An adaptive accelerated solution method and system for power spot market clearing

Abstract

The application provides a self-adaptive accelerated solving method and system for power spot market clearing, and belongs to the field of power market optimal dispatch. The method comprises the following steps: constructing a historical database, extracting node features and edge features through feature engineering, constructing a graph neural network prediction model, adjusting the prediction model after training in combination with boundary data, and generating a predicted solution through prediction; based on the predicted solution, an adaptive time period reduction method containing a statistical envelope index and a consistency discrimination criterion is used to establish a variable scale time period model and obtain a coarse-grained solution; the coarse-grained solution is expanded into an initial fine-grained solution, the model scale is reduced, and the initial fine-grained solution is sequentially adjusted through start-stop structure adjustment, scale-intra feasibility repair and scale-inter feasibility repair to obtain a fine feasible solution meeting the full physical constraints of the original time period, and the solving of the large-scale power spot market clearing is completed. Under the premise of ensuring the physical feasibility and economic rationality of the clearing result, the solving efficiency and stability are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of power market optimization and dispatching technology, and in particular relates to an adaptive accelerated solution method and system for power spot market clearing. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the deepening of power market reform, the operation mode of the power system is undergoing a fundamental shift from centralized dispatch based on planning to collaborative optimization centered on market mechanisms. In pilot provinces of the spot market, multiple entities such as power generation, energy storage, and virtual power plants are widely participating in the trading of various market commodities, including electricity, reserve, and ancillary services. This has promoted the development of a joint clearing mechanism coupled with multiple trading commodities, significantly increasing the scale and complexity of market clearing models. Against this backdrop, the power system optimization problem has evolved from the traditional single power balance problem into a high-dimensional mixed-integer programming problem involving multiple time periods, multiple entities, multiple commodities, and multiple complex constraints. At the same time, with the gradual liberalization of power generation and consumption plans, the uncertainty of unit output, load response, and high proportion of renewable energy output is increasing, leading to a significant increase in the number of discrete variables representing unit start-up and shutdown in the model, highly coupled various constraints, and an exponential expansion of the search space.

[0004] The aforementioned changes pose a severe challenge to the computational techniques for clearing the spot market. Traditional exact solution methods based on mathematical programming, such as mixed-integer linear programming or quadratic programming, while theoretically guaranteeing the global optimality and price consistency of the clearing result and serving as the core algorithmic foundation of current market operation systems, are highly susceptible to the "combinatorial explosion" problem when faced with large-scale scenarios involving hundreds of thousands of optimization variables and millions of constraints.

[0005] To overcome the computational bottleneck of accurate solutions, academia and industry have introduced solution strategies based on decomposition and coordination. These methods uncover the structural characteristics of the problem in time, space, or functional dimensions, decomposing the original large-scale problem into several relatively independent sub-problems for parallel solution, and then coordinating them through a master-sub-problem framework (such as Lagrange relaxation or Benders decomposition) or hierarchical strategies. While these methods alleviate computational pressure to some extent, their effectiveness highly depends on prior knowledge of the problem structure and sophisticated decomposition design, lacking a universal framework. Furthermore, convergence between master and sub-problems often requires multiple iterations, which can lead to oscillations or slow convergence in scenarios with tight constraints or high uncertainty, potentially increasing the overall solution time.

[0006] Another type of research turns to heuristic or approximate solution methods, which significantly improve computational speed by sacrificing some optimality through simplified models, relaxed constraints, variable aggregation, or staged approximation. However, the global optimality of the clearing results obtained by these methods is difficult to guarantee, and the economic consistency and physical interpretability of price signals are relatively weak. In scenarios involving complex cybersecurity constraints and multi-product coupling, approximation may lead to deviations in resource allocation from the optimal state, and even introduce potential security risks. Summary of the Invention

[0007] To overcome the shortcomings of the existing technology, this invention provides an adaptive accelerated solution method and system for electricity spot market clearing, aiming to solve the technical problems of large-scale spot market clearing models, such as large computational scale, long solution time, complex constraint coupling, and lack of transparency in intermediate decision-making processes.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0009] The first aspect of this invention provides an adaptive accelerated solution method for clearing the electricity spot market;

[0010] An adaptive accelerated solution method for electricity spot market clearing includes:

[0011] A historical database covering multi-dimensional operation and market information is constructed. Node features and edge features are extracted through feature engineering. A graph neural network prediction model that integrates physical constraint loss function is constructed. After training, it is adjusted in combination with boundary data to predict and generate unit operation schemes, output plans, predicted solutions, and minimum available output of the system.

[0012] Based on the predicted solution, an adaptive time period reduction method with statistical envelope index and consistency judgment criterion is adopted to merge the original multiple time periods into a variable-scale time period, establish a variable-scale time period model and solve it to obtain a coarse-grained solution.

[0013] The coarse-grained solution is expanded into an initial fine-grained solution. The model size is reduced by identifying effective variables and redundant constraints. The solution is then adjusted for start-stop structure, intra-scale feasibility repair, and inter-scale feasibility repair to obtain a fine-grained feasible solution that satisfies all physical constraints of the original time period, thus completing the solution for clearing the large-scale electricity spot market.

[0014] As a further technical solution, the feature engineering includes unit feature modeling, edge feature modeling, and label construction;

[0015] The specific modeling of the unit features is as follows: For any node in the power system topology diagram, construct a node feature vector. The vector includes the upper and lower limits of unit output, operating cost, start-up cost, shutdown cost, operating status, minimum continuous start-up and shutdown time, ramp-up capability, output in the previous period, reserve capacity, node load level, number of units connected to the node, and predicted output of new energy sources.

[0016] The edge feature modeling specifically involves: constructing an edge feature vector for any edge connecting two nodes in the topology graph, wherein the vector includes the upper limit capacity of line thermal stability transmission, the actual power flow during the time period, and the line load rate.

[0017] The label construction specifically involves: based on historical market clearing results, constructing supervised learning labels for each node in the corresponding time period, wherein the labels include the unit operating status and active power output.

[0018] As a further technical solution, the construction of the graph neural network prediction model that integrates the physical constraint loss function includes:

[0019] The power system is abstracted as a graph structure, where the set of nodes represents generation nodes, load nodes or their equivalent combinations, and the set of edges represents transmission lines or equivalent connections.

[0020] A multi-layer graph convolutional neural network is constructed. The hidden feature vector of a node is updated by aggregating the information of neighboring nodes layer by layer. The system-level hidden features are gathered by an aggregation function in the form of attention pooling. The minimum available output of the system is predicted by combining external prediction information.

[0021] The introduction of a joint loss function enhances the operational feasibility of the power system. The joint loss function includes a basic loss function and a physical constraint loss function.

[0022] As a further technical solution, the statistical envelope indicators include load curvature envelope, ramp cumulative pressure index, start-stop trigger risk envelope, and standby structure stability index.

[0023] The consistency criteria include numerical stability criteria, structural fluctuation criteria, discrete decision invariance criteria, and comprehensive risk consistency criteria.

[0024] As a further technical solution, the establishment of the variable-scale time-period model and the solution to obtain a coarse-grained solution include:

[0025] With the objective function of minimizing the total operating cost and start-up cost in each variable-scale period, constraints are introduced, including system load constraints, unit output upper and lower limit constraints, ramp rate constraints, system positive and negative reserve constraints, line transmission capacity constraints, and minimum continuous start-up and shutdown time of units. Among them, the system load constraints cover the average, maximum, and minimum system load in the variable-scale period and the maximum increase or decrease in load in adjacent original periods.

[0026] The obtained predicted solution is used as the initial solution of the variable-scale time period model. It is substituted into the model for iterative solution to obtain the unified start-up and shutdown variables and output variables of each unit in each variable-scale time period. The variables constitute a coarse-grained solution. The start-up and shutdown status and output of each unit remain unchanged in the variable-scale time period. A single variable-scale time period corresponds to one or more original time periods.

[0027] As a further technical solution, the model size is reduced through effective variable identification and redundant constraint identification, including:

[0028] The relaxed feasible region is obtained by continuously relaxing the original mixed integer clearing model. The Lagrangian function of the relaxed model is constructed and the unit structural perturbation direction of the start and stop variables is defined. The structural performance index is constructed by constraining the activation indicator function, the intensity of the constraint activation response, and the equivalent second-order structural response of the objective function with respect to the start and stop variables.

[0029] For the coupling constraints in the model, a process of first relaxing and then adding is adopted. By solving the relaxation subproblem with the goal of maximizing the power flow expression of a specific line and the condition of the basic system operation constraints, the theoretical and lower bounds of the power flow of the line in the feasible domain of the whole network are calculated.

[0030] By comparing the calculated theoretical upper and lower bounds with the actual transmission capacity limit of the line, if the theoretical upper bound is lower than the line capacity upper bound, the upper bound constraint is determined to be a redundant constraint; if the theoretical lower bound is higher than the line capacity lower bound, the lower bound constraint is determined to be a redundant constraint. Redundant constraints are eliminated, and non-redundant constraints are reintroduced into the corresponding repair model to simplify the constraint set.

[0031] As a further technical solution, the start-stop structure adjustment step is as follows: take the key start-stop variables as optimization variables, fix the other start-stop variables, and within their influence period, with the goal of minimizing the total system cost, re-solve a mixed integer programming subproblem containing only the key variables to obtain the adjusted start-stop scheme;

[0032] The specific steps for feasibility repair within the scale are as follows: Under the premise of fixed start-up and shutdown status, for all original time periods within each variable scale time period, establish a continuous variable optimization model, and under the condition of satisfying all physical and market constraints within the time period, further optimize the unit output to obtain the optimal output for each original time period.

[0033] The specific steps of the inter-scale coordination and repair are as follows: for the boundary area between adjacent variable-scale time periods, a buffer zone containing several original time periods is defined. Under the premise of fixing the time period solutions outside the buffer zone, only the unit output within the buffer zone is finely adjusted to eliminate the problems of ramp mismatch and standby discontinuity caused by time period merging and expansion, so as to ensure the smoothness and feasibility of the entire time series solution.

[0034] A second aspect of the present invention provides an adaptive accelerated solution system for clearing the electricity spot market.

[0035] An adaptive accelerated solution system for electricity spot market clearing includes:

[0036] The predictive hot start module is configured to: construct a historical database covering multi-dimensional operation and market information, extract node and edge features through feature engineering, construct a graph neural network prediction model that integrates physical constraint loss function, adjust it after training with boundary data, and predict and generate unit operation schemes, output plans, predicted solutions and minimum available output of the system;

[0037] The variable-scale time period reduction module is configured to: based on the predicted solution, adopt an adaptive time period reduction method containing statistical envelope index and consistency judgment criterion to merge the original multiple time periods into a variable-scale time period, establish a variable-scale time period model and solve it to obtain a coarse-grained solution;

[0038] The coarse-grained solution refinement and reconstruction module is configured to: expand the coarse-grained solution into an initial fine-grained solution, reduce the model size through effective variable identification and redundant constraint identification, and sequentially perform start-stop structure adjustment, intra-scale feasibility repair and inter-scale feasibility repair to obtain a fine feasible solution that satisfies all physical constraints of the original time period, thereby completing the solution for clearing the large-scale electricity spot market.

[0039] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of an adaptive accelerated solution method for electricity spot market clearing as described in the first aspect of the present invention.

[0040] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of an adaptive accelerated solution method for clearing the electricity spot market as described in the first aspect of the present invention.

[0041] The above one or more technical solutions have the following beneficial effects:

[0042] (1) This invention fundamentally changes the traditional approach of directly solving ultra-large-scale mixed-integer models through a hierarchical, progressive solution framework of prediction guidance, time period reduction, and structural repair. It utilizes machine learning to predict and generate high-quality initial solutions, effectively guiding the branching and bounding process and reducing meaningless searches. Adaptive time period merging based on physical context significantly reduces the problem dimensionality, directly shrinking the model size. In the solution repair phase, targeted optimization is performed only on identified key variables and effective constraints, avoiding repeated iterations across the entire variable space. This series of collaborative designs reduces the overall solution time by an order of magnitude compared to traditional methods, meeting the stringent requirements of the spot market for computational timeliness.

[0043] (2) This invention deeply embeds the physical laws of the power system into each link of the algorithm. In the prediction stage, the model is trained by integrating the loss function of constraints such as power balance and unit capacity to ensure that the predicted solution has good physical consistency. In the time period reduction stage, the designed statistical envelope index and discrimination criteria have the core objective of not triggering new operational risks when merging time periods, thereby ensuring that the coarse-grained solution is still located near the fine feasible region. In the solution repair stage, through multi-stage fine reconstruction, all original constraints are gradually and strictly restored. The final fine solution output fully meets all physical and market rule requirements and the economic efficiency is close to the global optimum.

[0044] (3) The machine learning model used in this invention can learn scheduling patterns under different boundary conditions from historical data and adapt to fluctuations in load and renewable energy output; the time-period reduction method dynamically constructs envelope indices based on real-time prediction data to achieve model dimensionality reduction "tailored to the scenario"; the intelligent identification mechanism of variables and constraints can automatically focus on key contradictions under the current operating mode. This adaptive characteristic reduces the dependence on manual parameter adjustment and empirical rules, improves the robustness and universality of the algorithm under different power grid structures and different market rules, and is more conducive to deployment and application in actual engineering systems.

[0045] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0046] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0047] Figure 1 This is a flowchart of the method in the first embodiment.

[0048] Figure 2 This is a flowchart of the multi-stage fine-scale reconstruction process for the coarse-grained solution in the first embodiment.

[0049] Figure 3 This is a flowchart of the redundancy constraint identification process in the first embodiment.

[0050] Figure 4 This is a schematic diagram of the greedy algorithm for identifying redundant cross-section power flow constraints in the first embodiment.

[0051] Figure 5 This is a system structure diagram of the second embodiment. Detailed Implementation

[0052] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0053] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0054] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0055] Example 1

[0056] This embodiment discloses an adaptive accelerated solution method for clearing the electricity spot market. It generates predictive solutions by training a machine learning model with historical operating data, thereby accelerating the convergence of branch bounds in the clearing calculation. Based on the physical connotation and mechanism of the clearing model, it achieves dimensionality reduction at the optimization time period and optimization object levels to generate coarse-grained solutions. Through variable structure adjustment and intra- and inter-scale feasibility repair, the coarse-grained solutions are repaired into fine-grained feasible solutions that satisfy complete physical constraints, thereby significantly improving the overall computational efficiency while ensuring clearing accuracy and feasibility.

[0057] Specifically, such as Figure 1 As shown, an adaptive accelerated solution method for electricity spot market clearing includes:

[0058] Step S1: Construct a historical database covering multi-dimensional operation and market information, extract node features and edge features through feature engineering, construct a graph neural network prediction model that integrates physical constraint loss function, and adjust it with boundary data after training to predict and generate unit operation schemes, output plans, predicted solutions and minimum available output of the system.

[0059] Step S101: First, construct a historical database covering multi-dimensional operational and market information. The database includes historical market clearing results (such as unit start-up and shutdown status, power generation output, nodal prices, and inter-market trading volume), as well as system operation boundary condition data, such as load curves, renewable energy forecast output curves, energy price signals, and network topology, forming a comprehensive characterization of the power system's operational status.

[0060] Based on this, using "historical days" as the basic sample unit, feature extraction and reconstruction are performed on the original data. For each historical operation scenario, key feature vectors that reflect the system's operational characteristics and scheduling difficulty are extracted, thereby mapping the high-dimensional, heterogeneous original operation data into a structured, learnable feature representation, laying the data foundation for subsequent agent training and inference.

[0061] Step S102: Extract node features and edge features using feature engineering, wherein the feature engineering includes unit feature modeling, edge feature modeling, and label construction.

[0062] During the unit feature modeling process, for any node in the topology graph During the period Construct the corresponding node feature vector Defined as:

[0063]

[0064] in: They represent the generating units. The upper and lower limits of the output are used to characterize the static feasible range of the unit; Indicates the unit Operating costs Indicates the unit Start-up costs, Indicates the unit Downtime costs; Indicates the time period of the unit The running status; , These represent the time periods of the generating units. Minimum continuous power-on / power-off duration; This indicates the unit's maximum uphill and downhill climbing capabilities; The power output of the unit in the previous time period is used to construct cross-time period coupling characteristics; Indicates the time period of the unit The available backup capacity reflects the system's support capabilities; Indicates the node in the time period The load level; Represents a node The number of connected generating units is used to characterize the local resource redundancy and adjustment potential of the node. This represents the predicted output of new energy sources. The above node characteristics, from multiple dimensions such as output boundaries, temporal coupling, flexibility constraints, reserve capacity, and node resource structure, systematically characterize the operating status and adjustability of nodes under a given time period.

[0065] In the process of edge feature modeling, for the connected nodes in the topological graph With nodes Construct the corresponding edge feature vector for any side. Defined as:

[0066]

[0067] in, Indicates the line thermally stable transmission upper limit capacity, Representing an edge During the period The actual trend The line load factor represents the line's operating status and the degree of constraint on its transmission capacity. The edge features describe the physical connections between nodes in the power grid and their transmission capacity constraints, providing a basis for power flow feasibility analysis in scheduling decisions.

[0068] For tag construction, based on historical market clearing results, for each node During the period Constructing supervised learning labels Defined as:

[0069]

[0070] in, Indicates the unit During the period The running status, Indicates the unit During the period The label reflects the contribution or effort made. It indicates the decision-making outcome obtained from a market clearing or scheduling optimization model, given system boundary conditions and network constraints.

[0071] Step S103: In this embodiment, to address the problems of complex power system topology, strong power coupling between nodes, and limited computational efficiency of traditional models, a prediction model based on Graph Neural Network (GNN) is designed to learn the nonlinear mapping relationship between power grid topology and power flow distribution.

[0072] First, the power system is abstracted into a graph structure. , where the set of nodes Represents a set of edges representing a generator node, a load node, or an equivalent combination thereof. This diagram represents the transmission lines or equivalent connections between nodes. This graph structure explicitly depicts the physical coupling relationships and network constraint characteristics between nodes in a power system. Each generating unit in the system... Mapped to its own node The units are indirectly coupled through the node-line topology.

[0073] Secondly, based on the above graph structure, a multi-layer graph convolutional neural network is constructed. By aggregating neighborhood node information layer by layer, the power flow distribution law of the power system is learned. For the first... Layered graph convolution, its nodes The hidden feature vector update method is defined as follows:

[0074]

[0075] in: Indicates the unit In the The feature vector after layer graph convolution, initially ; Indicates the unit The set of neighboring units; , These represent the degree of the node to which the unit belongs, and are used to normalize the neighborhood information to avoid the adverse effects of node degree differences on feature propagation. Indicates the first The layer is a trainable weight matrix used to perform linear transformations on the features of neighboring nodes; Indicates the bias term; This represents a nonlinear activation function used to enhance the nonlinear expressive power of the model; the ReLU activation function is preferred.

[0076] At the system level, aggregation functions are used to gather the high-level hidden features of all units:

[0077]

[0078] in: For time period System-level hidden feature representation; It is an aggregation function used to map a variable-length set of unit features to a fixed-dimensional system state vector; READOUT uses attention pooling. Based on system-level features and external prediction information, it predicts the system's state over a given time period. Minimum available output:

[0079]

[0080] in: The predicted minimum available conventional unit output for the system; For the system in time period The load forecast value; For the system in time period The predicted output of new energy sources; These correspond to the weight parameters respectively; It is a system-level prediction bias term; It is a non-negative activation function, ReLU, used to ensure the physical plausibility of the prediction results.

[0081] Furthermore, in this embodiment, to avoid the problem that traditional data-driven models rely solely on statistical correlation and cannot guarantee the feasibility of power system operation, a joint loss function is introduced to enhance the feasibility of power system operation. The joint loss function includes a basic loss function and a physical constraint loss function.

[0082] The basic loss function design addresses the characteristic of scheduling decisions involving both discrete variables (unit start-up and shutdown status) and continuous variables (generation output). A multi-task learning loss function is constructed, including:

[0083] For the task of predicting the unit's operating status, a binary cross-entropy loss function is used. The binary cross-entropy loss function, used to measure the difference between the predicted start-up status and the historical clearing results, is defined as follows:

[0084]

[0085] in: This represents the total number of units participating in scheduling within the system. This refers to the number of time periods within the scheduling cycle. For generator sets... During the period The task of predicting the running status has : The actual operational status label obtained from historical market clearing or scheduling optimization; Model-predicted units During the period The power-on status output.

[0086] For the task of predicting the active power output of the generating unit, the mean square error loss function is used. This is used to characterize the deviation between the predicted output and the actual clearing output. Mean squared error loss function. Defined as:

[0087]

[0088] Among them, for the unit During the period The task of predicting the contribution of personnel. The actual unit output value obtained from historical market clearing or scheduling optimization; For the units predicted by the model During the period The results of their efforts.

[0089] The aforementioned basic loss function can ensure that the model statistically approximates the historical market clearing results, but it is not sufficient to constrain the prediction results to meet the physical operating laws of the power system.

[0090] Based on the basic loss function, a physical constraint loss term is introduced. This is used to characterize the degree to which the prediction results violate the basic physical constraints of the power system, and a physical constraint loss function is constructed, which is defined as:

[0091]

[0092] in: Indicates the units predicted by the model. During the period Those who have made meritorious contributions; Indicates the system during the time period The predicted output of new energy sources; Indicates the system during the time period Load demand; Indicates the unit The maximum power generation capacity limit.

[0093] The physical loss function contains two types of core physical constraints: (1) Power balance constraint loss term; the first term is used to constrain the system to meet the power balance relationship in each time period, so that the sum of the output of conventional units and the output of new energy can match the system load demand, thereby guiding the model to learn the basic operating law of the power system "source-load balance". (2) Unit output upper limit constraint loss term; the second term punishes the situation where the predicted output exceeds the physical upper limit of the unit, constrains the model output to meet the unit capacity boundary conditions, and avoids generating unexecutable scheduling schemes.

[0094] Taking into account both prediction accuracy and physical feasibility, the final training loss function is constructed as follows:

[0095]

[0096] in, These are weighting coefficients used to balance data fitting error with the strength of physical constraints.

[0097] Step S104: Before using the predictive agent, it is necessary to modify and adjust the unit-related data information, such as the unit's specified status, unit maintenance plan, maximum and minimum operating modes of the unit group, and minimum advance notice start-up time. These adjustments ensure the accuracy of the unit information at the input end.

[0098] Let the set of units be The time set is The decision variable is the unit state:

[0099]

[0100] in Indicates the unit During the period It is currently running.

[0101] To ensure that the predicted initial solution is consistent with the actual boundary conditions on the operating day, the following constraints are explicitly adjusted:

[0102] (1) Unit designated status adjustment:

[0103]

[0104] in: This is the designated state of the generator set. It is a set of state machines.

[0105] (2) Adjustment of unit maintenance plan:

[0106]

[0107] in: The maintenance plan involves a group of generating units. It refers to the time period of the maintenance plan involving the units.

[0108] (3) Maximum / minimum operating mode constraints for the unit group:

[0109]

[0110] in: It is a group of generator units The minimum number of units allowed to operate simultaneously at any given time. It is a group of generator units The maximum number of units allowed to operate simultaneously at any given time.

[0111] (4) Minimum advance notice start time constraint:

[0112]

[0113] in: , indicating the unit During the period The startup advance notification status variable, 1 indicates that the startup notification has been completed in this time period, and 0 indicates that it has not been completed; Indicates the unit The minimum advance notice start time is the shortest time interval (in scheduling periods) required from the issuance of the start command to the time when the unit is allowed to start.

[0114] The above modifications fix some of the unit operation variables in the model and also ensure that the predictive agent does not change the fixed variables.

[0115] After completing model training based on historical data and adjusting the input boundary data, the graph neural network model is deployed in an actual operation or simulation environment to predict unit operation schemes. Forecast unit output plan Prediction of minimum available output of the system Predicted solution This enables rapid prediction and auxiliary decision support for power system dispatch decisions.

[0116] Step S2: Based on the predicted solution, an adaptive time period reduction method with statistical envelope index and consistency criterion is adopted to merge the original multiple time periods into a variable-scale time period, establish a variable-scale time period model and solve it to obtain a coarse-grained solution.

[0117] The predicted solution is obtained based on online prediction and application. The system's minimum available output is predicted using the subsequent scaling-down step as the initial solution for the model. This is used to construct the start / stop trigger risk envelope indicator for subsequent statistical envelope indicators.

[0118] Based on the above-mentioned agent prediction mechanism, the predicted operating solution, the output set, and the prediction of the minimum available output of the system, this invention uses the statistical envelope adaptive time period reduction method to merge and reduce the original multiple time periods according to the predicted solution. Then, a variable-scale time period model is established and the coarse-grained variable-scale optimal solution is obtained.

[0119] Step S201: Adopt an adaptive time-period reduction method based on statistical envelope, assuming the original unit combination model includes... Each equal-length original time period Indicates the system during the time period The load, whose system load sequence is represented as ,

[0120] The load sequence is preprocessed as follows to calculate the load change between adjacent time periods:

[0121]

[0122] Step S2011: Construct the statistical envelope index. This can be done using any candidate time scale. Define the statistical envelope characteristic indicators within this scale, using units as the unit, including:

[0123] (1) Load curvature envelope:

[0124] Define the second-order difference of the load within the scale:

[0125]

[0126] Construction curvature risk indicators:

[0127]

[0128] Define scale The equivalent reserve requirement range within:

[0129]

[0130] in: This is a scale-dependent risk amplification factor used to characterize the system's conservatism regarding load curvature risk. At scale... Within, based on the average level of the predicted load. Construct an equivalent spare constraint interval:

[0131]

[0132] in, .

[0133] (2) Cumulative Stress Index During Climbing:

[0134]

[0135] in: This refers to the unit's hill-climbing capability.

[0136] (3) Start-stop trigger risk envelope:

[0137] Because the predictive agent has already predicted the initial unit start-up and shutdown states and adjusted them based on boundary data... and the minimum available conventional unit output of the prediction system :

[0138]

[0139]

[0140]

[0141]

[0142]

[0143]

[0144]

[0145] The above formulas mainly define the predicted start-stop ground state. Define the minimum output envelope of the prediction system. Define the predictive start-stop deadlock coefficient. And construct a system-level conservative correction factor and the corrected system minimum output envelope Construction load-start-stop margin sequence Define the start-stop risk index within the specified scale. If this value is close to 0, it indicates a risk of triggering the start-up and shutdown of a new unit. Specifically: For the unit Minimum continuous power-on time; For the unit The minimum continuous downtime.

[0146] (4) Stability index of the backup structure:

[0147] Assess whether the backup structure remains structurally consistent across the entire scale. This is the maximum output of the unit's technology.

[0148]

[0149] Step S2012 involves designing the envelope consistency criterion. To determine whether multiple original time periods can be merged into the same variable-scale time period, a statistical envelope preservation constraint is introduced.

[0150] For any candidate scale The following conditions must be met simultaneously:

[0151] (1) Numerical stability criteria:

[0152] scale The following standby constraints can be uniformly written as:

[0153]

[0154]

[0155] (2) Structural fluctuation discrimination:

[0156]

[0157] (3) Discrete decision invariance criterion:

[0158]

[0159] (4) Comprehensive risk consistency judgment, defining the comprehensive operational risk function:

[0160]

[0161] Judgment criteria:

[0162]

[0163] in, , , These are adjustable threshold parameters that can be adaptively set based on system load fluctuation characteristics or the unit's ramp-up capability ratio. It is to predict the output of the generating unit, The above constraints ensure that load changes do not trigger new unit start-ups or extreme ramp-up demands within a variable-scale period, based on the predicted output of new energy units.

[0164] Step S2013: The original time periods are merged and reduced using a sequential scanning adaptive aggregation strategy. First, input the number of original time periods. Load forecast curve Corrected state set Maximum ascent, descent, and hill climbing ,unit Minimum continuous power-on / power-off time , Predicting unit output Maximum output of the unit And predicting the output of new energy units .

[0165] Initialize the starting point of the variable-scale time period, and gradually expand the scale backward from the first original time period, calculating the statistical envelope index in real time; when the envelope consistency criterion is violated for the first time, the current scale is identified as a complete variable-scale time period; record its length. And the current time period is used as the starting point of the next variable-scale time period;

[0166] Repeat the above process until the entire original time period is covered. Finally, generate... There are several variable-scale time periods, and the lengths of each variable-scale time period may be unequal.

[0167] Finally, output each variable-scale time period. Corresponding original time period set and the length of the variable-scale time period .

[0168] Step S202: Based on the above results of merging and reducing variable-scale time periods, establish a variable-time-scale unit combination model. Each variable-scale time period includes one or more original time periods, and it is assumed that the start-up and shutdown status and output of each unit remain unchanged within the variable-scale time period, with each unit modeled using a single variable. The objective function of this model is as follows:

[0169]

[0170] The optimization objective of the model is to minimize the total operating cost and startup cost at each time period. , The units During variable scale periods The output and start / stop variables, Number of generating units For the number of time periods with varying scales, For variable scale time period The length (in units of the original time period). This is the original time period duration. and The units The running cost function and the startup cost function, Typically, a piecewise linear convex function with respect to the unit's output is chosen. A constant cost is typically used for each startup of the unit.

[0171] The constraints of the model are as follows:

[0172]

[0173]

[0174]

[0175]

[0176]

[0177]

[0178]

[0179]

[0180]

[0181] in, , , , , Specific to variable scale time periods The system load average, maximum, minimum, maximum increase in adjacent original time periods, and maximum decrease in adjacent original time periods. , , , Generator sets The upper and lower limits of output and the upper and lower limits of climbing speed. , This represents the system's positive and negative reserve rates. For the line Transmission capacity, , The units and nodes Corresponding route The power flow transfer distribution factor For nodes During variable scale periods The average load, The number of system lines, This represents the number of system nodes. , Representing time periods of varying scales Starting from the unit The effective time range of the minimum continuous start-up and shutdown time constraint can be determined by the unit. Minimum continuous start-stop time and To obtain.

[0182] By using the agent's predicted solution as the initial solution for iterative solving, the results for each unit during the variable-scale time period can be obtained. and At this point, an optimal solution for a coarse-grained time period is obtained. , .

[0183] Step S3: Expand the coarse-grained solution into an initial fine-grained solution. Reduce the model size by identifying effective variables and redundant constraints. Then, proceed with start-stop structure adjustment, intra-scale feasibility repair, and inter-scale feasibility repair to obtain a fine-grained feasible solution that satisfies all physical constraints of the original time period, thus completing the solution for clearing the large-scale electricity spot market.

[0184] The above steps yielded a coarse-grained optimal solution. However, since it did not refine the solutions for each original time period, this solution still has significant room for optimization and improvement. The following section employs a multi-stage refined reconstruction logic to refine the time periods of the coarse-grained optimal solution, such as... Figure 2 As shown.

[0185] Because the solution process involves start-stop structure adjustment models, intra-scale feasible repair models, and inter-scale feasible repair models, an effective variable identification method is used to select a set of candidate adjustment state variables for the start-stop structure adjustment model, reducing the scale of discrete variables in the model. A constraint identification method is used to identify effective constraints for both the intra-scale and inter-scale repair models, further reducing the model size. The specific process is as follows:

[0186] Step S301, the first step in refining the coarse-grained variable-scale solution is to expand the variable-scale solution and use it as the initial solution for each model. For any unit... With any original time period Define the initial values ​​of the start / stop variables during the time expansion phase as follows:

[0187] For any unit , any The initial expansion of the output variable is defined as follows:

[0188]

[0189] To expand the perturbation variables, which characterize the redistribution of output within the scale. To ensure that the macroeconomics of the variable-scale solution is not compromised, the perturbation variables must satisfy:

[0190]

[0191] After the above time expansion, the initial fine-grained solution is obtained. .

[0192] Step S302 involves identifying and extracting effective variables and constraints from the sub-models during the reconstruction process, specifically including:

[0193] Step S3021: The variable scale of large-scale spot clearing models and their subproblems is enormous, especially when dealing with mixed-integer models, which further increases the difficulty of solving them. This embodiment proposes a variable identification method based on structural perturbation response to pre-select effective discrete variables. The specific steps are as follows:

[0194] The original mixed integer clearing model is continuously relaxed, and the start and stop variables are... Relax as While maintaining the integrity of all coupling constraints (power balance, redundancy, ramping, minimum start-stop time, etc.), the relaxed feasible region is obtained. This relaxation solution is considered as the structural baseline state of the system under the current operating scenario.

[0195] (1) Construction of structural perturbation operator

[0196] Constructing the Lagrangian function for the relaxation model:

[0197]

[0198] in For continuous variables, These are the start and stop variables. Objective function. This represents the system operation objective function of the original mixed-integer relaxation clearing model; This represents the set of indices for all constraints in the model. Indicates the number is A single constraint function; Lagrange multipliers Representing constraints The Lagrange multiplier corresponding to the relaxation solution.

[0199] Define the structural perturbation direction for start and stop variables, for any start and stop variable The perturbation direction of the unit structure is introduced. This perturbation does not represent the actual value change and is only used for sensitivity analysis.

[0200]

[0201] (2) Geometric deformation measurement of feasible region

[0202] Define the constraint activation indicator function:

[0203]

[0204] Define start / stop variables The constraint activation response strength is a quantity that characterizes the "tight constraint set reconstruction capability" triggered by small perturbations in variables.

[0205]

[0206] in: Represents start / stop variables Will small perturbations lead to constraints? This indicates a transition from a non-tight constraint to a tight constraint or from a tight constraint to a non-tight constraint. This indicates the weight of the economic impact of the constraint on the objective function once it is activated.

[0207] Define the equivalent second-order structural response of the objective function with respect to start and stop variables. This index is used to characterize whether the switching of variable values ​​will cause the descent path of the objective function to "bend" or "change direction".

[0208]

[0209] in: The term represents the first-order sensitivity of the start and stop variables to the Lagrange function, while the second-order term describes the...

[0210] Will switching the start / stop variable cause a significant bend or jump in the target's descent direction?

[0211] Finally, construct the structural strength index:

[0212]

[0213] in: To constrain structural reconfiguration capabilities; The third term represents the target path curvature response; the fourth term represents the direct economic driver. express Direct marginal economic impact on the original objective function For normalized weights. If but It is identified as a start / stop variable for the structure to function.

[0214] Step S3022, model reduction based on redundancy constraint identification, wherein the flowchart of redundancy constraint identification is as follows: Figure 3 As shown.

[0215] In solving the mathematical model, a process criterion of first relaxing and then adding redundant constraints is adopted, aiming to add constraints that actually have an effect on the model, thereby improving the efficiency of subsequent solution. This invention selects a highly coupled constraint, the power flow constraint, as a demonstration of redundant constraint removal.

[0216] Based on the cross-sectional power flow redundancy constraint analysis of unit output and balance constraints, the cross-sectional power flow constraints in the unit combination model can be abstracted into the following form:

[0217] Cross-sectional power flow constraints:

[0218]

[0219] in , and The constant coefficients, and The upper and lower boundaries of the cross-sectional current flow are constrained. and These are the unit and tie-line variables involved in the constraint. For ease of expression, the constraint can be further rewritten as:

[0220]

[0221]

[0222] make Let be the feasible region of the SCUC problem, and define:

[0223]

[0224]

[0225] That is, within the entire feasible region The upper and lower bounds of the variable. At this point, we can observe that for... ,if

[0226]

[0227] At this point, no feasible solution to the original problem will cause the upper bound of the cross section to be violated.

[0228]

[0229] At this point, no feasible solution to the original problem will cause the lower bound of the cross section to be violated.

[0230] In other words, all cross-sectional constraints that satisfy the above two conditions are redundant constraints, therefore they can be... and Used to prove the redundancy of constraints. In actual implementation, due to the feasible region of the original problem... Extremely complex, requiring solutions for each cross section. , It's generally impossible, so consider The issue of slack in the surface.

[0231]

[0232]

[0233] because At this point, we get , It is better than an exact solution in proving redundancy. , Weaker, but reasonable It can still provide non-trivial redundant constraint judgments.

[0234] Redundancy constraint analysis subproblem:

[0235]

[0236]

[0237]

[0238]

[0239]

[0240] Based on this, we consider solving this subproblem. First, we observe that in the optimal solution to this problem... Must satisfy:

[0241]

[0242] in For indicator functions, that is... Fixed to maximum The boundary. Also, after simplification, regarding... The problem

[0243]

[0244]

[0245]

[0246]

[0247] A greedy algorithm can be used to solve this problem directly, such as... Figure 4 As shown. Soon Sort by size from largest to smallest and gradually increase the corresponding values. Until the balance constraint is satisfied.

[0248] Based on the above methods, redundant constraints are gradually screened out and their effects are identified and added to the corresponding model, effectively reducing the complex and large constraint set of the original model.

[0249] Step S303 involves refined repair based on variable-scale solutions. By employing the proposed method for identifying and extracting effective variables and constraints, the effective variables and constraints of each model can be identified, further reducing the model size. This study's refined repair based on variable-scale solutions involves start-stop structure adjustment models and intra- / inter-scale feasibility repair models. Applying the methods for identifying effective variables and constraints to model solving can provide efficient acceleration for the model.

[0250] Assume the original unit combination model includes The variable-scale model divides the time period into equal-length intervals into... A set of time periods with varying scales: In the variable-scale model, each variable-scale period The optimal solution for an equivalent operating state in an aggregate sense is expressed as: .

[0251] Step S3031: Since the start-up and shutdown variable solutions in the variable-scale solution are coarse-grained and uniform, that is, the variable is highly consistent within the unit variable scale, it can be found that the variable will inevitably cause a certain unit to delay start-up or delay shutdown, which will increase the objective value for minimizing the model. Based on this problem, with the goal of reducing the objective function, this study proposes a start-up and shutdown structure adjustment method to finely adjust the start-up and shutdown variables of the unit in each time period.

[0252] For any unit Define continuous power-on time status Continuous downtime status Its update rules are as follows:

[0253]

[0254] This state is recursively derived over time, forming a cross-scale time state chain. The start-stop solution is obtained from the variable-scale model. First, map to the original time period:

[0255]

[0256] This mapping serves only as a reference solution for the start-up and shutdown structure and is not directly used as the final start-up and shutdown decision. At the original time-scale, the following validity check is performed on each unit:

[0257] Minimum continuous power-on time determination, if: Then the following must be satisfied:

[0258]

[0259] Minimum continuous downtime determination, if: Then the following must be satisfied:

[0260]

[0261] in , These are the minimum continuous start-up and shutdown times of the unit. Then, using a variable identification method based on structural disturbance response, the variables are divided into the following states:

[0262] Frozen state variables If the start-stop switching meets the minimum continuous time constraint, but its adjustment would compromise output feasibility; or if standby or network constraints fail, then the unit's start-stop status is marked as: This setting must be strictly fixed during subsequent repairs and cannot be changed. (The settings are...) .

[0263]

[0264] Adjustable state If the start-up and shutdown process meets the minimum continuous time constraint, and the unit is at low load or near minimum output during that period; and the system has redundancy for standby or ramp-up, then the unit is marked as: This allows for adjustments during subsequent start-up and shutdown optimization phases. The defined values ​​are: , .

[0265]

[0266] critical state If the start-up and shutdown states meet the minimum continuous time constraint, but: the unit is a necessary output unit for the system; or a critical support unit for network constraints and standby constraints; then the unit is marked as: This type of state is only allowed to participate in optimization in restricted integer repair models. The definition is as follows: , and Using the settings from the first two states, that is: , .

[0267]

[0268] For sets Defined as a candidate set for start-stop structure adjustment For the decision variables of the subsequent model Candidate set for start-stop structure adjustment ,Right now Define the unit During the period The effective impact range of start / stop adjustments:

[0269]

[0270] in: The duration of the start-stop effect is defined as:

[0271]

[0272] Construct a set of overall start-up and shutdown adjustment impact periods, which represents the cross-time intervals jointly covered by the start-up and shutdown adjustments of multiple candidate units. This is the time range that needs to be included in both modeling and solving in the local repair model.

[0273]

[0274] Introducing a long-view start-stop economic benefit function:

[0275]

[0276] in: The fixed costs of starting or stopping the unit It is the unit's no-load operating cost, It is the start-up and shutdown of the unit based on the ground state solution (i.e., the variable scaling solution). Time period Start-stop status It is a generator set During the period The output cost of the ground state solution. If This indicates that adjusting the start-up and shutdown of the unit can reduce costs. This refers to the one-time cost of starting and stopping the generator unit.

[0277] This function is incorporated into the start-stop structure repair model, namely:

[0278]

[0279] Constraints are considered only those related to load balance, unit start-up and shutdown status, and safety, including: system load balance constraints, unit output upper and lower limit constraints, unit minimum start-up and shutdown duration constraints, and power flow section constraints. A variable-metric solution is used as the initial solution for solving the problem, obtaining the start-up and shutdown structure adjustment variables and combining them with the unadjusted variables to form... And the obtained unit output .

[0280] Step S3032: Perform feasibility repair within the scale. For any variable-scale time period... Construct a scale-specific feasibility repair subproblem, which only involves: scale-specific... The original time period; the start / stop status has been fixed as follows. Traditional generator sets; all variables are continuous. At scale... The following repair variables are introduced: unit In the original time period Power generation Unit positive and negative standby capacity Slope relaxation variables , .

[0281] Define the intra-scale repair model:

[0282]

[0283] ST.

[0284]

[0285]

[0286]

[0287]

[0288]

[0289]

[0290] This model is based on a unit with fixed discrete variables and mainly considers constraints on the smooth transition of unit output, including: power balance constraints, upper and lower limits of unit output constraints, scale-based ramping constraints, system reserve constraints, and linear power flow constraints. For the piecewise linear or convex operating cost function declared by the unit, This represents the original time period load; The decision variable is the unit's output; It is the unit's technical maximum and minimum output; This refers to the unit's ability to climb steep inclines and declines. This is the maximum reserve capacity of the generating unit; It is the system's backup ratio coefficient; It is a line upper limit capacity, It is a generator set For the line The current transfer factor.

[0291] After constructing the above model, a model reduction method based on redundancy constraint identification is used to remove the ineffective constraints in the model. Then, the model is solved to obtain the unit output decision variables within the scale. .

[0292] Step S3033: Feasibility repair between scales. After completing the feasibility repair within each variable scale, adjacent scales... Systematic conflicts may still exist at their time boundaries due to timescale reduction. Unlike intra-scale conflicts, scale boundary conflicts have a significant "time state inheritance" characteristic, mainly reflected in: the ramp-up continuity of unit output between the end and beginning of the scale; the transformation of unit start-up and shutdown states at the scale boundary; the cross-scale consistency of minimum continuous start-up / shutdown time constraints; and the risk of abrupt changes in reserve and network constraints at the boundary period.

[0293] Let the first Each variable scale corresponds to the original time period set as follows: Adjacent scales are ,in After completing the intra-scale repair, for each adjacent scale pair... The boundary time is determined as follows:

[0294] Criteria for Climbing Constraint Conflict:

[0295] If there are units satisfy:

[0296]

[0297] If the scale splicing violates the unit ramping constraint, it will be determined that the scale splicing violates the unit ramping constraint.

[0298] Criteria for determining start-stop boundary conflicts:

[0299] If the start / stop variable is:

[0300]

[0301] This represents the boundary time period between two scales, during which the unit's operating status changes, and it is necessary to fix discrete variables for each time period in the buffer band.

[0302] If during the boundary period :

[0303]

[0304] Indicates the start point of the next scaled period of the system The required system-level reserve requirement indicates that if the constraint cannot be eliminated by adjusting small continuous variables while keeping the repair results unchanged within the scale, then a system-level conflict is considered to exist at this boundary.

[0305] For each adjacent scale Construct a boundary buffer zone at their intersection:

[0306]

[0307] in, Set the buffer length parameter. Only permitted in the buffer zone. The continuous variables are readjusted internally, while the solutions remain unchanged for other time periods. Under the premise of fixed start / stop states, a fixed marginal buffer zone is established. marginal original time period and Output This ensures the feasibility of smooth adjustments and connections in output within and between scales. (This is for boundary buffer zones.) Construct the following inter-scale coordinated repair model and optimize the objective function (minimum boundary perturbation):

[0308]

[0309] in, This is the original solution after the intra-scale repair is completed. These are the decision variables. Constraints include: power balance constraints (per time period), unit output upper and lower limits constraints, start-stop-output logic constraints, ramp-up constraints (including cross-scale boundaries), reserve constraints, and power flow constraints.

[0310] Inter-scale boundary repair is performed scale by scale in chronological order:

[0311]

[0312] After each pair of scales is repaired, the repair result is used as input for the subsequent scale to ensure consistency and feasibility across the global time series. The final solution yields the results for each unit. Each time period of .

[0313] Example 2

[0314] This embodiment discloses an adaptive accelerated solution system for clearing the electricity spot market;

[0315] like Figure 5 As shown, an adaptive accelerated solution system for electricity spot market clearing includes:

[0316] The predictive hot start module is configured to: construct a historical database covering multi-dimensional operation and market information, extract node and edge features through feature engineering, construct a graph neural network prediction model that integrates physical constraint loss function, adjust it after training with boundary data, and predict and generate unit operation schemes, output plans and predictive solutions.

[0317] The variable-scale time period reduction module is configured to: based on the predicted solution, adopt an adaptive time period reduction method containing statistical envelope index and consistency judgment criterion to merge the original multiple time periods into a variable-scale time period, establish a variable-scale time period model and solve it to obtain a coarse-grained solution;

[0318] The coarse-grained solution refinement and reconstruction module is configured to: expand the coarse-grained solution into an initial fine-grained solution, reduce the model size through effective variable identification and redundant constraint identification, and sequentially perform start-stop structure adjustment, intra-scale feasibility repair and inter-scale feasibility repair to obtain a fine feasible solution that satisfies all physical constraints of the original time period, thereby completing the solution for clearing the large-scale electricity spot market.

[0319] Example 3

[0320] The purpose of this embodiment is to provide a computer-readable storage medium.

[0321] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of an adaptive accelerated solution method for electricity spot market clearing as described in Example 1.

[0322] Example 4

[0323] The purpose of this embodiment is to provide an electronic device.

[0324] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in an adaptive accelerated solution method for clearing the electricity spot market as described in Embodiment 1.

[0325] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0326] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0327] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An adaptive accelerated solution method for electricity spot market clearing, characterized in that, include: A historical database covering multi-dimensional operation and market information is constructed. Node features and edge features are extracted through feature engineering. A graph neural network prediction model that integrates physical constraint loss function is constructed. After training, it is adjusted in combination with boundary data to predict and generate unit operation schemes, output plans, predicted solutions, and minimum available output of the system. The construction of the graph neural network prediction model that incorporates the physical constraint loss function includes: The power system is abstracted as a graph structure, where the set of nodes represents generation nodes, load nodes or their equivalent combinations, and the set of edges represents transmission lines or equivalent connections. A multi-layer graph convolutional neural network is constructed. The hidden feature vector of a node is updated by aggregating the information of neighboring nodes layer by layer. The system-level hidden features are gathered by an aggregation function in the form of attention pooling. The minimum available output of the system is predicted by combining external prediction information. A joint loss function is introduced to enhance the operational feasibility of the power system. The joint loss function includes a basic loss function and a physical constraint loss function. Based on the predicted solution, an adaptive time-segment reduction method incorporating statistical envelope indices and consistency criteria is used to merge the original multiple time periods into a variable-scale time period. A variable-scale time-segment model is then established and solved to obtain a coarse-grained solution, including: With the objective function of minimizing the total operating cost and start-up cost in each variable-scale period, constraints are introduced, including system load constraints, unit output upper and lower limit constraints, ramp rate constraints, system positive and negative reserve constraints, line transmission capacity constraints, and minimum continuous start-up and shutdown time of units. Among them, the system load constraints cover the average, maximum, and minimum system load in the variable-scale period and the maximum increase or decrease in load in adjacent original periods. The obtained predicted solution is used as the initial solution of the variable-scale time period model. It is substituted into the model for iterative solution to obtain the unified start-up and shutdown variables and output variables of each unit in each variable-scale time period. The variables constitute a coarse-grained solution. The start-up and shutdown status and output of each unit remain unchanged in the variable-scale time period. A single variable-scale time period corresponds to one or more original time periods. The statistical envelope indicators include load curvature envelope, ramp cumulative pressure index, start-stop trigger risk envelope, and standby structure stability index. The consistency criteria include numerical stability criteria, structural fluctuation criteria, discrete decision invariance criteria, and comprehensive risk consistency criteria. The coarse-grained solution is expanded into an initial fine-grained solution. The model size is reduced through effective variable identification and redundancy constraint identification. This is followed by start-stop structure adjustment, intra-scale feasibility repair, and inter-scale feasibility repair to obtain a refined feasible solution that satisfies all physical constraints of the original time period, thus completing the large-scale electricity spot market clearing solution. The reduction of model size through effective variable identification and redundancy constraint identification includes: The relaxed feasible region is obtained by continuously relaxing the original mixed integer clearing model. The Lagrangian function of the relaxed model is constructed and the unit structural perturbation direction of the start and stop variables is defined. The structural performance index is constructed by constraining the activation indicator function, the intensity of the constraint activation response, and the equivalent second-order structural response of the objective function with respect to the start and stop variables. For the coupling constraints in the model, a process of first relaxing and then adding is adopted. By solving the relaxation subproblem with the goal of maximizing the power flow expression of a specific line and the condition of the basic system operation constraints, the theoretical and lower bounds of the power flow of the line in the feasible domain of the whole network are calculated. By comparing the calculated theoretical upper and lower bounds with the actual transmission capacity limit of the line, if the theoretical upper bound is lower than the line capacity upper bound, the upper bound constraint is determined to be a redundant constraint; if the theoretical lower bound is higher than the line capacity lower bound, the lower bound constraint is determined to be a redundant constraint. Redundant constraints are eliminated, and non-redundant constraints are reintroduced into the corresponding repair model to simplify the constraint set.

2. The adaptive accelerated solution method for electricity spot market clearing as described in claim 1, characterized in that, The feature engineering includes unit feature modeling, edge feature modeling, and label construction; The specific modeling of the unit features is as follows: For any node in the power system topology diagram, construct a node feature vector. The vector includes the upper and lower limits of unit output, operating cost, start-up cost, shutdown cost, operating status, minimum continuous start-up and shutdown time, ramp-up capability, output in the previous period, reserve capacity, node load level, number of units connected to the node, and predicted output of new energy sources. The edge feature modeling specifically involves: constructing an edge feature vector for any edge connecting two nodes in the topology graph, wherein the vector includes the upper limit capacity of line thermal stability transmission, the actual power flow during the time period, and the line load rate. The label construction specifically involves: based on historical market clearing results, constructing supervised learning labels for each node in the corresponding time period, wherein the labels include the unit operating status and active power output.

3. The adaptive accelerated solution method for electricity spot market clearing as described in claim 1, characterized in that, The specific steps for adjusting the start-stop structure are as follows: taking the key start-stop variables as optimization variables, fixing the other start-stop variables, and within the scope of their influence period, resolving a mixed integer programming subproblem containing only the key variables with the goal of minimizing the total system cost, to obtain the adjusted start-stop scheme; The specific steps for feasibility repair within the scale are as follows: Under the premise of fixed start-up and shutdown status, for all original time periods within each variable scale time period, establish a continuous variable optimization model, and under the condition of satisfying all physical and market constraints within the time period, further optimize the unit output to obtain the optimal output for each original time period. The specific steps of the inter-scale coordination and repair are as follows: for the boundary area between adjacent variable-scale time periods, a buffer zone containing several original time periods is defined. Under the premise of fixing the time period solutions outside the buffer zone, only the unit output within the buffer zone is finely adjusted to eliminate the problems of ramp mismatch and standby discontinuity caused by time period merging and expansion, so as to ensure the smoothness and feasibility of the entire time series solution.

4. An adaptive accelerated solution system for electricity spot market clearing, characterized in that, include: The predictive hot start module is configured to: construct a historical database covering multi-dimensional operation and market information, extract node and edge features through feature engineering, construct a graph neural network prediction model that integrates physical constraint loss function, adjust it after training with boundary data, and predict and generate unit operation schemes, output plans, predicted solutions and minimum available output of the system; The construction of the graph neural network prediction model that incorporates the physical constraint loss function includes: The power system is abstracted as a graph structure, where the set of nodes represents generation nodes, load nodes or their equivalent combinations, and the set of edges represents transmission lines or equivalent connections. A multi-layer graph convolutional neural network is constructed. The hidden feature vector of a node is updated by aggregating the information of neighboring nodes layer by layer. The system-level hidden features are gathered by an aggregation function in the form of attention pooling. The minimum available output of the system is predicted by combining external prediction information. A joint loss function is introduced to enhance the operational feasibility of the power system. The joint loss function includes a basic loss function and a physical constraint loss function. The variable-scale time-segment reduction module is configured to: based on the predicted solution, employ an adaptive time-segment reduction method including statistical envelope indices and consistency criteria to merge the original multiple time periods into a variable-scale time period, establish a variable-scale time-segment model, and solve for a coarse-grained solution, including: With the objective function of minimizing the total operating cost and start-up cost in each variable-scale period, constraints are introduced, including system load constraints, unit output upper and lower limit constraints, ramp rate constraints, system positive and negative reserve constraints, line transmission capacity constraints, and minimum continuous start-up and shutdown time of units. Among them, the system load constraints cover the average, maximum, and minimum system load in the variable-scale period and the maximum increase or decrease in load in adjacent original periods. The obtained predicted solution is used as the initial solution of the variable-scale time period model. It is substituted into the model for iterative solution to obtain the unified start-up and shutdown variables and output variables of each unit in each variable-scale time period. The variables constitute a coarse-grained solution. The start-up and shutdown status and output of each unit remain unchanged in the variable-scale time period. A single variable-scale time period corresponds to one or more original time periods. The statistical envelope indicators include load curvature envelope, ramp cumulative pressure index, start-stop trigger risk envelope, and standby structure stability index. The consistency criteria include numerical stability criteria, structural fluctuation criteria, discrete decision invariance criteria, and comprehensive risk consistency criteria. The coarse-grained solution refinement and reconstruction module is configured to: expand the coarse-grained solution into an initial fine-grained solution; reduce the model size through effective variable identification and redundant constraint identification; and sequentially perform start-stop structure adjustment, intra-scale feasibility repair, and inter-scale feasibility repair to obtain a refined feasible solution that satisfies all physical constraints of the original time period, thus completing the large-scale electricity spot market clearing solution; the reduction of model size through effective variable identification and redundant constraint identification includes: The relaxed feasible region is obtained by continuously relaxing the original mixed integer clearing model. The Lagrangian function of the relaxed model is constructed and the unit structural perturbation direction of the start and stop variables is defined. The structural performance index is constructed by constraining the activation indicator function, the intensity of the constraint activation response, and the equivalent second-order structural response of the objective function with respect to the start and stop variables. For the coupling constraints in the model, a process of first relaxing and then adding is adopted. By solving the relaxation subproblem with the goal of maximizing the power flow expression of a specific line and the condition of the basic system operation constraints, the theoretical and lower bounds of the power flow of the line in the feasible domain of the whole network are calculated. By comparing the calculated theoretical upper and lower bounds with the actual transmission capacity limit of the line, if the theoretical upper bound is lower than the line capacity upper bound, the upper bound constraint is determined to be a redundant constraint; if the theoretical lower bound is higher than the line capacity lower bound, the lower bound constraint is determined to be a redundant constraint. Redundant constraints are eliminated, and non-redundant constraints are reintroduced into the corresponding repair model to simplify the constraint set.

5. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps of the adaptive accelerated solution method for electricity spot market clearing as described in any one of claims 1-3.

6. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the adaptive accelerated solution method for electricity spot market clearing as described in any one of claims 1-3.

Images