segments, providing a directly usable and fully informed source-load regulation potential envelope for subsequent power grid dispatch. It clearly depicts the spatiotemporal evolution of the regulation capacity of various power resources over a period of time, not only indicating how much can be regulated, but also specifying how long and how quickly it can be regulated. This provides accurate and reliable quantitative input for the power system to conduct multi-timescale, multi-objective safety-constrained economic dispatch.

[0021] The second aspect of this application provides a source-load regulation potential assessment system based on time-varying data management, the system comprising: a data acquisition module, a data fusion module, a regulation boundary prediction module, and an adjustable parameter generation module; The data acquisition module is used to acquire multi-dimensional time-varying data of various power resources; wherein, the multi-dimensional time-varying data includes static attribute data, operating status data, environmental data and power market data of various power resources; The data fusion module is used to perform data source verification and weighted fusion on the multi-dimensional time-varying data to obtain the fused time-varying sequence and the acquisition level related to the data acquisition frequency; The adjustment boundary prediction module is used to capture the spatiotemporal correlation between the fused time-varying sequence and the power resources through a preset source-load potential assessment model, and obtain the adjustment potential boundary and constraints of the power resources within the prediction time window; The adjustable parameter generation module is used to analyze the adjustable parameters corresponding to each time segment within the prediction time window based on the adjustment potential boundary and the constraint conditions, so as to obtain the source-load adjustment potential assessment result of the power resources.

[0022] A third aspect of this application provides a terminal device, the device comprising: a terminal device including a processor and a memory, the memory storing a computer program, wherein the processor executes the computer program to implement the steps of the source-load regulation potential assessment method based on time-varying data management as described in any one of the embodiments of this application. Attached Figure Description

[0023] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the specific process of a source-load regulation potential assessment method based on time-varying data management provided in an embodiment of this application; Figure 2This is a structural diagram of a source-load regulation potential assessment system based on time-varying data management, provided in one embodiment of this application. Figure 3 This is a structural diagram of a terminal device provided in an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0027] First Embodiment Existing models for assessing the regulation potential of multiple power resources often employ complex model structures, resulting in high computational loads and large inference delays, making it difficult to meet the second-level real-time requirements of online regulation. If lightweight models are used to meet real-time requirements, limited model capacity or inaccurate training data labels often lead to decreased assessment accuracy and insufficient generalization ability. Therefore, existing resource regulation potential assessment methods struggle to strike a balance between real-time performance, assessment accuracy, and engineering practical efficiency.

[0028] This application addresses the aforementioned problems by employing a grouped query attention mechanism and transferring weights from a complex teacher model to a source load potential assessment model through knowledge distillation. While ensuring high assessment accuracy, it significantly reduces model complexity and inference latency, thus meeting the real-time requirements of online control.

[0029] like Figure 1 As shown, to address the problems in existing technologies where power grid resource data from multiple data sources cannot be effectively integrated, and where a balance cannot be struck between data processing accuracy and efficiency, the first embodiment of this application provides a detailed flowchart of a source-load regulation potential assessment method based on time-varying data management. This embodiment's source-load regulation potential assessment method based on time-varying data management includes steps S1 to S4, detailed below: Step S1: Obtain multi-dimensional time-varying data for each power resource.

[0030] First, multi-dimensional time-varying data for assessing the source-load regulation potential is acquired. This multi-dimensional time-varying data includes static attribute data, operating status data, environmental data, and electricity market data for each power resource.

[0031] The principle behind assessing the regulation potential of power resources lies in the fact that the regulation capacity of resources depends not only on their inherent equipment parameters but also on the dynamic constraints of their real-time operating conditions and the surrounding environment and market conditions. Static attribute data of power resources refers to the inherent parameters of equipment that do not change or change slowly over time, such as the rated power and ramp rate of generator sets, the total capacity and maximum charging and discharging power of energy storage devices, and the rated power and maximum allowable interruption duration of interruptible loads; these data constitute the physical boundary basis of the resource's regulation capacity. Operating status data of power resources refers to rapidly changing data collected in real-time or near real-time, reflecting the current operating conditions of the resources, such as the current actual power generation of photovoltaic power plants or wind farms, the current state of charge of energy storage systems, the current power consumption and operating conditions of adjustable loads, and the current battery charge and charging status of electric vehicles; these data determine the available regulation margin of the resources at the current moment. The environmental data and electricity market data refer to dynamic information from outside the resources that affects their regulation behavior or regulation value. For example, ultra-short-term wind speed and solar intensity forecasts for the assessment time, ambient temperature, real-time nodal marginal electricity prices, ancillary service market clearing price signals, and regulation demand instructions issued by the power grid. These data reflect the external driving factors and constraints that affect the willingness and feasibility of resource regulation.

[0032] Step S2: Perform data source verification and weighted fusion on the multi-dimensional time-varying data to obtain the fused time-varying sequence and the acquisition level related to the data acquisition frequency.

[0033] This step implements hierarchical management based on the inherent characteristics of the data and system requirements, and uses multi-source cross-validation and adaptive weighting strategies to generate a single, continuous, and reliable fused time-varying sequence that best reflects the true state of resources.

[0034] First, the multi-dimensional time-varying data is divided into three acquisition levels according to the preset data change rate and acquisition urgency. These three acquisition levels are the millisecond-level urgent and critical layer, the minute-level dynamic change layer, and the hour-level slowly changing background layer.

[0035] Specifically, the emergency critical layer includes event-type data that requires millisecond-level response, such as grid frequency deviation and circuit breaker tripping signals; the dynamic change layer includes continuously changing operational data ranging from seconds to minutes, such as photovoltaic output, load power, energy storage status of charge, and real-time electricity price; and the gradual change background layer includes status and environmental data updated on hourly or longer time scales, such as weather trend forecasts, equipment health indicators, and medium- and long-term market expectations.

[0036] The aforementioned hierarchical division forms the basis for implementing differentiated data management. For each acquisition level, the acquisition frequency and sampling density are adjusted based on the current control mode of the power system and the historical trends of various time-varying data.

[0037] For example, when the power system needs to make rapid frequency response, the data acquisition cycle that is strongly related to frequency regulation in the dynamic change layer will be automatically compressed; while for temperature data with slow historical changes, the acquisition cycle can be appropriately extended, thereby optimizing system communication and computing resources while ensuring data timeliness.

[0038] Then, the multi-dimensional time-varying data is validated to assess the reliability of each data source. For each type of power resource, similar time-varying data from at least three independent data sources are acquired in parallel. These independent data sources may include local equipment monitoring systems, station-level control systems, and power grid dispatch automation systems, etc.

[0039] Based on the data source type, data real-time performance, and historical accuracy records, a corresponding dynamic reliability weight is calculated for each independent data source. The data types include direct measurement and indirect estimation, with direct measurement having a higher weight factor than indirect estimation. Data real-time performance is related to the data reporting latency of similar time-varying data; the lower the data reporting latency, the higher the weight factor for data real-time performance. Historical accuracy records are obtained by comparing with a set benchmark data set.

[0040] The system compares the same data item from different time-varying data sources in real time. When a value in a data item deviates from other values ​​beyond a preset reasonable error range, arbitration is conducted using the majority consensus principle, and this value is defined as an outlier. Outliers are then corrected using methods based on time series forecasting or interpolation from adjacent data sources to ensure the continuity of the data sequence, resulting in the multi-dimensional time-varying data verified by the data sources.

[0041] Finally, the multi-dimensional time-varying data verified by the data sources is weighted and fused using preset dynamic reliability weights to generate a fused time-varying sequence for each power resource. During the fusion process, for each data point in time, the final fused value is equal to the weighted average of the values ​​reported by each data source. This generates a fused time-varying sequence for each resource, which serves as the standard input to the source-load potential assessment model, effectively reducing the interference of noise and errors on the assessment results.

[0042] Step S3: Using a preset source-load potential assessment model, capture the spatiotemporal correlation between the fused time-varying sequence and the power resources, and obtain the adjustment potential boundary and constraints of the power resources within the prediction time window.

[0043] This application embodiment constructs a pre-trained, lightweight, and efficient source-load potential assessment model to analyze the generated fused time-varying sequence, thereby extracting and quantifying the regulation potential information therein, and quickly and accurately obtaining the regulation potential boundary of each power resource in the future time period.

[0044] The source load potential assessment model adopts a grouped query attention mechanism as the core computing unit. This mechanism learns from a more complex and accurate teacher model through knowledge distillation technology, which enables it to capture the complex spatiotemporal relationship between time-varying data and adjustment potential. This results in a source load potential assessment model that combines high inference speed and high assessment accuracy, meeting the real-time requirements of online regulation.

[0045] The source load potential assessment model mainly consists of a group query attention mechanism and a loss function. The group query attention mechanism is mainly constructed using the provided teacher model, and the specific construction steps are as follows: 1. Obtain the model weights of the multi-head attention module of the teacher model, wherein the multi-head attention module contains a third threshold number of attention heads; 2. Divide the third threshold number of attention heads into multiple groups, each group containing a fourth threshold number of consecutive attention heads; wherein, each group contains the same number of attention heads, and each group is only used to process data from one of the acquisition layers; 3. For each group, the model weights of the fourth threshold attention heads contained therein are averaged and aggregated to generate a first model weight for a corresponding group query attention head; 4. The group query attention mechanism is constituted by the group query attention heads corresponding to all groups.

[0046] Specifically, extract all model weights of the multi-head attention module in the teacher model, assuming that this module contains... N Each head of attention Corresponding to three sets of weight matrices: Query weight Key weight Value weight Determine the number of groups in the student model (i.e., the source load potential assessment model). The teacher model The order of size is as follows: There are 1 group, each containing 1 group. A continuous header. For each group , and include The weight matrices corresponding to each head are averaged and aggregated to generate the weights of a group head in the source load potential assessment model: , , ;in, The index representing the teacher model's attention head; Indicates the allocation to the first The set of indices of all attention heads for each group; , , The respective sources and loads potential assessment models represent the first... The query, key, and value weight matrix for each grouped query attention header; Indicates the feature dimension of the model; This represents the key dimensions of each attention head.

[0047] The above polymerization method yielded Construct a grouped query attention mechanism for the source load potential assessment model.

[0048] The other model weights of the teacher model are transferred to the source load potential assessment model, and then the group query attention mechanism is trained through a preset knowledge distillation strategy to obtain the trained source load potential assessment model.

[0049] The knowledge distillation process is as follows: 1. Obtain the model weights of all modules (e.g., feedforward networks) in the teacher model except for the first model weights to obtain the second model weights; assign the second model weights to the corresponding parts of the source load potential assessment model, and fix all model weights transferred from the teacher model except for the first model weights; 2. Construct a loss function using the knowledge distillation strategy. The loss function includes a first loss term and a second loss term. The first loss term is used to calculate the difference in output distribution between the teacher model and the source load potential assessment model for the same training sample. The second loss term is used to calculate the difference in output distribution between the multi-head attention module and the grouped query attention mechanism. 3. Using the teacher model as the teacher and the source load potential assessment model as the student, and with fixed model weights, the first model weights of the grouped query attention mechanism are trained using the loss function with the goal of minimizing the attention weight matrix of the teacher model and the source load potential assessment model.

[0050] The loss function is used to simultaneously supervise the output distribution and internal attention features of the source-load potential assessment model, and its specific expression is: ; in, and These are hyperparameters used to balance the weights of the output distribution loss and the attention feature loss in the total loss, respectively. Distillation loss for attention features Let the loss function be... This is to account for distillation losses.

[0051] Output distillation loss Use with temperature T The KL divergence forces students to learn the teacher's output probability distribution. The specific expression for the output distillation loss is: ; in, , Teacher and student models at different temperatures The output probability distribution is as follows. The index represents the output category. In the potential assessment task, the output is discretized into multiple potential interval categories. and These represent the teacher and student models for the categories, respectively. The predicted probability; It is a temperature parameter used to soften the probability distribution, making knowledge transfer smoother.

[0052] Attention feature distillation loss The mean squared error is used to minimize the difference in the attention weight matrix between the teacher model and the source load potential assessment model, and its specific expression is as follows: ; in, For the number of floors, It is the student's number Layer Group attention matrix, It is the teacher's number Corresponding groups in the layer The average of multiple head attention matrices.

[0053] Therefore, the teacher model is a fixed guide, and the source load potential assessment model is a trainable object. The weights of the grouped query attention mechanism are updated by minimizing the loss function through an optimizer, so that its behavior approximates the teacher model.

[0054] Furthermore, the knowledge distillation strategy employs a progressive training strategy, including a first training phase and a second training phase. These two phases represent increasing difficulty to comprehensively improve the model's robustness and generalization ability.

[0055] In the first training phase, the source load potential assessment model is trained using training samples with a completeness higher than a first threshold and a prediction time window length lower than a second threshold. In this phase, the source load potential assessment model outputs the adjustment potential boundaries of all time segments using a parallel inference mode; wherein, the time segment is the time segment within the set prediction time window.

[0056] In the second training phase, the source load potential assessment model that has passed the first training phase is trained using training samples with a completeness lower than the first threshold or a prediction time window length higher than the second threshold. In this phase, the source load potential assessment model adopts a stepwise inference mode to calculate the adjustment potential boundary of the next time segment based on the output value of the previous time segment.

[0057] The first training phase employs parallel inference mode, using high-quality, short-cycle training samples. The goal is to rapidly establish the model's basic evaluation capabilities. Sample selection criteria include: data completeness exceeding a first threshold (e.g., a missing rate of less than 5% for key time-varying data within a single sample), and a prediction window length shorter than a second threshold (e.g., a total rolling window duration not exceeding 15 minutes). The first threshold is determined based on historical data quality statistics to cover most periods of good operating conditions; the second threshold is set based on the typical response cycle of the power system to ultra-short-term regulation. In this phase, the student model is configured for parallel inference mode. In this mode, the model's output layer outputs all data within the future rolling window at once. The potential boundary value of a time segment. That is, given the current time. Fusion of time-varying data sequences The model performs one forward propagation and directly outputs a vector. ,in Including cross-section The upward / downward adjustment potential boundary.

[0058] The second training phase employs a stepwise inference mode, introducing more challenging training samples to improve the model's evaluation performance under non-ideal conditions. The samples used meet the following criteria: data integrity is below a first threshold, simulating scenarios with missing data or high noise levels, or the prediction time window length is longer than a second threshold (e.g., a rolling window of one hour). The student model that converged in the previous phase is used as the initialization starting point for this phase, and its inference mode is switched to stepwise inference mode. In this mode, the model operates in an autoregressive manner: first, based on the current time... Input The model predicts the next cross-section. Potential boundaries Subsequently, the predicted (or some hidden state representation thereof) as additional input information, and New external data that may be acquired at any time (which may use real or simulated values ​​during training) are combined and used as input for prediction. potential boundary of the cross section This process is repeated recursively until the prediction sequence for the entire window is generated. This mode forces the model to learn to handle long-range dependencies, compensate for uncertainties in the input data, and ensure the temporal coherence and reasonableness of multi-step predictions by explicitly feeding the prediction results of the previous section back to the model. Training at this stage effectively enhances the model's robustness and accuracy in handling complex, long-term evaluation tasks.

[0059] Based on the input fused time-varying sequence, the constraints of each power resource at each time section within the prediction time window are analyzed by the trained source-load potential assessment model.

[0060] Firstly, in practical applications, the trained source-load potential assessment model uses the grouped query attention mechanism as its core computational unit, receiving the fused time-varying sequence as input. Internally, the grouped query attention mechanism first calculates the query, key, and value vectors: ; in, , , The first The query, key, and value matrix corresponding to each group; The fused time-varying sequence is referred to as .

[0061] Then, the group attention output is calculated. Head g : ; in, This is a scaling factor used to prevent the gradient of the Softmax function from vanishing due to an excessively large dot product result.

[0062] The outputs of all groups are concatenated and linearly transformed before being processed by subsequent network layers to finally output structured potential assessment results. This process efficiently captures the time-series dependencies and correlations between different feature dimensions in the input sequence, providing direct data for obtaining the regulation potential boundaries and constraints of the power resources within the prediction time window.

[0063] First, the source-load potential assessment model analyzes the constraints of each power resource at each time segment based on the fused time-varying sequence, including real-time operating status constraints, physical regulation capability constraints, and external environmental incentive constraints.

[0064] From the input fused time-varying sequence, three key constraints affecting the potential assessment are separated and quantified. The real-time operating state constraints are the instantaneous adjustment margins determined by the current operating point of the power resource; for example, for a power unit with a rated power of... And the current output is The instantaneous upward adjustment range of the generator set is limited by The downward adjustment space is limited by , Minimize the effort required for the technology.

[0065] The physical regulation capability constraints are the regulation capabilities and process limitations determined by the inherent physical characteristics of the equipment, including maximum charging and discharging power (for energy storage), maximum interruption power and duration (for load), ramp rate (maximum power change per unit time), and total energy for sustainable regulation (such as the capacity limit of energy storage).

[0066] The external environmental incentives and constraints refer to the willingness to regulate and the economic boundaries driven by external factors such as electricity market signals, policies, or contractual terms. The most typical examples are real-time electricity prices or ancillary service clearing prices. For instance, resources are only economically adjustable when the benefits of regulation exceed their costs.

[0067] Then, based on real-time operating state constraints and physical regulation capacity constraints, the maximum theoretically adjustable power range that each power resource can provide at each time segment is calculated. This is a dynamic calculation process that considers time coupling. Taking the upward regulation potential as an example, its theoretical maximum value... It is not only constrained by the instantaneous margin at the current moment, but also by the time from the current moment to... The cumulative impact of the maximum permissible climb rate within a given period must be considered, while ensuring that the total energy consumed or increased during the entire adjustment process does not exceed its sustainable energy. This calculation can be formally represented as solving a constrained optimization problem, with the model approximating the solution through its internal network. Therefore, the specific formula for calculating the upper bound of the maximum adjustable power range is: ; in, Time section The upper bound of the instantaneous power margin, From the current moment arrive The maximum cumulative increase in power based on the maximum climbing rate. For the current remaining sustainable energy regulation, To adjust the duration.

[0068] The lower bound of the maximum adjustable power range The calculation is the same as above.

[0069] Based on external environmental incentive constraints, the theoretically maximum adjustable power range is modified for incentive compatibility to obtain the effective adjustable power range for each time segment considering economic efficiency or market signal response.

[0070] Specifically, the maximum adjustable power range reflects technical feasibility, while the effective adjustable power range reflects cost feasibility. Therefore, the source-load potential assessment model introduces external environmental incentive constraints to modify the theoretical range.

[0071] For example, define a value related to real-time electricity prices. Related excitation response function ,in As the cost threshold for this resource to participate in regulation, when the real-time electricity price Less than the cost threshold At times, resources may be unwilling to provide regulation services, and the effective regulation potential approaches zero. The model learns this nonlinear relationship, mapping the theoretical potential to the effective potential. The effective adjustable power range is expressed as: ; in, For power resources in time cross-section t k The effective adjustable power range, For power resources in time cross-section t k The maximum adjustable power range, f up For upward excitation function, This is the lower bound of the effective adjustable power range. This is the lower bound of the maximum adjustable power range. f down This is an upward activation function. The specific form of the activation function is learned by the model from historical market feedback information during training.

[0072] The upper and lower limits of the effective adjustable power range are used to determine the upward and downward adjustment potential boundaries for each time segment.

[0073] Specifically, the upper and lower bounds of the obtained effective adjustable power range are directly quantified into the evaluation results for each time segment, and the calculation formula is as follows: ; in, For the downward adjustment potential boundary, This is the upward adjustment potential boundary.

[0074] These two boundary values ​​together define the power regulation range that the power resources can safely and economically provide at that time interval. .

[0075] Step S4: Based on the adjustment potential boundary and the constraint conditions, analyze the adjustable parameters corresponding to each time segment within the prediction time window to obtain the source-load adjustment potential assessment result of the power resources.

[0076] The previously calculated parameters of the regulation potential boundaries and constraints, scattered across various time segments, are integrated into a structured, time-evolving complete description, namely, the source-load regulation potential envelope of each power resource. This envelope represents the source-load regulation potential assessment result for each power resource. This envelope is the final presentation form of the source-load regulation potential assessment result obtained in this application embodiment, providing the dispatch system with a directly usable and comprehensive list of regulation capabilities.

[0077] Specifically, for a coverage N A rolling evaluation window spanning multiple time segments; the model outputs a time-series vector for each evaluated object. : ; Each time section t k Corresponding adjustable parameters It is a subvector containing four key indicators, with the specific expression as follows: ; To sustainably regulate total energy, it represents the time span from the current moment to that time section. After considering the continuous adjustment process, the net adjustment energy that resources can still provide can be obtained by subtracting the energy expected to be consumed (or increased) up to this section from the initial sustainable energy. The maximum allowable adjustment rate represents the effective adjustment rate that resources are actually available near this cross section, which may be lower than the nominal maximum ramp rate due to equipment operating conditions or external limitations.

[0078] As can be seen from the above formula, the key indicators are the maximum adjustable power upward, the maximum adjustable power downward, the total energy for sustainable adjustment, and the maximum allowable adjustment rate.

[0079] In summary, the machine-readable time-series vector form of the source-load regulation potential envelope characterizes the spatiotemporal evolution of resource regulation capacity over a future period. It not only provides the amount that can be regulated, but also clarifies the duration and speed of regulation, providing accurate and reliable quantitative input for the power system to carry out multi-timescale, multi-objective safety-constrained economic dispatch.

[0080] Implementing the embodiments of this application has the following beneficial effects: To address the heterogeneity conflicts of multi-dimensional time-varying data from different data sources, this embodiment verifies the data sources before weighted fusion. Time-varying data is managed hierarchically based on data characteristics and collection urgency, and the reliability of the data sources is verified. The fusion generates single, continuous, and reliable time-series data that best reflects the true state of resources, along with the corresponding collection levels for each data point. Adaptive hierarchical collection is achieved through these collection levels. Then, based on a preset source-load potential assessment model, the current and future adjustment capabilities of each resource are accurately evaluated, yielding potential values ​​that more closely reflect the actual dispatching needs of the power grid and meet the real-time requirements of online control. Furthermore, constructing the source-load potential assessment model through the collection levels significantly reduces model parameters, lowers computational complexity, and improves the efficiency of adjustment potential assessment. Finally, the output source-load adjustment potential envelope unifies multiple physical constraints such as power, energy, and rate into a time-series representation, providing a high-precision and refined decision-making basis for power grid dispatch.

[0081] Second Embodiment Furthermore, in order to implement the source-load regulation potential assessment system based on time-varying data management corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects, Figure 2 A structural diagram of a source-load regulation potential assessment system based on time-varying data management is provided. For ease of explanation, only the parts relevant to this embodiment are shown. The source-load regulation potential assessment system based on time-varying data management provided in this application embodiment includes: The data acquisition module 201 is used to acquire multi-dimensional time-varying data of various power resources; wherein, the multi-dimensional time-varying data includes static attribute data, operating status data, environmental data and power market data of various power resources.

[0082] In this embodiment of the application, multi-dimensional time-varying data for source-load regulation potential assessment is first obtained. The multi-dimensional time-varying data includes static attribute data, operating status data, environmental data, and electricity market data of each power resource.

[0083] The principle behind assessing the regulation potential of power resources lies in the fact that the regulation capacity of resources depends not only on their inherent equipment parameters but also on the dynamic constraints of their real-time operating conditions and the surrounding environment and market conditions. Static attribute data of power resources refers to the inherent parameters of equipment that do not change or change slowly over time, such as the rated power and ramp rate of generator sets, the total capacity and maximum charging and discharging power of energy storage devices, and the rated power and maximum allowable interruption duration of interruptible loads; these data constitute the physical boundary basis of the resource's regulation capacity. Operating status data of power resources refers to rapidly changing data collected in real-time or near real-time, reflecting the current operating conditions of the resources, such as the current actual power generation of photovoltaic power plants or wind farms, the current state of charge of energy storage systems, the current power consumption and operating conditions of adjustable loads, and the current battery charge and charging status of electric vehicles; these data determine the available regulation margin of the resources at the current moment. The environmental data and electricity market data refer to dynamic information from outside the resources that affects their regulation behavior or regulation value. For example, ultra-short-term wind speed and solar intensity forecasts for the assessment time, ambient temperature, real-time nodal marginal electricity prices, ancillary service market clearing price signals, and regulation demand instructions issued by the power grid. These data reflect the external driving factors and constraints that affect the willingness and feasibility of resource regulation.

[0084] The data fusion module 202 is used to perform data source verification and weighted fusion on the multi-dimensional time-varying data to obtain a fused time-varying sequence and a collection level related to the data collection frequency.

[0085] In this embodiment, hierarchical management is implemented based on the inherent characteristics of the data and system requirements, and a single, continuous and reliable fused time-varying sequence that best reflects the true state of resources is generated through multi-source cross-validation and adaptive weighting strategies.

[0086] First, the multi-dimensional time-varying data is divided into three acquisition levels according to the preset data change rate and acquisition urgency. These three acquisition levels are the millisecond-level urgent and critical layer, the minute-level dynamic change layer, and the hour-level slowly changing background layer.

[0087] Specifically, the emergency critical layer includes event-type data that requires millisecond-level response, such as grid frequency deviation and circuit breaker tripping signals; the dynamic change layer includes continuously changing operational data ranging from seconds to minutes, such as photovoltaic output, load power, energy storage status of charge, and real-time electricity price; and the gradual change background layer includes status and environmental data updated on hourly or longer time scales, such as weather trend forecasts, equipment health indicators, and medium- and long-term market expectations.

[0088] The aforementioned hierarchical division forms the basis for implementing differentiated data management. For each acquisition level, the acquisition frequency and sampling density are adjusted based on the current control mode of the power system and the historical trends of various time-varying data.

[0089] For example, when the power system needs to make rapid frequency response, the data acquisition cycle that is strongly related to frequency regulation in the dynamic change layer will be automatically compressed; while for temperature data with slow historical changes, the acquisition cycle can be appropriately extended, thereby optimizing system communication and computing resources while ensuring data timeliness.

[0090] Then, the multi-dimensional time-varying data is validated to assess the reliability of each data source. For each type of power resource, similar time-varying data from at least three independent data sources are acquired in parallel. These independent data sources may include local equipment monitoring systems, station-level control systems, and power grid dispatch automation systems, etc.

[0091] Based on the data source type, data real-time performance, and historical accuracy records, a corresponding dynamic reliability weight is calculated for each independent data source. The data types include direct measurement and indirect estimation, with direct measurement having a higher weight factor than indirect estimation. Data real-time performance is related to the data reporting latency of similar time-varying data; the lower the data reporting latency, the higher the weight factor for data real-time performance. Historical accuracy records are obtained by comparing with a set benchmark data set.

[0092] The system compares the same data item from different time-varying data sources in real time. When a value in a data item deviates from other values ​​beyond a preset reasonable error range, arbitration is conducted using the majority consensus principle, and this value is defined as an outlier. Outliers are then corrected using methods based on time series forecasting or interpolation from adjacent data sources to ensure the continuity of the data sequence, resulting in the multi-dimensional time-varying data verified by the data sources.

[0093] Finally, the multi-dimensional time-varying data verified by the data sources is weighted and fused using preset dynamic reliability weights to generate a fused time-varying sequence for each power resource. During the fusion process, for each data point in time, the final fused value is equal to the weighted average of the values ​​reported by each data source. This generates a fused time-varying sequence for each resource, which serves as the standard input to the source-load potential assessment model, effectively reducing the interference of noise and errors on the assessment results.

[0094] The adjustment boundary prediction module 203 is used to capture the spatiotemporal correlation between the fused time-varying sequence and the power resources through a preset source-load potential assessment model, and obtain the adjustment potential boundary and constraints of the power resources within the prediction time window.

[0095] In this embodiment of the application, the source load potential evaluation model is constructed based on a preset grouped query attention mechanism and loss function; Based on the fused time-varying sequence, the constraints of each power resource at each time segment within the prediction time window are analyzed by the source-load potential assessment model; wherein, the constraints include real-time operating status constraints, physical regulation capability constraints, and external environmental incentive constraints. Based on the real-time operating status constraints and physical adjustment capability constraints, the maximum adjustable power range of each power resource at each time segment is calculated. Based on the external environmental incentive constraints, and by taking into account historical feedback information from the electricity market, the maximum adjustable power range is adjusted to obtain the effective adjustable power range for each time segment. The upper and lower limits of the effective adjustable power range are used to determine the adjustment potential boundary for each time segment.

[0096] The adjustable parameter generation module 204 is used to analyze the adjustable parameters corresponding to each time segment within the prediction time window based on the adjustment potential boundary and the constraint conditions, so as to obtain the source-load adjustment potential assessment result of the power resources.

[0097] In this embodiment, the previously calculated parameters of the regulation potential boundaries and constraints, scattered across various time segments, are integrated into a structured, time-evolving complete description, namely, the source-load regulation potential envelope of each power resource. This envelope represents the source-load regulation potential assessment result for each power resource. This envelope is the final presentation of the source-load regulation potential assessment result obtained in this embodiment, providing the dispatch system with a directly usable and comprehensive list of regulation capabilities.

[0098] Specifically, for a coverage N A rolling evaluation window spanning multiple time segments; the model outputs a time-series vector for each evaluated object. : ; Each time section t k Corresponding adjustable parameters It is a subvector containing four key indicators, with the specific expression as follows: ; To sustainably regulate total energy, it represents the time span from the current moment to that time section. After considering the continuous adjustment process, the net adjustment energy that resources can still provide can be obtained by subtracting the energy expected to be consumed (or increased) up to this section from the initial sustainable energy. The maximum allowable adjustment rate represents the effective adjustment rate that resources are actually available near this cross section, which may be lower than the nominal maximum ramp rate due to equipment operating conditions or external limitations.

[0099] As can be seen from the above formula, the key indicators are the maximum adjustable power upward, the maximum adjustable power downward, the total energy for sustainable adjustment, and the maximum allowable adjustment rate.

[0100] In summary, the machine-readable time-series vector form of the source-load regulation potential envelope characterizes the spatiotemporal evolution of resource regulation capacity over a future period. It not only provides the amount that can be regulated, but also clarifies the duration and speed of regulation, providing accurate and reliable quantitative input for the power system to carry out multi-timescale, multi-objective safety-constrained economic dispatch.

[0101] In some embodiments, the adjustment boundary prediction module 203 specifically comprises: By constructing a pre-trained, lightweight, and efficient source-load potential assessment model, the generated fused time-varying sequence is analyzed to extract and quantify the regulation potential information, thereby quickly and accurately obtaining the regulation potential boundaries of each power resource in future time periods.

[0102] The source load potential assessment model adopts a grouped query attention mechanism as the core computing unit. This mechanism learns from a more complex and accurate teacher model through knowledge distillation technology, which enables it to capture the complex spatiotemporal relationship between time-varying data and adjustment potential. This results in a source load potential assessment model that combines high inference speed and high assessment accuracy, meeting the real-time requirements of online regulation.

[0103] The source load potential assessment model mainly consists of a group query attention mechanism and a loss function. The group query attention mechanism is mainly constructed using the provided teacher model, and the specific construction steps are as follows: 1. Obtain the model weights of the multi-head attention module of the teacher model, wherein the multi-head attention module contains a third threshold number of attention heads; 2. Divide the third threshold number of attention heads into multiple groups, each group containing a fourth threshold number of consecutive attention heads; wherein, each group contains the same number of attention heads, and each group is only used to process data from one of the acquisition layers; 3. For each group, the model weights of the fourth threshold attention heads contained therein are averaged and aggregated to generate a first model weight for a corresponding group query attention head; 4. The group query attention mechanism is constituted by the group query attention heads corresponding to all groups.

[0104] Specifically, extract all model weights of the multi-head attention module in the teacher model, assuming that this module contains... N Each head of attention Corresponding to three sets of weight matrices: Query weight Key weight Value weight Determine the number of groups in the student model (i.e., the source load potential assessment model). The teacher model The order of size is as follows: There are 1 group, each containing 1 group. A continuous header. For each group , and include The weight matrices corresponding to each head are averaged and aggregated to generate the weights of a group head in the source load potential assessment model: , , ;in, The index representing the teacher model's attention head; Indicates the allocation to the first The set of indices of all attention heads for each group; , , The respective sources and loads potential assessment models represent the first... The query, key, and value weight matrix for each grouped query attention header; Indicates the feature dimension of the model; This represents the key dimensions of each attention head.

[0105] The above polymerization method yielded Construct a grouped query attention mechanism for the source load potential assessment model.

[0106] The other model weights of the teacher model are transferred to the source load potential assessment model, and then the group query attention mechanism is trained through a preset knowledge distillation strategy to obtain the trained source load potential assessment model.

[0107] The knowledge distillation process is as follows: 1. Obtain the model weights of all modules (e.g., feedforward networks) in the teacher model except for the first model weights to obtain the second model weights; assign the second model weights to the corresponding parts of the source load potential assessment model, and fix all model weights transferred from the teacher model except for the first model weights; 2. Construct a loss function using the knowledge distillation strategy. The loss function includes a first loss term and a second loss term. The first loss term is used to calculate the difference in output distribution between the teacher model and the source load potential assessment model for the same training sample. The second loss term is used to calculate the difference in output distribution between the multi-head attention module and the grouped query attention mechanism. 3. Using the teacher model as the teacher and the source load potential assessment model as the student, and with fixed model weights, the first model weights of the grouped query attention mechanism are trained using the loss function with the goal of minimizing the attention weight matrix of the teacher model and the source load potential assessment model.

[0108] The loss function is used to simultaneously supervise the output distribution and internal attention features of the source-load potential assessment model, and its specific expression is: ; in, and These are hyperparameters used to balance the weights of the output distribution loss and the attention feature loss in the total loss, respectively. Distillation loss for attention features Let the loss function be... This is to account for distillation losses.

[0109] Output distillation loss Use with temperature T The KL divergence forces students to learn the teacher's output probability distribution. The specific expression for the output distillation loss is: ; in, , Teacher and student models at different temperatures The output probability distribution is as follows. The index represents the output category. In the potential assessment task, the output is discretized into multiple potential interval categories. and These represent the teacher and student models for the categories, respectively. The predicted probability; It is a temperature parameter used to soften the probability distribution, making knowledge transfer smoother.

[0110] Attention feature distillation loss The mean squared error is used to minimize the difference in the attention weight matrix between the teacher model and the source load potential assessment model, and its specific expression is as follows: ; in, For the number of floors, It is the student's number Layer Group attention matrix, It is the teacher's number Corresponding groups in the layer The average of multiple head attention matrices.

[0111] Therefore, the teacher model is a fixed guide, and the source load potential assessment model is a trainable object. The weights of the grouped query attention mechanism are updated by minimizing the loss function through an optimizer, so that its behavior approximates the teacher model.

[0112] Furthermore, the knowledge distillation strategy employs a progressive training strategy, including a first training phase and a second training phase. These two phases represent increasing difficulty to comprehensively improve the model's robustness and generalization ability.

[0113] In the first training phase, the source load potential assessment model is trained using training samples with a completeness higher than a first threshold and a prediction time window length lower than a second threshold. In this phase, the source load potential assessment model outputs the adjustment potential boundaries of all time segments using a parallel inference mode; wherein, the time segment is the time segment within the set prediction time window.

[0114] In the second training phase, the source load potential assessment model that has passed the first training phase is trained using training samples with a completeness lower than the first threshold or a prediction time window length higher than the second threshold. In this phase, the source load potential assessment model adopts a stepwise inference mode to calculate the adjustment potential boundary of the next time segment based on the output value of the previous time segment.

[0115] The first training phase employs parallel inference mode, using high-quality, short-cycle training samples. The goal is to rapidly establish the model's basic evaluation capabilities. Sample selection criteria include: data completeness exceeding a first threshold (e.g., a missing rate of less than 5% for key time-varying data within a single sample), and a prediction window length shorter than a second threshold (e.g., a total rolling window duration not exceeding 15 minutes). The first threshold is determined based on historical data quality statistics to cover most periods of good operating conditions; the second threshold is set based on the typical response cycle of the power system to ultra-short-term regulation. In this phase, the student model is configured for parallel inference mode. In this mode, the model's output layer outputs all data within the future rolling window at once. The potential boundary value of a time segment. That is, given the current time. Fusion of time-varying data sequences The model performs one forward propagation and directly outputs a vector. ,in Including cross-section The upward / downward adjustment potential boundary.

[0116] The second training phase employs a stepwise inference mode, introducing more challenging training samples to improve the model's evaluation performance under non-ideal conditions. The samples used meet the following criteria: data integrity is below a first threshold, simulating scenarios with missing data or high noise levels, or the prediction time window length is longer than a second threshold (e.g., a rolling window of one hour). The student model that converged in the previous phase is used as the initialization starting point for this phase, and its inference mode is switched to stepwise inference mode. In this mode, the model operates in an autoregressive manner: first, based on the current time... Input The model predicts the next cross-section. Potential boundaries Subsequently, the predicted (or some hidden state representation thereof) as additional input information, and New external data that may be acquired at any time (which may use real or simulated values ​​during training) are combined and used as input for prediction. potential boundary of the cross section This process is repeated recursively until the prediction sequence for the entire window is generated. This mode forces the model to learn to handle long-range dependencies, compensate for uncertainties in the input data, and ensure the temporal coherence and reasonableness of multi-step predictions by explicitly feeding the prediction results of the previous section back to the model. Training at this stage effectively enhances the model's robustness and accuracy in handling complex, long-term evaluation tasks.

[0117] Based on the input fused time-varying sequence, the constraints of each power resource at each time section within the prediction time window are analyzed by the trained source-load potential assessment model.

[0118] Firstly, in practical applications, the trained source-load potential assessment model uses the grouped query attention mechanism as its core computational unit, receiving the fused time-varying sequence as input. Internally, the grouped query attention mechanism first calculates the query, key, and value vectors: ; in, , , The first The query, key, and value matrix corresponding to each group; The fused time-varying sequence is referred to as .

[0119] Then, the group attention output is calculated. Head g : ; in, This is a scaling factor used to prevent the gradient of the Softmax function from vanishing due to an excessively large dot product result.

[0120] The outputs of all groups are concatenated and linearly transformed before being processed by subsequent network layers to finally output structured potential assessment results. This process efficiently captures the time-series dependencies and correlations between different feature dimensions in the input sequence, providing direct data for obtaining the regulation potential boundaries and constraints of the power resources within the prediction time window.

[0121] First, the source-load potential assessment model analyzes the constraints of each power resource at each time segment based on the fused time-varying sequence, including real-time operating status constraints, physical regulation capability constraints, and external environmental incentive constraints.

[0122] From the input fused time-varying sequence, three key constraints affecting the potential assessment are separated and quantified. The real-time operating state constraints are the instantaneous adjustment margins determined by the current operating point of the power resource; for example, for a power unit with a rated power of... And the current output is The instantaneous upward adjustment range of the generator set is limited by The downward adjustment space is limited by , Minimize the effort required for the technology.

[0123] The physical regulation capability constraints are the regulation capabilities and process limitations determined by the inherent physical characteristics of the equipment, including maximum charging and discharging power (for energy storage), maximum interruption power and duration (for load), ramp rate (maximum power change per unit time), and total energy for sustainable regulation (such as the capacity limit of energy storage).

[0124] The external environmental incentives and constraints refer to the willingness to regulate and the economic boundaries driven by external factors such as electricity market signals, policies, or contractual terms. The most typical examples are real-time electricity prices or ancillary service clearing prices. For instance, resources are only economically adjustable when the benefits of regulation exceed their costs.

[0125] Then, based on real-time operating state constraints and physical regulation capacity constraints, the maximum theoretically adjustable power range that each power resource can provide at each time segment is calculated. This is a dynamic calculation process that considers time coupling. Taking the upward regulation potential as an example, its theoretical maximum value... It is not only constrained by the instantaneous margin at the current moment, but also by the time from the current moment to... The cumulative impact of the maximum permissible climb rate within a given period must be considered, while ensuring that the total energy consumed or increased during the entire adjustment process does not exceed its sustainable energy. This calculation can be formally represented as solving a constrained optimization problem, with the model approximating the solution through its internal network. Therefore, the specific formula for calculating the upper bound of the maximum adjustable power range is: ; in, Time section The upper bound of the instantaneous power margin, From the current moment arrive The maximum cumulative increase in power based on the maximum climbing rate. For the current remaining sustainable energy regulation, To adjust the duration.

[0126] The lower bound of the maximum adjustable power range The calculation is the same as above.

[0127] Based on external environmental incentive constraints, the theoretically maximum adjustable power range is modified for incentive compatibility to obtain the effective adjustable power range for each time segment considering economic efficiency or market signal response.

[0128] Specifically, the maximum adjustable power range reflects technical feasibility, while the effective adjustable power range reflects cost feasibility. Therefore, the source-load potential assessment model introduces external environmental incentive constraints to modify the theoretical range.

[0129] For example, define a value related to real-time electricity prices. Related excitation response function ,in As the cost threshold for this resource to participate in regulation, when the real-time electricity price Less than the cost threshold At times, resources may be unwilling to provide regulation services, and the effective regulation potential approaches zero. The model learns this nonlinear relationship, mapping the theoretical potential to the effective potential. The effective adjustable power range is expressed as: ; in, For power resources in time cross-section t k The effective adjustable power range, For power resources in time cross-section t k The maximum adjustable power range, f up For upward excitation function, This is the lower bound of the effective adjustable power range. This is the lower bound of the maximum adjustable power range. f down This is an upward activation function. The specific form of the activation function is learned by the model from historical market feedback information during training.

[0130] The upper and lower limits of the effective adjustable power range are used to determine the upward and downward adjustment potential boundaries for each time segment.

[0131] Specifically, the upper and lower bounds of the obtained effective adjustable power range are directly quantified into the evaluation results for each time segment, and the calculation formula is as follows: ; in, For the downward adjustment potential boundary, This is the upward adjustment potential boundary.

[0132] These two boundary values ​​together define the power regulation range that the power resources can safely and economically provide at that time interval. .

[0133] Implementing the embodiments of this application has the following beneficial effects: To address the heterogeneity conflicts of multi-dimensional time-varying data from different data sources, this embodiment verifies the data sources before weighted fusion. Time-varying data is managed hierarchically based on data characteristics and collection urgency, and the reliability of the data sources is verified. The fusion generates single, continuous, and reliable time-series data that best reflects the true state of resources, along with the corresponding collection levels for each data point. Adaptive hierarchical collection is achieved through these collection levels. Then, based on a preset source-load potential assessment model, the current and future adjustment capabilities of each resource are accurately evaluated, yielding potential values ​​that more closely reflect the actual dispatching needs of the power grid and meet the real-time requirements of online control. Furthermore, constructing the source-load potential assessment model through the collection levels significantly reduces model parameters, lowers computational complexity, and improves the efficiency of adjustment potential assessment. Finally, the output source-load adjustment potential envelope unifies multiple physical constraints such as power, energy, and rate into a time-series representation, providing a high-precision and refined decision-making basis for power grid dispatch.

[0134] Furthermore, Figure 3 This is a structural diagram of a terminal device provided in one embodiment of this application. Figure 3 As shown, the terminal device 3 of this embodiment includes: at least one processor 30 (in... Figure 3 (Only one is shown in the image) and a memory 31 and a computer program 32 stored in the memory 31 and executable on the at least one processor, wherein when the processor 30 executes the computer program 32, it can implement the steps of a source-load regulation potential assessment method based on time-varying data management as described in any one of the embodiments of this application.

[0135] The terminal device 3 may be a computing device such as a desktop computer, a cloud server, or a laptop computer, and the computing device may include, but is not limited to, a processor 30 and a memory 31. Figure 3 This is merely an example of terminal device 3 and does not constitute a limitation on terminal device 3. It may include more or fewer components than those shown in the figure.

[0136] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. In particular, it should be noted that any modifications, equivalent substitutions, or improvements made by those skilled in the art within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for assessing source-load regulation potential based on time-varying data management, characterized in that, include: Acquire multi-dimensional time-varying data for each power resource; wherein, the multi-dimensional time-varying data includes static attribute data, operating status data, environmental data, and power market data for each power resource; The multi-dimensional time-varying data is validated and weighted to obtain a fused time-varying sequence and a collection level related to the data collection frequency. By using a pre-defined source-load potential assessment model, the spatiotemporal correlation between the fused time-varying sequence and the power resources is captured, and the adjustment potential boundary and constraints of the power resources within the prediction time window are obtained; wherein, the source-load potential assessment model is constructed using the acquisition level; Based on the adjustment potential boundary and the constraints, the adjustable parameters corresponding to each time segment within the prediction time window are analyzed to obtain the source-load adjustment potential assessment results of the power resources.

2. The source-load regulation potential assessment method based on time-varying data management according to claim 1, characterized in that, The process of performing data source verification and weighted fusion on the multi-dimensional time-varying data to obtain the fused time-varying sequence and the acquisition level related to the data acquisition frequency is as follows: Based on preset data change rates and collection urgency, the multi-dimensional time-varying data is hierarchically divided to obtain the collection levels of the multi-dimensional time-varying data; wherein, the collection levels include a millisecond-level urgent critical layer, a minute-level dynamic change layer, and an hour-level slowly changing background layer; Based on the current control mode of the power system and the historical trend of the multi-dimensional time-varying data, adjust the acquisition frequency and sampling density of each acquisition level; By using preset dynamic credibility weights, the multi-dimensional time-varying data verified by the data source is weighted and fused to generate the fused time-varying sequence for each power resource.

3. The source-load regulation potential assessment method based on time-varying data management according to claim 2, characterized in that, The multi-dimensional time-varying data after data source verification is specifically as follows: Obtain similar time-varying data from at least a first threshold number of independent data sources for each power resource, and determine the dynamic reliability weight of the independent data sources based on the data type, real-time performance, and historical accuracy records of the independent data sources. The data types include direct measurement type and indirect estimation type, with the weighting factor of direct measurement type being greater than that of indirect estimation type; the data real-time performance is related to the data reporting delay of the same type of time-varying data, and the smaller the data reporting delay, the higher the weighting factor of the data real-time performance. Compare the same data item of each of the same time-varying data. If the difference between a value in the same data item and other values ​​exceeds a preset reasonable error range, then the value is defined as an outlier. The outliers are repaired, and the multi-dimensional time-varying data after data source verification is obtained based on the repair results.

4. The source-load regulation potential assessment method based on time-varying data management according to claim 1, characterized in that, The method involves using a pre-defined source-load potential assessment model to capture the spatiotemporal correlation between the fused time-varying sequence and the power resources, thereby obtaining the adjustment potential boundary and constraints of the power resources within the prediction time window. Specifically: The source load potential assessment model is constructed based on the preset grouped query attention mechanism and loss function; Based on the fused time-varying sequence, the constraints of each power resource at each time segment within the prediction time window are analyzed by the source-load potential assessment model; wherein, the constraints include real-time operating status constraints, physical regulation capability constraints, and external environmental incentive constraints. Based on the real-time operating status constraints and physical adjustment capability constraints, the maximum adjustable power range of each power resource at each time segment is calculated. Based on the external environmental incentive constraints, and by taking into account historical feedback information from the electricity market, the maximum adjustable power range is adjusted to obtain the effective adjustable power range for each time segment. The upper and lower limits of the effective adjustable power range are used to determine the adjustment potential boundary for each time segment.

5. The source-load regulation potential assessment method based on time-varying data management according to claim 4, characterized in that, The process of constructing the source load potential assessment model based on a preset grouped query attention mechanism and loss function is as follows: A multi-head attention module of a preset teacher model is obtained. The attention heads of the multi-head attention module are divided according to the acquisition level to obtain several groups. Each group contains the same number of attention heads, and each group is only used to process data from one acquisition level. The model weights of the attention heads of each group are averaged and aggregated to generate the first model weight of the attention head of each group. Based on the attention heads of all groups, the group query attention mechanism is constructed. Obtain the model weights of all modules in the teacher model except for the first model weights to obtain the second model weights; Based on the second model weights, the grouped query attention mechanism is trained using the loss function to obtain the trained source load potential evaluation model.

6. The source-load regulation potential assessment method based on time-varying data management according to claim 5, characterized in that, The process of training the grouped query attention mechanism based on the second model weights and using the loss function to obtain the trained source load potential evaluation model is as follows: The second model weights are assigned to the corresponding parts of the source load potential assessment model. Then, using a preset knowledge distillation strategy, the weight coefficients of the group query attention mechanism are trained by minimizing the attention weight matrix of the teacher model and the source load potential assessment model through the loss function. The loss function includes a first loss term and a second loss term. The first loss term is used to calculate the difference in output distribution between the teacher model and the source load potential assessment model for the same training sample. The second loss term is used to calculate the difference in output distribution between the multi-head attention module and the grouped query attention mechanism.

7. The source-load regulation potential assessment method based on time-varying data management according to claim 6, characterized in that, The knowledge distillation strategy includes a first training phase and a second training phase. In the first training phase, the source load potential assessment model is trained using training samples with a completeness higher than a first threshold and a prediction time window length lower than a second threshold, and the adjustment potential boundary of all time sections is output. In the second training phase, the source load potential assessment model that has passed the first training phase is trained using training samples with a completeness lower than the first threshold or a prediction time window length higher than the second threshold, and the adjustment potential boundary of the next time segment is calculated based on the output value of the previous time segment.

8. The source-load regulation potential assessment method based on time-varying data management according to claim 1, characterized in that, The process involves analyzing the adjustable parameters corresponding to each time segment within the prediction time window based on the adjustment potential boundary and the constraints, to obtain the source-load adjustment potential assessment result of the power resources. Specifically: Based on the adjustment potential boundary and the constraints, the adjustable parameters for each time segment are constructed; the adjustable parameters include sub-vectors of four key indicators, namely, maximum adjustable upward power, maximum adjustable downward power, total sustainable adjustment energy, and maximum allowable adjustment rate; By integrating the vector elements of all the time sections, a time-varying source load regulation potential envelope is constructed to obtain the source load regulation potential assessment result.

9. A source-load regulation potential assessment system based on time-varying data management, characterized in that, include: The module includes a data acquisition module, a data fusion module, an adjustable boundary prediction module, and an adjustable parameter generation module. The data acquisition module is used to acquire multi-dimensional time-varying data of various power resources; wherein, the multi-dimensional time-varying data includes static attribute data, operating status data, environmental data and power market data of various power resources; The data fusion module is used to perform data source verification and weighted fusion on the multi-dimensional time-varying data to obtain the fused time-varying sequence and the acquisition level related to the data acquisition frequency; The adjustment boundary prediction module is used to capture the spatiotemporal correlation between the fused time-varying sequence and the power resources through a preset source-load potential assessment model, and obtain the adjustment potential boundary and constraints of the power resources within the prediction time window; The adjustable parameter generation module is used to analyze the adjustable parameters corresponding to each time segment within the prediction time window based on the adjustment potential boundary and the constraint conditions, so as to obtain the source-load adjustment potential assessment result of the power resources.

10. A terminal device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the steps of the source-load regulation potential assessment method based on time-varying data management according to any one of claims 1 to 8.

Images