Method and system for predicting unit combination based on fusion of physical graph network and cross attention

Abstract

The application discloses a kind of unit combination pre-judgment method and system fusing physical graph network and cross attention, the method obtains target and historical scheduling period system data and unit data, constructs unit combination model and power grid physical information graph, and constructs training data set with physical information graph as input, historical start-stop plan as label;Establish a prediction model containing a space-time graph convolution module, a spatial graph convolution module, a cross attention module and a variable mapping module, extract the time sequence characteristics of power grid load and unit topology characteristics respectively, realize feature interaction and fusion through cross attention, and output unit start-stop state prediction probability;Based on confidence threshold and fixed variable quantity mechanism, fix part of start-stop variables, convert the original model into a simplified model, thereby reducing the scale of solution, and speeding up the optimization solution of unit combination problem.

Description

Technical Field

[0001] This invention belongs to the field of power system technology, and in particular relates to a method and system for predicting unit combination by integrating physical graph networks and cross-attention. Background Technology

[0002] In the field of power systems, the Unit Commitment (UC) problem is a key optimization task involving generator start-up and shutdown decisions and power allocation. Its mathematical model is Mixed Integer Linear Programming (MILP), a typical NP-hard problem. With the high proportion of renewable energy grid integration and the deepening reform of the power market, the complexity of the UC problem has significantly increased: the participation of multiple stakeholders leads to an increase in dispatchable objects, the strong volatility of renewable energy requires more granular scheduling time, and the multi-stage clearing mechanism of the spot market places higher demands on solution efficiency. Currently, power market clearing relies heavily on commercial solvers (such as Gurobi and CPLEX), but these solvers are prone to the curse of dimensionality and combinatorial explosion when dealing with large-scale provincial power grid UC problems, resulting in a significant decrease in solution efficiency, which has become a potential bottleneck restricting the operational efficiency of the power market.

[0003] Currently, solutions for enhancing UC problem solving using artificial intelligence technology mainly fall into two categories: one is the end-to-end direct solution method, which attempts to bypass the traditional optimization process by training a neural network to fit the mapping relationship between input parameters and output solutions; the other is the fusion enhancement method, which helps the solver make optimizations in key decision-making stages by mining historical data features. For example, it can predict the start-up and shutdown status of the unit based on deep learning and fix variables according to confidence levels, thereby simplifying the problem size and reducing the solution space.

[0004] However, existing technologies still have significant shortcomings. End-to-end methods struggle to guarantee the feasibility and security of prediction results, and the output solutions often fail to meet power system operational constraints. Fusion enhancement methods, on the other hand, lack sufficient representation capabilities for multi-source heterogeneous power grid data, failing to fully utilize the power grid's physical topology (such as unit electrical connections and transmission line capacity limitations) and spatiotemporal evolution patterns, resulting in limited prediction accuracy. Furthermore, existing variable fixing mechanisms are relatively simple and lack fine-grained control, easily leading to model infeasibility due to too many fixed variables, or ineffective solution acceleration due to too few fixed variables. These problems restrict the engineering application of artificial intelligence technology in the electricity spot market, necessitating an accelerated solution method that can deeply integrate power grid physical characteristics, accurately predict start-up and shutdown states, and possess a robust variable fixing mechanism. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method and system for predicting unit combination by integrating physical graph networks and cross-attention, which can deeply integrate the physical topology and spatiotemporal characteristics of the power grid, achieve high-precision start-up and shutdown state prediction and robust variable fixation, and significantly improve the solution efficiency while ensuring the feasibility of the solution.

[0006] Technical solution: The unit combination prediction method integrating physical graph networks and cross-attention as described in this invention includes the following steps:

[0007] Unit Combination Data Acquisition: Acquire system data and unit data of the power system within the target scheduling cycle and historical scheduling cycles; Based on the system data and unit data, construct a unit combination problem model and a power grid physical information map for the corresponding scheduling cycle; Using data from historical scheduling cycles, with the power grid physical information map as input features and unit start-up and shutdown plans as output labels, construct an artificial intelligence training dataset;

[0008] Generator unit start-up and shutdown status prediction: A prediction model integrating physical information graph network and cross-attention is constructed. The prediction model includes a spatiotemporal graph convolution module, a spatial graph convolution module, a cross-attention module, and a variable mapping module. Spatiotemporal graph data from the power grid physical information graph are input into the spatiotemporal graph convolution module, and spatial graph data from the power grid physical information graph are input into the spatial graph convolution module. The outputs of the spatiotemporal graph convolution module and the spatial graph convolution module are input into the cross-attention module for feature interaction fusion. The output of the cross-attention module is input into the variable mapping module to generate the predicted probability of generator unit start-up and shutdown status.

[0009] Unit start-stop status fixed: Based on the predicted probability, according to the confidence threshold and the variable fixing mechanism of the total number of fixed variables, some generator start-stop status variables in the unit combination problem model of the target scheduling cycle are fixed to 0 or 1, and the original-scale unit combination problem model is transformed into a simplified-scale unit combination problem model to accelerate the solver solution.

[0010] This invention achieves high-precision prediction and robust variable fixation of unit start-up and shutdown states by deeply integrating the physical topology of the power grid with multi-dimensional temporal features, significantly improving the solution efficiency of unit combination problems. First, the unit combination data acquisition step constructs a dataset linked by the physical information graph of the power grid, providing an input foundation for subsequent models that combines physical structure and scheduling cycle characteristics. Second, the unit start-up and shutdown state prediction step utilizes a model that integrates physical graph networks and cross-attention. A spatiotemporal graph convolution module captures the evolutionary patterns of power grid nodes in the time dimension, a spatial graph convolution module extracts spatial topological relationships, and a cross-attention module achieves dynamic interaction and enhancement of multi-dimensional features, thereby generating accurate and reliable start-up and shutdown probability predictions. Finally, the unit start-up and shutdown state fixation step introduces a fixation mechanism based on the predicted probability, incorporating confidence thresholds and constraints on the total number of variables. This simplifies the original large-scale combinatorial optimization problem while ensuring no loss of feasible solution space, allowing the solver to quickly focus on key decision variables. This significantly improves the overall solution efficiency of unit combination problems while ensuring both feasibility and optimality of the solution.

[0011] Preferably, the unit combination problem model is a mixed-integer linear programming model, and its objective function is to minimize the operating cost and start-up cost of the units in the power system.

[0012]

[0013] Where T is the total number of time periods in the corresponding scheduling cycle; N is the number of generating units; M is the number of new energy generating units; S i,t and C i,t Let F be the start-up and shutdown costs of the i-th generator unit in time period t; i,t Let be the power generation cost of the i-th generator unit in time period t;

[0014] The constraints of the unit combination problem model include:

[0015] Start-up and shutdown cost calculation constraints:

[0016]

[0017]

[0018] Among them, I i,t K represents the start-up and shutdown status of the i-th generator unit during time period t, where 1 indicates start-up and 0 indicates shutdown; i J represents the single start-up cost of the i-th generator set; i Let be the cost of a single shutdown of the i-th generator unit;

[0019] Output upper and lower limit constraints:

[0020]

[0021] Among them, P i,t P represents the output of the i-th generator unit during time period t; i,min and P i,max These are the minimum and maximum outputs of the i-th generator set, respectively;

[0022] Slope climbing and landslide constraints:

[0023]

[0024]

[0025] Where, r u,i and r d,i Let be the gradeability and landslide rate of the i-th generator unit, respectively;

[0026] Initial state constraints:

[0027]

[0028]

[0029] Among them, U i D represents the number of time periods during which the i-th generator unit needs to remain running from its initial state; i T represents the number of time periods during which the i-th generator unit needs to remain offline in its initial state; on,i and T off,i These represent the minimum start-up time and minimum downtime of the i-th generator unit, respectively; I i,0 This represents the initial start-stop state of the i-th generator unit;

[0030] Minimum boot time constraint:

[0031]

[0032]

[0033] Minimum downtime constraint:

[0034]

[0035]

[0036] System power balance constraints:

[0037]

[0038] Among them, P Load,t Let t be the system load during time period t;

[0039] System spin-off standby constraints:

[0040]

[0041]

[0042] Among them, S u,i,t and S d,i,t Δt represents the positive and negative spinning reserve capacity that generator i can provide during time period t; Δt is the spinning reserve response time of generator i; Δu and Δd are the upper and lower deviations of load forecast, respectively.

[0043] By constructing a mixed-integer linear programming model with the objective of minimizing overall operating costs, and systematically integrating key physical and economic constraints such as generator start-up and shutdown costs, output range, ramp rate, minimum start-up and shutdown time, system power balance, and spinning reserve, a rigorous and comprehensive mathematical characterization of the unit combination problem is provided. This model not only accurately reflects the operating characteristics and cost structure of thermal power units, but also ensures the accurate calculation of start-up and shutdown operation costs through start-up and shutdown cost calculation constraints; it strictly guarantees the physical feasibility and safety of unit operation through upper and lower output limits, ramp and slippage constraints, and minimum start-up and shutdown time constraints; simultaneously, the integration of system power balance and spinning reserve constraints effectively addresses load fluctuations and forecast uncertainties, enhancing the robustness of the dispatching scheme. This complete constraint system provides high-quality training labels and optimization foundations that conform to actual operating rules for subsequent start-up and shutdown state prediction based on physical graph networks, ensuring that both the prediction results and the solver's solution process strictly adhere to the physical laws and operating requirements of the power system.

[0044] Preferably, the power grid physical information map includes a spatiotemporal map and a spatial map, used to characterize the spatiotemporal and spatial characteristics of the power system;

[0045] The spatiotemporal diagram is defined as It is a dynamic graph used to characterize the changes in grid load and renewable energy output over time, where D is the corresponding set of bus nodes. E is equal to the number of busbars, where E is the corresponding set of edges. The number of transmission lines is equal to the number of lines, and A is the adjacency matrix of the graph; the net load characteristic of bus node i is represented as... ,in This represents the system load on the i-th bus. This represents the new energy output on the i-th bus; the spatial diagram is defined as follows: , is a static graph used to characterize the inherent properties of generator sets and transmission lines in a power grid, where G is the corresponding set of generator set nodes. E is equal to the number of generator sets, and E is the corresponding set of edges. The number of transmission lines is equal to the number of nodes in the graph, and A is the adjacency matrix of the graph; the physical characteristics of generator node i are represented as follows: ,in This represents the segmented power generation capacity of the i-th generator unit. This represents the segmented price quote for the i-th generator set. and Let i represent the gradeability and landslide rate of the i-th generator unit, respectively. This indicates the initial start-stop state of the i-th generator unit. Let m represent the initial active power output of the i-th generator unit; the physical characteristics of edge m are represented as follows: ,in and Let represent the maximum forward and negative transmission capacities of the m-th transmission line, respectively. and Let represent the reactance and susceptance of the m-th transmission line, respectively; the adjacency matrix A is defined as follows: if node i is connected to node j, then ,otherwise .

[0046] By constructing a spatiotemporal graph and a spatial graph, a decoupled representation and deep integration of the dynamic operating characteristics and static inherent attributes of the power system are achieved. The spatiotemporal graph, with the bus node as its core, dynamically depicts the temporal evolution of load and renewable energy output through net load characteristics, providing the model with time-dimensional information to capture system operating trends. The spatial graph, with generator units and transmission lines as nodes, fully presents the regulation capacity of power generation resources and the transmission bottlenecks of the power grid through static descriptions of unit physical characteristics and line electrical parameters. This dual-graph structure allows subsequent prediction models to extract features from both spatiotemporal evolution and spatial topology levels, and accurately reflect the physical connections between nodes through adjacency matrices. This provides a clear and comprehensive input foundation for high-precision predictions that integrate physical graph networks and cross-attention, ensuring the physical logic and credibility of start-up and shutdown state prediction results.

[0047] Preferably, the spatiotemporal graph convolution module includes two spatiotemporal convolutional blocks and a first fully connected layer, used to capture the dynamic patterns and complex dependencies that evolve over time between different nodes;

[0048] Each spatiotemporal convolutional block adopts an architecture with a spatial graph convolutional layer in the middle and a temporally gated convolutional layer connected in series before and after it.

[0049] The time-gated convolutional layer is used to extract the temporal features of the bus net load, consisting of a width of K. t It consists of a one-dimensional causal convolution kernel and gated linear units, and performs temporal convolution operation on the load time series of each node in the graph:

[0050]

[0051] Where D represents the node feature tensor input in the spatiotemporal graph; W and V are the tensors of size . The convolution kernel, K t Where is the width of the time-gated convolution kernel, and T is the time step. Output channel dimension; Indicates the convolution operation; This indicates element-wise multiplication; This represents the sigmoid gate function;

[0052] The spatial graph convolutional layer is used to capture the dynamic impact of adjacent buses on the current bus load. It is composed of a Chebyshev spectral convolutional network and aggregates spatial information for each node in the graph.

[0053]

[0054] in, This is the node embedding matrix output by the spatial graph convolutional layer; The size of the convolution kernel; This is the k-th weight matrix; This is the k-th order Chebyshev polynomial obtained by the recursive formula; The recursive calculation formula is:

[0055]

[0056] in, , The node embedding matrix at time t is the output of the time-gated convolutional layer; The scaled and normalized Laplace matrix is ​​defined as follows:

[0057]

[0058] Where L is the original Laplacian matrix, A is the adjacency matrix of the graph; L is the largest eigenvalue of the Laplacian matrix L; I is the identity matrix.

[0059] When using a first-order approximation, the spatial graph convolutional layer simplifies to:

[0060]

[0061] in, This represents the normalized adjacency matrix; and These are learnable weight parameters.

[0062] By employing a dual-block stacking structure and a collaborative design of in-layer gated temporal convolution and spectral graph convolution, a deep coupled modeling of the dynamic spatiotemporal characteristics of the power grid is achieved. Its temporally gated convolutional layer utilizes causal convolution and gated linear units to efficiently extract the evolution of bus net load over time while ensuring the causality of temporal dependencies. The spatial graph convolutional layer, based on a Chebyshev spectral convolutional network, effectively aggregates spatial correlation information between adjacent buses through spectral decomposition of the Laplace matrix and polynomial approximation, accurately capturing the impact of topology on load dynamics. The simplified form under the first-order approximation reduces computational complexity while maintaining core aggregation capabilities. This spatiotemporally interwoven feature extraction mechanism enables the model to fully learn the complex dependencies between power grid nodes that change over both time and space, providing node representations rich in dynamic spatiotemporal semantics for subsequent cross-attention modules. This significantly improves the prediction accuracy and robustness of unit start-up and shutdown status trends.

[0063] Preferably, the spatial graph convolution module includes two edge-conditional convolutional layers and a second fully connected layer, used to identify and integrate spatial dependencies and interpret the interaction mechanism between nodes based on the features of the connected edges;

[0064] The edge-conditional convolutional layer is used to process graphs with edge features, propagating and aggregating generator set features based on transmission line features; in this layer, node features... Update to features using the following operations. :

[0065]

[0066] in, This represents the input feature vector of node i; This represents the updated feature vector of node i; This is the weight matrix; Represents the set of neighboring nodes of node i; This represents the input feature vector of neighbor node j; This represents a multilayer perceptron neural network used for propagating edge features. Its learnable parameters; This represents the feature vector of the edge connecting node i and node j.

[0067] By stacking edge-conditional convolutional layers, a deep fusion and interactive modeling of generator unit attributes and line features in the power grid spatial topology is achieved. This module utilizes a learnable multilayer perceptron to dynamically process the electrical parameters of transmission lines as edge features. This allows node features to fully consider the constraints and influences of physical attributes such as line transmission capacity and reactance on energy interaction when aggregating neighbor information, thereby accurately depicting the coupling relationship between generator units based on the actual power grid topology. This edge-feature-aware graph convolution mechanism breaks through the limitation of traditional graph convolution that only utilizes adjacency relationships. It enables the model to identify differentiated spatial dependencies caused by line physical characteristics, providing generator unit node representations rich in physical topological semantics for the subsequent cross-attention module, significantly enhancing the prediction accuracy of spatial constraints in generator start-up and shutdown decisions.

[0068] Preferably, the cross-attention module adopts a bidirectional cross-attention mechanism, which realizes the interactive fusion between features through attention calculations in two directions: from the spatiotemporal graph to the spatial graph and from the spatial graph to the spatiotemporal graph.

[0069] The calculation process of the cross-attention module is as follows:

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076] In this context, the subscripts st and s represent the spatiotemporal graph and the spatial graph, respectively; The output features of the spatiotemporal graph convolution module; The output features of the spatial graph convolution module; norm(⋅) represents the normalization operation; , , These are the weight parameters for the query matrix, key matrix, and value matrix, respectively. , , These are the corresponding bias parameters; Qs and Qs are query matrices generated from spatiotemporal graph features and spatial graph features, respectively; and These are the key matrices generated from spatiotemporal graph features and spatial graph features, respectively; and These are value matrices generated from spatiotemporal graph features and spatial graph features, respectively.

[0077] Calculate the attention output in both directions:

[0078]

[0079]

[0080] in, Attention output from the spatiotemporal map to the spatial map; Attention output in the direction from the spatial map to the spatiotemporal map; It is a normalized exponential function; This is a scaling factor used to prevent the inner product from becoming too large;

[0081] The attention outputs in both directions are residually connected to the original input features respectively:

[0082]

[0083]

[0084] in, This is the output from the spacetime map to the spatial map. This is the output from the spatial map to the spatiotemporal map.

[0085] Will and The features are concatenated along the feature dimension to form a fused feature vector, and then the two features are fused through a third fully connected layer.

[0086] Through a bidirectional interaction mechanism, deep coupling and synergistic enhancement of dynamic features in the spatiotemporal graph and static attributes in the spatial graph are achieved. This module first maps the two types of heterogeneous features into query and key-value matrices, respectively. Through attention calculations in both directions—from the spatiotemporal graph to the spatial graph and from the spatial graph to the spatiotemporal graph—dynamic temporal information can focus on static physical attributes to identify key constraint nodes, while static topological features can perceive temporal changes to adjust dependency weights. The introduction of residual connections ensures the effective preservation of original information and avoids feature degradation in deep networks. Finally, feature concatenation and fully connected layers achieve the organic fusion of the two types of semantics. This bidirectional interaction mechanism enables the model to understand the power grid operation patterns from both spatiotemporal evolution and physical topology dimensions, providing a rich and semantically aligned unified feature representation for accurate prediction of the final start-up and shutdown states, significantly improving the modeling capability for unit behavior patterns under complex coupling relationships.

[0087] Preferably, the variable mapping module consists of a Chebyshev convolutional layer and a fourth fully connected layer;

[0088] The Chebyshev convolutional layer is used to propagate information from the fused features of the nodes in the graph to obtain aggregated features;

[0089] The fourth fully connected layer transforms the aggregated features to generate predicted probabilities of unit start-up and shutdown states.

[0090] By cascading Chebyshev convolutional layers and fully connected layers, a precise mapping from fused features to unit start-up and shutdown probabilities is achieved. The Chebyshev convolutional layers, based on graph theory, utilize Chebyshev polynomials to recursively aggregate and propagate information from the fused multidimensional features within the node neighborhood, effectively capturing long-range dependencies and local smoothness between nodes in the graph, ensuring that the topological context information upon which start-up and shutdown decisions depend is fully integrated. Subsequently, the fourth fully connected layer performs nonlinear transformations and dimensionality compression on the aggregated high-order features, generating a probability output reflecting the start-up and shutdown possibilities of each generator unit. This design enables the model to comprehensively consider the power grid physical topology, dynamic temporal information, and inherent unit attributes, ultimately outputting a physically consistent and highly discriminative start-up and shutdown probability distribution, providing a reliable basis for subsequent variable fixation based on confidence thresholds.

[0091] Preferably, the unit start-up and shutdown state prediction step further includes training the prediction model:

[0092] Construct a loss function, treat the prediction of unit start-up and shutdown status as a binary classification task, minimize the difference between the model output and the true label, so that the unit start-up and shutdown status predicted by the model is as close as possible to the unit start-up and shutdown plan in the optimal solution.

[0093] The specific form of the loss function is as follows:

[0094]

[0095] in, The value of the loss function. This represents all learnable parameters of the prediction model; N is the total number of instances in the training dataset; n represents the number of binary variables in a single instance, i.e., the total number of generator start-stop state variables to be predicted; i is the instance index; d is the variable index. This represents the probability that the model predicts the value of the d-th variable to be 1 in the i-th instance; represents the true label of the d-th variable in the i-th instance, and the true label is derived from the unit start-up and shutdown plan in the optimal solution of the historical scheduling cycle unit combination problem; The loss function represents the natural logarithm function; the prediction model is trained based on the training dataset and the loss function, and the loss function value is minimized through an optimization algorithm. The training of the unit start-up and shutdown state prediction model integrating physical information graph network and cross attention was completed, and the model structure and parameters were saved to obtain the trained unit start-up and shutdown state prediction model integrating physical information graph network and cross attention.

[0096] By constructing a binary cross-entropy loss function, the prediction of start-up and shutdown states is transformed into minimizing the difference between the model's output probability and the actual start-up and shutdown plans in the historical optimal solution, thereby guiding the model to learn the intrinsic laws of optimal power system scheduling decisions. This loss function measures prediction bias in both instance and variable dimensions, enabling the model to perform refined optimization for the start-up and shutdown states of each generator unit, ensuring a high degree of consistency between the probability output and the binary decision in the optimal solution. Based on iterative training and optimization algorithms using a large-scale historical scheduling dataset, the model can effectively converge and fully learn the complex mapping relationship between the power grid's physical topology, spatiotemporal characteristics, and generator start-up and shutdown decisions. The final saved model structure and parameters provide reliable and highly generalizable decision support for accurate start-up and shutdown probability prediction in subsequent target scheduling cycles, significantly improving the accuracy and robustness of fixed variables.

[0097] Preferably, the variable fixing mechanism for the confidence threshold and the total number of fixed variables includes:

[0098] The power grid physical information graph of the target scheduling cycle is input into the trained fusion physical information graph network and cross-attention generator start-up and shutdown state prediction model to obtain the prediction probability of the start-up and shutdown state of each generator unit.

[0099] Set the first confidence threshold. Second confidence threshold and the total number of the first fixed variables The second total number of fixed variables ;

[0100] Predicted probability of generator start-up and shutdown states Where i is the generator set index and t is the time period index: when When, variable I i,t The value is fixed at 1; when When, variable I i,t The value is fixed at 0;

[0101] If satisfied If the number of variables is greater than the total number of the first fixed variables k1, then only the values ​​of the first k1 variables with the highest predicted probabilities are fixed to 1; if the following conditions are met... If the number of variables is greater than the total number of the second fixed variables k0, then only the values ​​of the first k0 variables with the lowest prediction probabilities are fixed to 0;

[0102] Where k1 and k0 are the parameters for the total number of fixed variables; The model outputs the predicted probability of the start-up / shutdown state of the i-th generator unit in time period t; I i,t Let be the start / stop state variable of the i-th generator unit during time period t.

[0103] By employing a dual control mechanism of confidence thresholds and total variable constraints, high-confidence screening and robust dimensionality reduction for unit start-up and shutdown states are achieved. This mechanism first sets high and low thresholds based on predicted probabilities, fixing variables with probabilities above the upper limit as start-up and those below the lower limit as shutdown, thus prioritizing the most confident decisions of the model. Simultaneously, a fixed total variable limit is introduced; when the number of high-confidence variables exceeds a preset upper limit, only the top few variables with the most extreme probabilities are fixed, effectively avoiding the loss of feasible solution space due to over-fixing. This threshold and total variable constraint strategy fully utilizes the confidence information of the prediction model while ensuring the relaxation and solution flexibility of the optimization problem through quantity control. While ensuring that the simplified unit combination model still contains the optimal solution, it minimizes the size of the binary variables to be solved, creating favorable conditions for rapid solver convergence.

[0104] Secondly, the unit combination prediction system integrating physical graph networks and cross-attention as described in this invention includes:

[0105] The unit combination data acquisition module is used to acquire system data and unit data of the power system within the target scheduling cycle and historical scheduling cycle; based on the system data and unit data, it constructs a unit combination problem model and power grid physical information map data for the corresponding scheduling cycle; using data from historical scheduling cycles, with the power grid physical information map data as input features and the unit start-up and shutdown plans as output labels, it constructs an artificial intelligence training dataset.

[0106] The generator start-up and shutdown state prediction module is used to construct a prediction model that integrates a physical information graph network and cross-attention. The prediction model includes a spatiotemporal graph convolution module, a spatial graph convolution module, a cross-attention module, and a variable mapping module. Spatiotemporal graph data from the power grid physical information graph is input to the spatiotemporal graph convolution module, and spatial graph data from the power grid physical information graph is input to the spatial graph convolution module. The outputs of the spatiotemporal graph convolution module and the spatial graph convolution module are input to the cross-attention module for feature interaction fusion. The output of the cross-attention module is input to the variable mapping module to generate the predicted probability of the generator start-up and shutdown states.

[0107] The unit start-stop state fixing module is used to fix some generator start-stop state variables in the unit combination problem model of the target scheduling cycle to 0 or 1 based on the predicted probability and according to the variable fixing mechanism of confidence threshold and total number of fixed variables, so as to transform the original-scale unit combination problem model into a simplified-scale unit combination problem model, thereby accelerating the solver's solution.

[0108] Thirdly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed by the unit combination prediction method of the fusion physical graph network and cross attention.

[0109] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for predicting unit combinations by integrating physical graph networks and cross-attention.

[0110] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. This invention constructs a prediction model that integrates a physical information graph network and cross-attention. It utilizes spatiotemporal graph convolution modules and spatial graph convolution modules to mine the temporal evolution of power grid load and the spatial topological dependence of generator units, respectively. Then, a cross-attention module enables deep interaction and adaptive fusion of these two types of features. This mechanism can comprehensively capture the complex physical dynamics of the power grid, thereby significantly improving the prediction accuracy of unit start-up and shutdown states, laying a reliable foundation for subsequent variable fixing. 2. Based on high-confidence prediction probabilities, this invention designs a variable fixing mechanism that combines a confidence threshold with the total number of fixed variables. By prioritizing the fixing of high-confidence start-up and shutdown state variables and controlling the total number of fixed variables, it can effectively avoid the infeasibility or suboptimal problems of unit combination models caused by oversimplification. While ensuring the quality of the solution, it transforms the original large-scale mixed integer programming model into a simplified model, thereby achieving several times the speedup on commercial solvers. 3. This invention innovatively combines an artificial intelligence prediction model with a traditional mathematical programming solver, forming an end-to-end intelligent decision-making architecture. This architecture utilizes historical data to train models that predict key variables, guiding the solver to focus on optimizing complex variables. It leverages the advantages of data-driven approaches in pattern recognition while retaining the rigor of mathematical programming in ensuring optimality, making the solution process more transparent, reliable, and efficient. 4. This invention simultaneously constructs a power grid physical information graph containing both spatiotemporal and spatial graphs, transforming multi-source, heterogeneous system data into a unified graph structure feature input. This data modeling method is naturally adaptable to power systems of different sizes and topologies, allowing the trained AI model to be applied to new networks without adjustment. It is also seamlessly compatible with various commercial solvers such as Gurobi and CPLEX, demonstrating good generalization ability and practical engineering value. Attached Figure Description

[0111] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0112] Figure 2 This is a schematic diagram of the prediction model that integrates physical information graph networks and cross-attention according to the present invention. Detailed Implementation

[0113] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0114] This invention provides a method for predicting crew combinations that integrates physical graph networks and cross-attention, such as... Figure 1 As shown, it includes the following steps:

[0115] The detailed technical solution of this invention includes the following process:

[0116] 1. Constructing models and graphical data for the UC problem in power systems.

[0117] The system and generating unit data of the power system are acquired within the target dispatch cycle and historical dispatch cycle. This data is collected by the power system. System data includes: system active load, positive reserve capacity, and negative reserve capacity demand. Generating unit data includes: the number of generating units and renewable energy units, the rated capacity and minimum technical output of generating units, the active power output of renewable energy units, the maximum ramp rate and maximum runaway rate of generating units, the minimum operating duration and minimum downtime of generating units, and the starting price, downtime price, and tiered pricing for generating units.

[0118] Based on the system data and unit data within the corresponding scheduling cycle, construct the UC problem model and power grid physical information diagram data for the corresponding scheduling cycle.

[0119] The UC problem model is a mixed-integer linear programming model, which is constructed by substituting system data and unit data within the corresponding scheduling cycle.

[0120] The objective function of the UC problem model is to minimize the operating cost and start-up cost of the units in the power system.

[0121]

[0122] The constraints of the UC problem model include start-up and shutdown cost calculation constraints, output upper and lower limit constraints, ramp and landslide constraints, initial state constraints, minimum start-up time constraints, minimum downtime constraints, system power balance constraints, and system spinning reserve constraints.

[0123] Start-up and shutdown cost calculation constraints:

[0124]

[0125]

[0126] Output upper and lower limit constraints:

[0127]

[0128] Slope climbing and landslide constraints:

[0129]

[0130]

[0131] Initial state constraints:

[0132]

[0133]

[0134] Minimum boot time constraint:

[0135]

[0136]

[0137] Minimum downtime constraint:

[0138]

[0139]

[0140] System power balance constraints:

[0141]

[0142] System spin-off standby constraints:

[0143]

[0144]

[0145] T represents the total number of time periods in the corresponding scheduling cycle; N represents the number of generator units; M represents the number of new energy generator units; S i,t and C i,t Let F be the start-up and shutdown costs of the i-th generator unit in time period t; i,t Let I be the power generation cost of the i-th generator unit in time period t; i,t P represents the start-up and shutdown status of the i-th generator unit during time period t (1 indicates start-up, 0 indicates shutdown); i,t P represents the output of the i-th generator unit during time period t; i,min and P i,maxThese are the minimum and maximum outputs of the i-th generator unit, respectively; r u,i and r d,i These represent the gradeability and landslide rate of the i-th generator unit, respectively; T on,i and T off,i These represent the minimum start-up time and minimum downtime of the i-th generator unit, respectively; P Load,t Let be the system load during time period t; s u,i,t and s d,i,t Δt represents the positive and negative spinning reserve capacity that generator i can provide during time period t; Δt is the spinning reserve response time of generator i; Δu and Δd are the upper and lower deviations of load forecast, respectively.

[0146] A commercial solver is invoked to solve the UC problem model for the historical scheduling cycle, and the unit start-up and shutdown plan within the historical scheduling cycle is determined based on the optimal solution obtained.

[0147] The power grid physical information map consists of a spatiotemporal diagram and a spatial diagram, constructed by extracting system and generating unit data within the corresponding scheduling cycle. It is used to characterize the spatiotemporal and spatial characteristics of the power system. The spatiotemporal diagram is a dynamic diagram that focuses on the changes in grid bus load and renewable energy output over time, defined as follows: A spatial diagram is a static diagram that focuses on the inherent properties of generator units and transmission lines in a power grid, and is defined as follows: .in It is the corresponding set of bus nodes. The number of busbars represents the net load characteristic of the busbars, as defined. ,in These represent the system load and new energy output on the i-th bus, respectively. It is the corresponding set of generator node nodes. Equal to the number of generator sets, representing the physical characteristics of the generator sets, defined. ,in These represent the segmented generating capacity and price of the i-th generator unit, respectively. Let i represent the gradeability and landslide rate of the i-th generator unit, respectively. This indicates the initial start-stop state of the i-th generator unit. This represents the initial active power output of the i-th generator unit. It is the corresponding set of edges. Equal to the number of transmission lines, representing the physical characteristics of transmission lines, defined. ,in Let represent the maximum forward and negative transmission capacities of the m-th transmission line, respectively. Let represent the reactance and susceptance of the m-th transmission line, respectively. Given the adjacency matrix of a graph, if the nodes... With nodes If connected, then Otherwise, it is 0.

[0148] Data from historical power system scheduling periods is used to generate the input features and output labels needed for training an artificial intelligence model. The dataset uses power grid physical information map data as input features and unit start-up and shutdown plans as output labels to train the artificial intelligence model to learn the mapping relationship between input features and output labels.

[0149] 2. Construct a prediction model that integrates physical information graph networks and cross-attention.

[0150] The constructed unit start-up and shutdown state prediction model, which integrates a physical information graph network and cross-attention, includes a spatiotemporal graph convolution module, a spatial graph convolution module, a cross-attention module, and a variable mapping module. For example... Figure 2 As shown, the spatiotemporal graph data in the power grid physical information graph is input into the spatiotemporal graph convolution module, the spatial graph data in the power grid physical information graph is input into the spatial graph convolution module, the outputs of the spatiotemporal graph convolution module and the spatial graph convolution module are input into the cross-attention module, and finally the prediction of the unit start-up and shutdown status is completed through the variable mapping module.

[0151] The spatiotemporal graph convolutional module comprises two spatiotemporal convolutional modules and a fully connected layer, effectively capturing dynamic patterns and complex dependencies that evolve over time between different nodes. Each spatiotemporal convolutional module employs a classic architecture with a spatial graph convolutional layer in the middle, preceded and followed by a temporally gated convolutional layer.

[0152] A time-gated convolutional layer is used to extract the temporal characteristics of the bus net load. It consists of layers with a width of K. t The layer consists of a one-dimensional causal convolution kernel and gated linear units. This layer performs temporal convolution operations on the load time series of each node in the graph.

[0153]

[0154] in, This represents the nodal feature tensor input to the spatiotemporal graph. Represents convolution kernels of the same size. Indicates the output channel dimension. This represents the sigmoid gate function.

[0155] The spatial graph convolutional layer is used to effectively capture the dynamic impact of adjacent buses on the current bus load. It consists of a Chebyshev spectral convolutional network, which aggregates spatial information for each node in the graph.

[0156]

[0157] in, The size of the convolution kernel is shown in the figure. For the k-th weight matrix, The k-th order Chebyshev polynomial is calculated by the recursive formula. :

[0158]

[0159] in, This represents the node embedding matrix at time t output by the time-gated convolutional layer. Represented as the Laplacian matrix of the graph after scaling and normalization. It is the Laplace matrix The largest eigenvalue. For the original Laplace matrix, . Let be the degree matrix of the graph. In the specific implementation, a first-order approximation is used to simplify the calculation:

[0160]

[0161] in, This represents the normalized adjacency matrix. This approximation significantly reduces computational complexity while maintaining performance.

[0162] The spatial graph convolution module contains two edge-conditional convolutional layers and one fully connected layer, which can effectively identify and integrate spatial dependencies and interpret the interaction mechanism between nodes based on the features of the connected edges.

[0163] Edge-conditional convolutional layers are used to process graphs with edge features, enabling the propagation and aggregation of generator unit features based on transmission line features. In this layer, node features... Update to features using the following operations. :

[0164]

[0165] in, Represents the weight matrix. Represents a node The set of neighboring nodes, This represents a multilayer perceptron neural network used for propagating edge features. Indicates the connection node and The eigenvectors of the edges.

[0166] The cross-attention module employs a bidirectional cross-attention mechanism, achieving interactive fusion between features through attention calculations in two directions (spatiotemporal-spatial and spatial-spatiotemporal). The calculation processes for the spatiotemporal-spatial attention output and the spatial-spatiotemporal attention output are shown below:

[0167]

[0168]

[0169]

[0170]

[0171]

[0172]

[0173]

[0174]

[0175] Among them, subscript and These represent the spacetime diagram and the spatial diagram, respectively. This represents the output of the spatiotemporal graph network module. This represents the output of the spatial graph network module.

[0176] The outputs from both directions are residually concatenated with the original input features to preserve the original information and alleviate the gradient vanishing problem. The calculation process is as follows:

[0177]

[0178]

[0179] Representation of these two directions and The features are concatenated along the feature dimension to form a fused feature vector, and then the two features are fused through a fully connected layer.

[0180] The variable mapping module consists of Chebyshev convolutional layers and fully connected layers. First, the Chebyshev convolutional layers are used to propagate the information of the fused features of the nodes in the graph to obtain aggregated features. Then, the aggregated features are input into the fully connected layers for transformation, and finally the feature representation with variables as the core is output, which is the predicted probability of the unit start-up and shutdown status.

[0181] A loss function was designed to train a generator start-up and shutdown state prediction model that integrates a physical information graph network and cross-attention. The prediction is treated as a binary classification task, aiming to minimize the difference between the model output and the true label, ensuring that the predicted generator start-up and shutdown state is as close as possible to the optimal generator start-up and shutdown plan. The specific form of the loss function is as follows:

[0182]

[0183] The training dataset contains N instances, where n represents the number of binary variables in a single instance. This indicates the probability that the model predicts the variable will take the value 1. This represents the true label of the variable.

[0184] Based on the training dataset and the corresponding loss function, the training of the unit start-up and shutdown state prediction model integrating physical information graph network and cross attention is completed. The model structure and parameters are saved to obtain the trained unit start-up and shutdown state prediction model integrating physical information graph network and cross attention.

[0185] 3. Construct a variable fixation mechanism based on confidence threshold and fixed total number of variables.

[0186] The power grid physical information graph of the target scheduling cycle is input into the trained fusion physical information graph network and cross-attention generator start-up and shutdown state prediction model to obtain the prediction probability of the start-up and shutdown state of each generator unit.

[0187] Based on the generator start-up and shutdown states predicted by the model, the generator start-up and shutdown state variables in the UC problem model for the target scheduling cycle are fixed to 0 or 1 according to a variable fixing mechanism based on a confidence threshold and the total number of fixed variables. A confidence threshold is set. and and the total number of fixed variables and When the predicted probability of generator start-up and shutdown states is greater than When the predicted probability of generator start-up and shutdown is less than 1, the value of this variable is fixed at 1; when the predicted probability of generator start-up and shutdown is less than 1, the value of this variable is fixed at 1. When the predicted probability of generator start-up and shutdown is greater than 0, the value of this variable is fixed at 0. The number of variables is greater than the total number of fixed variables. At that time, only the one with the highest prediction probability is fixed. The value of each variable is 1; when the predicted probability of generator start-up / shutdown state is less than The number of variables is greater than the total number of fixed variables. At that time, only the one with the lowest predicted probability is fixed. The value of each variable is 0.

[0188] By fixing the values ​​of the selected high-confidence variables, the original large-scale UC problem model is transformed into a simplified UC problem model, thereby improving the solution efficiency, shortening the solution time, and obtaining the unit start-up and shutdown plan within the target scheduling cycle.

[0189] Drawing upon a similar inventive concept, embodiments of the present invention provide a system for predicting crew combination using a fusion of physical graph networks and cross-attention, corresponding to a method for predicting crew combination using fusion of physical graph networks and cross-attention, comprising:

[0190] The unit combination data acquisition module is used to acquire system data and unit data of the power system within the target scheduling cycle and historical scheduling cycle; based on the system data and unit data, it constructs a unit combination problem model and power grid physical information map data for the corresponding scheduling cycle; using data from historical scheduling cycles, with the power grid physical information map data as input features and the unit start-up and shutdown plans as output labels, it constructs an artificial intelligence training dataset.

[0191] The generator start-up and shutdown state prediction module is used to construct a prediction model that integrates a physical information graph network and cross-attention. The prediction model includes a spatiotemporal graph convolution module, a spatial graph convolution module, a cross-attention module, and a variable mapping module. Spatiotemporal graph data from the power grid physical information graph is input to the spatiotemporal graph convolution module, and spatial graph data from the power grid physical information graph is input to the spatial graph convolution module. The outputs of the spatiotemporal graph convolution module and the spatial graph convolution module are input to the cross-attention module for feature interaction fusion. The output of the cross-attention module is input to the variable mapping module to generate the predicted probability of the generator start-up and shutdown states.

[0192] The unit start-stop state fixing module is used to fix some generator start-stop state variables in the unit combination problem model of the target scheduling cycle to 0 or 1 based on the predicted probability and according to the variable fixing mechanism of confidence threshold and total number of fixed variables, so as to transform the original-scale unit combination problem model into a simplified-scale unit combination problem model, thereby accelerating the solver's solution.

[0193] The experimental examples were validated using the SG-126 testing system. The SG-126 model, originating from the State Grid Dispatch AI Innovation Competition, includes 126 buses and 54 generating units. The original dataset comprises 100,000 AC power flow sections, with power flow characteristics consistent with actual power grid operation. The maximum fluctuation range of total renewable energy output and total load can reach 20-50%. The experimental platform was configured with one Intel mobile processor (14 cores, 20 threads @ 2.3GHz) CPU, 24GB of RAM, and one NVIDIA GeForce RTX 3070 Laptop GPU. The experimental scenario was set as a day-ahead UC problem, with a time granularity of 15 minutes, corresponding to 96 time periods in 24 hours. 100 test samples under different load conditions were randomly generated and tested using the mixed integer programming algorithm of the academic version of Gurobi V12 and the commercial CPLEX V20 solver. The baseline method directly uses the commercial solver to solve the UC problem, and the performance value of the baseline method is considered as unit 1. Additional comparisons were made with the k-nearest neighbor method and the graph neural network variable prediction method. The confidence thresholds were uniformly set to p1=0.95 and p0=0.1, with the total number of fixed variables being k1=1500 and k0=7000. The ratios of different methods to the basic method were recorded to evaluate the performance impact of this invention on accelerating the solution of the UC problem.

[0194] Tables 1 and 2 compare the performance of different methods in solving UC problems in a 96-time scenario using the commercial solvers Gurobi and CPLEX as benchmarks, respectively.

[0195] Table 1. Gurobi's solution performance in the 96-time scenario of the SG126 system.

[0196] index Gurobi KNN GNN The method proposed in this invention Average proportion of fixed integer variables - 89.80% 82.65% 82.67% Total forward time (seconds) for all samples - - 58.79 1.85 Average solution speedup 1 3.50 3.13 3.53 The proportion of problems for which feasible solutions can be obtained 100% 93.8% 100% 100% Average relative error of the target value of the optimal solution <0.01% <0.01% <0.01% <0.01%

[0197] Table 2. Solving performance of CPLEX in the SG126 system under a 96-time scenario.

[0198] index CPLEX KNN GNN The method proposed in this invention Average proportion of fixed integer variables - 89.80% 82.65% 82.67% Total forward time for all samples (seconds) - - 58.79 1.85 Average solution speedup 1 4.31 4.22 4.99 The proportion of problems for which feasible solutions can be obtained 100% 93.8% 100% 100% Average relative error of the target value of the optimal solution <0.01% <0.01% <0.01% <0.01%

[0199] As can be seen from Tables 1 and 2:

[0200] The performance of this invention is significantly better than the benchmark and comparative methods, exhibiting the highest speedup ratio, the highest proportion of problems yielding feasible solutions, and the ability to fix a larger proportion of integer variables. This demonstrates that the size of integer variables is a key factor affecting the solution time of UC problems, and reducing the size of integer variables is an effective method for acceleration.

[0201] The relative error of the optimal solution objective value of the present invention remains at an extremely low level, proving that the present invention can guarantee the accuracy of solving the UC problem.

[0202] The present invention demonstrates significant acceleration on both the Gurobi and CPLEX commercial solvers, indicating its good versatility.

[0203] The present invention also discloses an electronic device.

[0204] Specifically, the electronic device can be a desktop computer, laptop computer, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. The processor and memory can be connected via a bus or other means. The processor can be a Central Processing Unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, graphics processing units (GPUs), embedded neural network processing units (NPUs) or other dedicated deep learning coprocessors, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0205] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules. The processor executes various functional applications and data processing by running non-transitory software programs, instructions, and modules stored in memory. Memory may include a program storage area and a data storage area. The program storage area may store the control unit and the application program required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, memory may include high-speed random access memory and non-transitory memory. In some embodiments, memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0206] The present invention also discloses a computer-readable storage medium.

[0207] Specifically, the computer-readable storage medium is used to store a computer program, which, when executed by a processor, implements the methods described in the above method implementation.

[0208] Those skilled in the art will understand that all or part of the processes in the methods described above can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

Claims

1. A method for predicting unit combination by integrating physical graph networks and cross-attention, characterized in that, Includes the following steps: Unit Combination Data Acquisition: Acquire system data and unit data of the power system within the target scheduling cycle and historical scheduling cycles; Based on the system data and unit data, construct a unit combination problem model and a power grid physical information map for the corresponding scheduling cycle; Using data from historical scheduling cycles, with the power grid physical information map as input features and unit start-up and shutdown plans as output labels, construct an artificial intelligence training dataset; Generator unit start-up and shutdown status prediction: A prediction model integrating physical information graph network and cross-attention is constructed. The prediction model includes a spatiotemporal graph convolution module, a spatial graph convolution module, a cross-attention module, and a variable mapping module. Spatiotemporal graph data from the power grid physical information graph are input into the spatiotemporal graph convolution module, and spatial graph data from the power grid physical information graph are input into the spatial graph convolution module. The outputs of the spatiotemporal graph convolution module and the spatial graph convolution module are input into the cross-attention module for feature interaction fusion. The output of the cross-attention module is input into the variable mapping module to generate the predicted probability of generator unit start-up and shutdown status. Unit start-stop status fixed: Based on the predicted probability, according to the confidence threshold and the variable fixing mechanism of the total number of fixed variables, some generator start-stop status variables in the unit combination problem model of the target scheduling cycle are fixed to 0 or 1, and the original-scale unit combination problem model is transformed into a simplified-scale unit combination problem model to accelerate the solver solution.

2. The method according to claim 1, characterized in that, The unit combination problem model is a mixed-integer linear programming model, and its objective function is to minimize the operating cost and start-up cost of the units in the power system. Where T is the total number of time periods in the corresponding scheduling cycle; N is the number of generating units; M is the number of new energy generating units; S i,t and C i,t Let F be the start-up and shutdown costs of the i-th generator unit in time period t; i,t Let be the power generation cost of the i-th generator unit in time period t; The constraints of the unit combination problem model include: Start-up and shutdown cost calculation constraints: ; Among them, I i,t K represents the start-up and shutdown status of the i-th generator unit during time period t, where 1 indicates start-up and 0 indicates shutdown; i J represents the single start-up cost of the i-th generator set; i Let be the cost of a single shutdown of the i-th generator unit; Output upper and lower limit constraints: Among them, P i,t P represents the output of the i-th generator unit during time period t; i,min and P i,max These are the minimum and maximum outputs of the i-th generator set, respectively; Slope climbing and landslide constraints: ; ; where r u,i and r d,i Let be the gradeability and landslide rate of the i-th generator unit, respectively; Initial state constraints: ; Among them, U i D represents the number of time periods during which the i-th generator unit needs to remain running from its initial state; i T represents the number of time periods during which the i-th generator unit needs to remain offline in its initial state; on,i and T off,i These represent the minimum start-up time and minimum downtime of the i-th generator unit, respectively; I i,0 This represents the initial start-stop state of the i-th generator unit; Minimum boot time constraint: ; Minimum downtime constraint: ; System power balance constraints: Among them, P Load,t Let t be the system load during time period t; System spin-off standby constraints: ; Among them, S u,i,t and S d,i,t Δt represents the positive and negative spinning reserve capacity that generator i can provide during time period t; Δt is the spinning reserve response time of generator i; Δu and Δd are the upper and lower deviations of load forecast, respectively.

3. The method according to claim 1, characterized in that, The power grid physical information diagram includes a spatiotemporal diagram and a spatial diagram, which are used to characterize the spatiotemporal and spatial characteristics of the power system. The spatiotemporal diagram is defined as It is a dynamic graph used to characterize the changes in grid load and renewable energy output over time, where D is the corresponding set of bus nodes. E is equal to the number of busbars, where E is the corresponding set of edges. The number of transmission lines is equal to the number of lines, and A is the adjacency matrix of the graph; the net load characteristic of bus node i is represented as... ,in This represents the system load on the i-th bus. This represents the new energy output on the i-th bus; the spatial diagram is defined as follows: , is a static graph used to characterize the inherent properties of generator sets and transmission lines in a power grid, where G is the corresponding set of generator set nodes. E is equal to the number of generator sets, and E is the corresponding set of edges. The number of transmission lines is equal to the number of nodes in the graph, and A is the adjacency matrix of the graph; the physical characteristics of generator node i are represented as follows: ,in This represents the segmented power generation capacity of the i-th generator unit. This represents the segmented price quote for the i-th generator set. and Let i represent the gradeability and landslide rate of the i-th generator unit, respectively. This indicates the initial start-stop state of the i-th generator unit. Let m represent the initial active power output of the i-th generator unit; the physical characteristics of edge m are represented as follows: ,in and Let represent the maximum forward and negative transmission capacities of the m-th transmission line, respectively. and Let represent the reactance and susceptance of the m-th transmission line, respectively; the adjacency matrix A is defined as follows: if node i is connected to node j, then ,otherwise .

4. The method according to claim 1, characterized in that, The spatiotemporal graph convolution module contains two spatiotemporal convolutional blocks and a first fully connected layer, which are used to capture the dynamic patterns and complex dependencies that evolve over time between different nodes. Each spatiotemporal convolutional block adopts an architecture with a spatial graph convolutional layer in the middle and a temporally gated convolutional layer connected in series before and after it. The time-gated convolutional layer is used to extract the temporal features of the bus net load, consisting of a width of K. t It consists of a one-dimensional causal convolution kernel and gated linear units, and performs temporal convolution operation on the load time series of each node in the graph: Where D represents the node feature tensor input in the spatiotemporal graph; W and V are the tensors of size . The convolution kernel, K t Where is the width of the time-gated convolution kernel, and T is the time step. Output channel dimension; Indicates the convolution operation; This indicates element-wise multiplication; This represents the sigmoid gate function; The spatial graph convolutional layer is used to capture the dynamic impact of adjacent buses on the current bus load. It is composed of a Chebyshev spectral convolutional network and aggregates spatial information for each node in the graph. ;in, This is the node embedding matrix output by the spatial graph convolutional layer; The size of the convolution kernel; This is the k-th weight matrix; This is the k-th order Chebyshev polynomial obtained by the recursive formula; The recursive calculation formula is: ;in, , The node embedding matrix at time t is the output of the time-gated convolutional layer; The scaled and normalized Laplace matrix is ​​defined as follows: Where L is the original Laplace matrix, A is the adjacency matrix of the graph; L is the largest eigenvalue of the Laplacian matrix L; I is the identity matrix. When using a first-order approximation, the spatial graph convolutional layer simplifies to: ;in, This represents the normalized adjacency matrix; and These are learnable weight parameters.

5. The method according to claim 1, characterized in that, The spatial graph convolution module includes two edge-conditional convolutional layers and a second fully connected layer, which are used to identify and integrate spatial dependencies and interpret the interaction mechanism between nodes based on the features of the connected edges. The edge-conditional convolutional layer is used to process graphs with edge features, propagating and aggregating generator set features based on transmission line features; In this layer, node features Update to features using the following operations. : ;in, This represents the input feature vector of node i; This represents the updated feature vector of node i; This is the weight matrix; Represents the set of neighboring nodes of node i; This represents the input feature vector of neighbor node j; This represents a multilayer perceptron neural network used for propagating edge features. Its learnable parameters; This represents the feature vector of the edge connecting node i and node j.

6. The method according to claim 1, characterized in that, The cross-attention module adopts a bidirectional cross-attention mechanism, which realizes the interactive fusion between features through attention calculations in two directions: from the spatiotemporal graph to the spatial graph and from the spatial graph to the spatiotemporal graph. The calculation process of the cross-attention module is as follows: ; ; ; ; ; ; where the subscripts st and s represent the spatiotemporal graph and the spatial graph, respectively; The output features of the spatiotemporal graph convolution module; The output features of the spatial graph convolution module; norm(⋅) represents the normalization operation; , , These are the weight parameters for the query matrix, key matrix, and value matrix, respectively. , , These are the corresponding bias parameters; Qs and Qs are query matrices generated from spatiotemporal graph features and spatial graph features, respectively; and These are the key matrices generated from spatiotemporal graph features and spatial graph features, respectively; and These are value matrices generated from spatiotemporal graph features and spatial graph features, respectively. Calculate the attention output in both directions: ; ;in, Attention output from the spatiotemporal map to the spatial map; Attention output in the direction from the spatial map to the spatiotemporal map; It is a normalized exponential function; This is a scaling factor used to prevent the inner product from becoming too large; The attention outputs in both directions are residually connected to the original input features respectively: ; ;in, This is the output from the spacetime map to the spatial map. This is the output from the spatial map to the spatiotemporal map. Will and The features are concatenated along the feature dimension to form a fused feature vector, and then the two features are fused through a third fully connected layer.

7. The method according to claim 1, characterized in that, The variable mapping module consists of a Chebyshev convolutional layer and a fourth fully connected layer; The Chebyshev convolutional layer is used to propagate information from the fused features of the nodes in the graph to obtain aggregated features; The fourth fully connected layer transforms the aggregated features to generate predicted probabilities of unit start-up and shutdown states.

8. The method according to claim 1, characterized in that, The unit start-up and shutdown status prediction step also includes training the prediction model: Construct a loss function, treat the prediction of unit start-up and shutdown status as a binary classification task, minimize the difference between the model output and the true label, so that the unit start-up and shutdown status predicted by the model is as close as possible to the unit start-up and shutdown plan in the optimal solution. The specific form of the loss function is as follows: ;in, The value of the loss function. This represents all learnable parameters of the prediction model; N is the total number of instances in the training dataset; n represents the number of binary variables in a single instance, i.e., the total number of generator start-stop state variables to be predicted; i is the instance index; d is the variable index. This represents the probability that the model predicts the value of the d-th variable to be 1 in the i-th instance; represents the true label of the d-th variable in the i-th instance, and the true label is derived from the unit start-up and shutdown plan in the optimal solution of the historical scheduling cycle unit combination problem; The loss function represents the natural logarithm function; the prediction model is trained based on the training dataset and the loss function, and the loss function value is minimized through an optimization algorithm. The training of the unit start-up and shutdown state prediction model integrating physical information graph network and cross attention was completed, and the model structure and parameters were saved to obtain the trained unit start-up and shutdown state prediction model integrating physical information graph network and cross attention.

9. The method according to claim 1, characterized in that, The variable fixing mechanism for the confidence threshold and the total number of fixed variables includes: The power grid physical information graph of the target scheduling cycle is input into the trained fusion physical information graph network and cross-attention generator start-up and shutdown state prediction model to obtain the prediction probability of the start-up and shutdown state of each generator unit. Set the first confidence threshold. Second confidence threshold and the total number of the first fixed variables The second total number of fixed variables ; Predicted probability of generator start-up and shutdown states Where i is the generator set index and t is the time period index: when When, variable I i,t The value is fixed at 1; when When, variable I i,t The value is fixed at 0; If satisfied If the number of variables is greater than the total number of the first fixed variables k1, then only the values ​​of the first k1 variables with the highest predicted probabilities are fixed to 1; if the following conditions are met... If the number of variables is greater than the total number of the second fixed variables k0, then only the values ​​of the first k0 variables with the lowest prediction probabilities are fixed to 0; Where k1 and k0 are the parameters for the total number of fixed variables; The model outputs the predicted probability of the start-up / shutdown state of the i-th generator unit in time period t; I i,t Let be the start / stop state variable of the i-th generator unit during time period t.

10. A unit combination prediction system integrating physical graph networks and cross-attention, characterized in that, include: The unit combination data acquisition module is used to acquire system data and unit data of the power system within the target scheduling cycle and historical scheduling cycle; Based on the system data and unit data, a unit combination problem model and power grid physical information map data for the corresponding scheduling cycle are constructed; using data from historical scheduling cycles, with the power grid physical information map data as input features and the unit start-up and shutdown plans as output labels, an artificial intelligence training dataset is constructed. The generator start-up and shutdown state prediction module is used to construct a prediction model that integrates a physical information graph network and cross-attention. The prediction model includes a spatiotemporal graph convolution module, a spatial graph convolution module, a cross-attention module, and a variable mapping module. Spatiotemporal graph data from the power grid physical information graph is input to the spatiotemporal graph convolution module, and spatial graph data from the power grid physical information graph is input to the spatial graph convolution module. The outputs of the spatiotemporal graph convolution module and the spatial graph convolution module are input to the cross-attention module for feature interaction fusion. The output of the cross-attention module is input to the variable mapping module to generate the predicted probability of the generator start-up and shutdown states. The unit start-stop state fixing module is used to fix some generator start-stop state variables in the unit combination problem model of the target scheduling cycle to 0 or 1 based on the predicted probability and according to the variable fixing mechanism of confidence threshold and total number of fixed variables, so as to transform the original-scale unit combination problem model into a simplified-scale unit combination problem model, thereby accelerating the solver's solution.

Images