Unit Commitment Scheduling Method and Related Device Based on Constrained Space-Time Graph Neural Network

Abstract

A unit combination scheduling method and related device based on a constrained spatiotemporal graph neural network includes collecting bus load data and power system load data from the power grid at various times of the day; constructing a unit combination problem on a daily basis and solving it under constraints to obtain the corresponding optimal unit power; using a sliding window to divide the bus load and the corresponding optimal unit power by time period; training a pre-established spatiotemporal graph neural network model based on an expanded training data set using the divided bus load as input data and the corresponding optimal unit power as a data label; dividing the bus load data to be predicted by time period and inputting the trained spatiotemporal graph neural network model to obtain a unit combination scheduling solution. The present invention improves the accuracy of solving unit combination solutions based on data-driven methods, reduces the time and computational complexity of calling the solver, constructs a linear constraint layer and post-processing method, and improves the reliability of the prediction results.

Description

Technical Field

[0001] The present invention belongs to the field of power system dispatching, and in particular relates to a unit combination dispatching method based on a constrained spatiotemporal graph neural network and related devices. Background Art

[0002] Unit commitment (UC) is a key issue in power system scheduling for power generation planning. This involves scheduling the start and stop times and output of units based on load forecasts, day ahead, to minimize power system operating costs. To ensure safe power system operation, unit commitment must also consider safety constraints, such as generator ramp rates, output limits, and load balancing constraints at each moment.

[0003] Currently, mainstream algorithms for solving UC problems include heuristic methods, mixed-integer linear programming, Lagrangian relaxation, machine learning, and deep learning. Heuristic methods simulate evolutionary or physical phenomena. They are computationally lightweight and relatively simple to implement, but they cannot guarantee global optimality. Mixed-integer linear programming is currently the mainstream algorithm for solving UC problems. Commonly used solutions include branch-and-bound and cutting plane methods. Mixed-integer linear programming offers high accuracy and is the foundation of commercial solvers such as Gurobi and Cplex. However, it also suffers from long solution times and high computational resource consumption. Lagrangian relaxation uses a dual ascent method to iteratively solve the original problem and update the dual variables, obtaining a set of feasible solutions by reducing the dual gap. However, optimality is also not guaranteed. In recent years, data-driven algorithms, such as machine learning and deep learning, have also been widely used for UC problems. A mainstream approach uses machine learning algorithms to predict unit startups and shutdowns and subsequently calculate unit power. This method significantly improves solution efficiency while obtaining a high-quality initial solution. However, existing data-driven algorithms break down the connections between nodes and are unable to directly map loads to approximate solutions. Data-driven algorithms typically focus on the data itself, not the physical connections between grid nodes. This makes it difficult to learn the complex mapping between bus loads and unit outputs. Furthermore, most existing data-driven algorithms for unit commitment problems require solvers to achieve economic dispatch, which remains time-consuming. Summary of the Invention

[0004] The purpose of the present invention is to address the problems in the above-mentioned prior art and provide a unit combination scheduling method and related devices based on a constrained spatiotemporal graph neural network, improve the accuracy of solving unit combination schemes based on data-driven, reduce the time and computational complexity of calling the solver, construct a linear constraint layer and post-processing method, and improve the reliability of the prediction results of the spatiotemporal graph neural network.

[0005] In order to achieve the above object, the present invention has the following technical solutions:

[0006] In a first aspect, an embodiment of the present invention provides a unit commitment scheduling method based on a constrained spatiotemporal graph neural network, comprising:

[0007] Collect bus load data and power system load data at all times of the day to establish a training data set;

[0008] Construct the unit combination problem on a daily basis and solve it under the constraints to obtain the corresponding optimal unit power;

[0009] Use a sliding window to divide the bus load and the corresponding optimal unit power into time periods to expand the training data set;

[0010] Based on the expanded training data set, the divided bus load is used as input data and the corresponding optimal unit power is used as the data label to train a pre-established spatiotemporal graph neural network model. The spatiotemporal graph neural network model adds a linear constraint layer to the end of the spatiotemporal graph convolution, uses the linear constraint layer to linearly constrain the output of the fully connected layer, and projects the predicted power of the neural network into a feasible region that meets the specified linear constraints.

[0011] The bus load data that needs to be predicted is divided into time periods and input into the trained space-time graph neural network model. The predicted power output by the space-time graph neural network model is restricted, including filtering the prediction results that are less than the minimum power and adjusting the prediction results that do not meet the start-stop requirements, to ensure the availability of the predicted power and obtain the unit combination scheduling plan.

[0012] As a preferred solution, the step of constructing the unit commitment problem on a daily basis includes:

[0013] Assume there are N nodes, N B busbar nodes, N G The total number of decision-making periods per day in the power system is N. T ; The objective function of the unit commitment problem is constructed as follows:

[0014]

[0015] In the formula, g, t represent the sequence number of the unit and time period respectively; N G and N T Represent the total number of units and time periods respectively; p g,t is the output of unit g in time period t; a g , b g , c gare the coefficients of the quadratic term, linear term and constant term of the operating cost function of unit g; SU g is the startup cost of unit g; y g,t It is an integer variable indicating whether the unit g has started in time period t. A value of 1 indicates that the unit g has started in time period t, and a value of 0 indicates that the unit g has not started.

[0016] As a preferred solution, in the step of solving the corresponding optimal unit power under the constraint conditions, the constraint conditions include:

[0017] Power balance constraints:

[0018]

[0019] Where b is the node number; N B is the total number of busbar nodes; D b,t is the load of node b in time period t;

[0020] Upper and lower limit constraints of unit output:

[0021]

[0022] Where, and are the minimum and maximum output of unit g respectively; u g,t is an integer variable indicating the on / off status of unit g in time period t. A value of 1 indicates that unit g is on in time period t, and a value of 0 indicates that unit g is off in time period t.

[0023] Unit climbing constraints:

[0024] -RD g ≤p g,t -p g,t-1 ≤RU g ,g=1,…,N G ,T=2,…,N T

[0025] Where, RD g and RU g They are the ramp rate limits for the unit’s g output decrease and increase, respectively;

[0026] Minimum start and stop time constraints:

[0027]

[0028] y g,t +z g,t ≤1,g=1,…,N G ,t=1,…,N T

[0029] Where TU g and TD g are the minimum start and stop time of unit g; y g,t , z g,t is an integer variable indicating whether the unit g has started or stopped in time period t. A value of 1 indicates that the unit g has started or stopped in time period t, and a value of 0 indicates that the unit g has not started or stopped.

[0030] Base-state branch active power flow security constraints:

[0031]

[0032] Where, l is the branch number; N L is the total number of branches; is the active power flow of branch l in time period t under the base state; T is the power transfer distribution factor; g∈G(b) is the set of units at node b; F l max is the long-term allowable current-carrying power of branch l.

[0033] As a preferred solution, the steps of training a pre-established spatiotemporal graph neural network model based on the expanded training data set, using the divided bus load as input data and the corresponding optimal unit power as a data label include:

[0034] Based on the expanded training dataset, the graph neural connection relationship is obtained from the real power grid topology;

[0035] The load D of all nodes in the power grid within a period of time F is input into the spatiotemporal graph convolutional network without a linear constraint layer. The graph convolution layer of the spatiotemporal graph convolutional network extracts the input spatial features. The graph convolution layer extracts and aggregates the load features of each node, learns the spatial features between the data, and expands the time series length of the data.

[0036] The graph data processed by the graph convolution layer is converted into time series data, and then causal dilation time convolution is used to extract the time series features of each node. Subsequently, a fully connected neural network layer is used to aggregate the time series features between different nodes, and the data format is converted to output the predicted power P of the thermal power unit within a period of time F. The neural network repeatedly executes the above steps by updating the loss function between the predicted power and the actual unit power scheduling data to complete the training of the spatiotemporal graph neural network model.

[0037] As a preferred solution, the step of projecting the predicted power of the neural network into a feasible region that meets the specified linear constraints includes:

[0038] make Represents the input value of the constrained activation layer, that is, the predicted power value of the spatiotemporal graph neural network model; let Represents the output of the constrained activation layer, that is, the power value that meets the linear constraint; where N G Represents the number of units, F represents the time series length; Construct a convex feasible set for all linear constraints in the unit combination problem The convex feasible set Ω is the intersection of a set of constraints in the form of m inequality constraints, and the mathematical expression is as follows:

[0039] Ω={x|h i (x)≤0,i=1,...,m}

[0040] Ensure that each constraint h i (x)≤0 is convex, that is, it conforms to the following mathematical expression:

[0041]

[0042] Determine a fixed point p in the convex feasible set Ω, that is, p∈Ω, then any point x in the set Ω is expressed as x=p+α·r, and calculate the scaling factor α≥0; is a vector from point p; for any p and r, the scaling factor α has an upper limit The mathematical expression is as follows:

[0043]

[0044] At this time, the line segment Belongs to the set Ω, upper limit The meaning of is the intersection of ray r and a constraint;

[0045] For the point z outside the constrained feasible region in the spatiotemporal graph neural network, by specifying the ray r(z) = zp, the scaling factor The calculation is performed through the following mathematical expression, that is, projecting the point z outside the feasible domain along the ray from z to p to the point g inside the feasible domain:

[0046]

[0047] As a preferred solution, the steps of limiting the predicted power output by the spatiotemporal graph neural network model include filtering the predicted results that are less than the minimum power and adjusting the predicted results that do not meet the start-stop requirements, and ensuring the availability of the predicted power include:

[0048] Set a threshold δ, first check and adjust the output power of the spatiotemporal graph neural network model for each unit, set the prediction result whose difference between the minimum power and the prediction result is less than the threshold δ as the minimum power of the corresponding unit, and set the prediction result whose difference between the minimum power and the prediction result is greater than the threshold δ to 0;

[0049] Check and adjust the minimum start and stop time of each unit in the prediction results;

[0050] The prediction results are re-passed through the linear constraint activation layer with branch active power flow security constraints added to ensure that the output power meets the branch active power flow security constraints.

[0051] As a preferred solution, the step of checking and adjusting the minimum start and stop time of each unit in the prediction result includes:

[0052] Record the duration of each start and stop of each unit;

[0053] Check the duration of each start and stop of each unit. For any start or stop that violates the minimum start or stop time constraint, compare the start or stop time with the previous and subsequent stop or start times in chronological order.

[0054] When the start and stop time of a constraint violation is shorter than the start and stop time on both sides, adjust the corresponding start and stop.

[0055] In a second aspect, a unit commitment scheduling system based on a constrained spatiotemporal graph neural network is provided, comprising:

[0056] The data collection module is used to collect bus load data and power system load data of the power grid at various times of the day to establish a training data set;

[0057] The optimal unit power solution module is used to construct the unit combination problem on a daily basis and solve the corresponding optimal unit power under the constraints;

[0058] The training data set expansion module is used to use a sliding window to divide the bus load and the corresponding optimal unit power according to time periods to expand the training data set;

[0059] A spatiotemporal graph neural network model training module is used to train a pre-established spatiotemporal graph neural network model based on the expanded training data set, using the divided bus load as input data and the corresponding optimal unit power as the data label. The spatiotemporal graph neural network model adds a linear constraint layer to the end of the spatiotemporal graph convolution, uses the linear constraint layer to linearly constrain the output of the fully connected layer, and projects the predicted power of the neural network into a feasible region that meets the specified linear constraints.

[0060] The predicted power post-processing output module is used to divide the bus load data that needs to be predicted according to time periods, input the trained space-time graph neural network model, and limit the predicted power output by the space-time graph neural network model, including filtering the prediction results that are less than the minimum power and adjusting the prediction results that do not meet the start-stop requirements, to ensure the availability of the predicted power and obtain the unit combination scheduling plan.

[0061] As a preferred solution, the steps of constructing the unit commitment problem on a daily basis in the optimal unit power solving module include:

[0062] Assume there are N nodes, N B busbar nodes, N G The total number of decision-making periods per day in the power system is N. T ; The objective function of the unit commitment problem is constructed as follows:

[0063]

[0064] In the formula, g, t represent the sequence number of the unit and time period respectively; N G and N T Represent the total number of units and time periods respectively; p g,t is the output of unit g in time period t; a g , b g , c g are the coefficients of the quadratic term, linear term and constant term of the operating cost function of unit g; SU g is the startup cost of unit g; y g,t It is an integer variable indicating whether the unit g has started in time period t. A value of 1 indicates that the unit g has started in time period t, and a value of 0 indicates that the unit g has not started.

[0065] As a preferred solution, when the optimal unit power solving module solves the corresponding optimal unit power under constraints, the constraints include:

[0066] Power balance constraints:

[0067]

[0068] Where b is the node number; N B is the total number of busbar nodes; D b,t is the load of node b in time period t;

[0069] Upper and lower limit constraints of unit output:

[0070]

[0071] Where, and are the minimum and maximum output of unit g respectively; u g,t is an integer variable indicating the on / off status of unit g in time period t. A value of 1 indicates that unit g is on in time period t, and a value of 0 indicates that unit g is off in time period t.

[0072] Unit climbing constraints:

[0073] -RD g ≤p g,t -p g,t-1 ≤RU g ,g=1,…,N G ,T=2,…,N T

[0074] Where, RD g and RU g They are the ramp rate limits for the unit’s g output decrease and increase, respectively;

[0075] Minimum start and stop time constraints:

[0076]

[0077]

[0078] y g,t +z g,t ≤1,g=1,…,N G ,t=1,…,N T

[0079] Where TU g and TD g are the minimum start and stop time of unit g; y g,t , z g,t is an integer variable indicating whether the unit g has started or stopped in time period t. A value of 1 indicates that the unit g has started or stopped in time period t, and a value of 0 indicates that the unit g has not started or stopped.

[0080] Base-state branch active power flow security constraints:

[0081]

[0082] Where, l is the branch number; N L is the total number of branches; is the active power flow of branch l in time period t under the base state; T is the power transfer distribution factor; g∈G(b) is the set of units at node b; F l max is the long-term allowable current-carrying power of branch l.

[0083] As a preferred solution, the spatiotemporal graph neural network model training module uses the divided bus load as input data and the corresponding optimal unit power as a data label. The steps of training the pre-established spatiotemporal graph neural network model include:

[0084] Based on the expanded training dataset, the graph neural connection relationship is obtained from the real power grid topology;

[0085] The load D of all nodes in the power grid within a period of time F is input into the spatiotemporal graph convolutional network without a linear constraint layer. The graph convolution layer of the spatiotemporal graph convolutional network extracts the input spatial features. The graph convolution layer extracts and aggregates the load features of each node, learns the spatial features between the data, and expands the time series length of the data.

[0086] The graph data processed by the graph convolution layer is converted into time series data, and then causal dilation time convolution is used to extract the time series features of each node. Subsequently, a fully connected neural network layer is used to aggregate the time series features between different nodes, and the data format is converted to output the predicted power P of the thermal power unit within a period of time F. The neural network repeatedly executes the above steps by updating the loss function between the predicted power and the actual unit power scheduling data to complete the training of the spatiotemporal graph neural network model.

[0087] As a preferred solution, the step of the spatiotemporal graph neural network model training module projecting the predicted power of the neural network into a feasible domain that meets the specified linear constraints includes:

[0088] make Represents the input value of the constrained activation layer, that is, the predicted power value of the spatiotemporal graph neural network model; let Represents the output of the constrained activation layer, that is, the power value that meets the linear constraint; where N G Represents the number of units, F represents the time series length; Construct a convex feasible set for all linear constraints in the unit combination problem The convex feasible set Ω is the intersection of a set of constraints in the form of m inequality constraints, and the mathematical expression is as follows:

[0089] Ω={x|h i (x)≤0,i=1,...,m}

[0090] Ensure that each constraint h i (x)≤0 is convex, that is, it conforms to the following mathematical expression:

[0091]

[0092] Determine a fixed point p in the convex feasible set Ω, that is, p∈Ω, then any point x in the set Ω is expressed as x=p+α·r, and calculate the scaling factor α≥0; is a vector from point p; for any p and r, the scaling factor α has an upper limit The mathematical expression is as follows:

[0093]

[0094] At this time, the line segment Belongs to the set Ω, upper limit The meaning of is the intersection of ray r and a constraint;

[0095] For the point z outside the constrained feasible region in the spatiotemporal graph neural network, by specifying the ray r(z) = zp, the scaling factor The calculation is performed through the following mathematical expression, that is, projecting the point z outside the feasible domain along the ray from z to p to the point g inside the feasible domain:

[0096]

[0097] As a preferred solution, the predicted power post-processing output module limits the predicted power output by the spatiotemporal graph neural network model, including filtering the predicted results that are less than the minimum power and adjusting the predicted results that do not meet the start-stop requirements. When ensuring the availability of the predicted power, a threshold δ is set, and the output power of the spatiotemporal graph neural network model is first checked and adjusted for each unit. The predicted result whose difference between the minimum power and the predicted result is less than the threshold δ is set as the minimum power of the corresponding unit, and the predicted result whose difference between the minimum power and the predicted result is greater than the threshold δ is set to 0;

[0098] Check and adjust the minimum start and stop time of each unit in the prediction results;

[0099] The prediction results are re-passed through the linear constraint activation layer with branch active power flow security constraints added to ensure that the output power meets the branch active power flow security constraints.

[0100] As a preferred solution, the step of the predicted power post-processing output module checking and adjusting the minimum start and stop time of each unit in the prediction result includes:

[0101] Record the duration of each start and stop of each unit;

[0102] Check the duration of each start and stop of each unit. For any start or stop that violates the minimum start or stop time constraint, compare the start or stop time with the previous and subsequent stop or start times in chronological order.

[0103] When the start and stop time of a constraint violation is shorter than the start and stop time on both sides, adjust the corresponding start and stop.

[0104] In a third aspect, an embodiment of the present invention provides an electronic device comprising a processor and a memory, wherein the processor is configured to execute a computer program stored in the memory to implement the unit combination scheduling method based on a constrained spatiotemporal graph neural network.

[0105] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by a processor, the unit combination scheduling method based on the constrained spatiotemporal graph neural network is implemented.

[0106] Compared with the prior art, the first aspect of the present invention has at least the following beneficial effects:

[0107] The unit combination scheduling method based on the constrained spatiotemporal graph neural network of the present invention collects bus load data and power system load data of the power grid at each time of the day, divides the bus load and the corresponding optimal unit power according to time periods, uses the divided bus load as input data, and uses the corresponding optimal unit power as a data label to train a pre-established spatiotemporal graph neural network model. The spatiotemporal graph neural network model is a framework that combines graph neural networks and temporal convolutional networks. It can simultaneously learn the physical topology of the power grid and the temporal information of historical data, and directly use the bus load to solve the unit power scheduling. In addition, in response to the problem that the data-driven methods currently applied to power systems do not have hard output constraints, the spatiotemporal graph neural network model of the present invention adds a linear constraint layer to the end of the spatiotemporal graph convolution, uses the linear constraint layer to linearly constrain the output result of the fully connected layer, and projects the predicted power of the neural network into a feasible domain that meets the specified linear constraints. In addition, the present invention designs a post-processing process, which limits the predicted power output by the spatiotemporal graph neural network model, including filtering the prediction results that are less than the minimum power and adjusting the prediction results that do not meet the start-stop requirements, ensuring the availability of the predicted power and improving the reliability of the spatiotemporal graph neural network prediction results. The present invention uses realistic node connection relationships and edge weights to convert the physical power grid topology into graph neural modeling. The graph neural network uses the physical power grid model to learn network topology information during the training process, achieving higher prediction accuracy with fewer samples. Many existing studies use neural network methods for auxiliary decision-making of unit commitment problems, but the present invention constructs a mapping relationship between known load inputs and generator unit power decisions, and uses the bus load of each node at each moment of the power grid to directly predict the output of the generator unit, which simplifies the calculation process and improves computational efficiency. The present invention combines a temporal convolutional network with a graph neural network, and simultaneously extracts the spatiotemporal characteristics in the data, thereby improving prediction accuracy and model robustness. Traditional data-driven algorithms cannot guarantee any constraints during the training process, but the present invention combines data-driven and model-driven methods, and introduces a linear constraint layer in the spatiotemporal graph neural network model, so that the neural network can add linear constraints to the results. The addition of the linear constraint layer simplifies the constraint verification and post-processing of the prediction results, and helps to find feasible solutions that meet the constraints. Combining the above characteristics, the unit combination scheduling method based on the constrained spatiotemporal graph neural network of the present invention can quickly and accurately solve the unit combination problem, and automatically check and adjust the feasibility and accuracy of the solution results. The method of the present invention not only significantly reduces the solution time of power system scheduling, but also greatly improves the level of intelligence in the field of power system scheduling.

[0108] It can be understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0110] Figure 1 A schematic diagram of the spatiotemporal graph neural network model architecture established for an embodiment of the present invention;

[0111] Figure 2 This is a flow chart of a unit commitment scheduling method based on a constrained spatiotemporal graph neural network according to an embodiment of the present invention;

[0112] Figure 3 Schematic diagram of the working principle of the linear constraint layer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0113] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0114] In order to get rid of the problem that the data-driven unit combination solution algorithm is not accurate enough and that calling the solver is time-consuming and computationally intensive, an embodiment of the present invention proposes a unit combination scheduling method based on a constrained spatiotemporal graph neural network, and constructs a framework that combines graph neural networks and temporal convolutional networks, namely, a spatiotemporal graph convolutional network. The spatiotemporal graph neural network model of the present invention can simultaneously learn the physical topology of the power grid and the temporal information of historical data, and directly use the bus load to solve the unit power scheduling. In addition, the data-driven methods currently applied to power systems do not have output hard constraints. The present invention innovatively constructs a linear constraint layer and designs a post-processing method to improve the reliability of the prediction results of the spatiotemporal graph neural network.

[0115] See also Figure 2 The unit commitment scheduling method based on a constrained spatiotemporal graph neural network in an embodiment of the present invention mainly includes:

[0116] S1. Collect bus load data and power system load data of the power grid at various times of the day to establish a training data set;

[0117] S2. Construct the unit combination problem on a daily basis and solve it under the constraints to obtain the corresponding optimal unit power;

[0118] S3. Use a sliding window to divide the bus load and the corresponding optimal unit power according to time periods to expand the training data set;

[0119] S4. Based on the expanded training data set, the divided bus load is used as input data and the corresponding optimal unit power is used as the data label to train a pre-established spatiotemporal graph neural network model; the spatiotemporal graph neural network model adds a linear constraint layer to the end of the spatiotemporal graph convolution, uses the linear constraint layer to linearly constrain the output result of the fully connected layer, and projects the predicted power of the neural network into a feasible region that meets the specified linear constraints;

[0120] S5. Divide the bus load data that needs to be predicted according to time periods and input the trained space-time graph neural network model. Limit the predicted power output by the space-time graph neural network model, including filtering the prediction results that are less than the minimum power and adjusting the prediction results that do not meet the start-stop requirements, to ensure the availability of the predicted power and obtain the unit combination scheduling plan.

[0121] The embodiment of the present invention takes the IEEE30-BUS system and the IEEE118-BUS system as examples to verify the effectiveness of the method.

[0122] In a possible implementation, step S1 collects bus load data and power system load data of the power grid at various times of the day, formats the data, generates training samples, and establishes a training data set.

[0123] In a possible implementation, the step of constructing the unit commitment problem on a daily basis in step S2 includes:

[0124] Assume there are N nodes, N B busbar nodes, N G The total number of decision-making periods per day in the power system is N. T ; The objective function of the unit commitment problem is constructed as follows:

[0125]

[0126] In the formula, g, t represent the sequence number of the unit and time period respectively; N G and N T Represent the total number of units and time periods respectively; p g,t is the output of unit g in time period t; a g , b g , c g are the coefficients of the quadratic term, linear term and constant term of the operating cost function of unit g; SU g is the startup cost of unit g; y g,tIt is an integer variable indicating whether the unit g has started in time period t. A value of 1 indicates that the unit g has started in time period t, and a value of 0 indicates that the unit g has not started.

[0127] Furthermore, in the step of solving the corresponding optimal unit power under the constraint conditions, the constraint conditions include:

[0128] Power balance constraints:

[0129]

[0130] Where b is the node number; N B is the total number of busbar nodes; D b,t is the load of node b in time period t;

[0131] Upper and lower limit constraints of unit output:

[0132]

[0133] Where, and are the minimum and maximum output of unit g respectively; u g,t is an integer variable indicating the on / off status of unit g in time period t. A value of 1 indicates that unit g is on in time period t, and a value of 0 indicates that unit g is off in time period t.

[0134] Unit climbing constraints:

[0135] -RD g ≤p g,t -p g,t-1 ≤RU g ,g=1,…,N G ,T=2,…,N T

[0136] Where, RD g and RU g They are the ramp rate limits for the unit’s g output decrease and increase, respectively;

[0137] Minimum start and stop time constraints:

[0138]

[0139] y g,t +z g,t ≤1,g=1,…,N G ,t=1,…,N T

[0140] Where TU g and TD g are the minimum start and stop time of unit g; yg,t , z g,t is an integer variable indicating whether the unit g has started or stopped in time period t. A value of 1 indicates that the unit g has started or stopped in time period t, and a value of 0 indicates that the unit g has not started or stopped.

[0141] Base-state branch active power flow security constraints:

[0142]

[0143] Where, l is the branch number; N L is the total number of branches; is the active power flow of branch l in time period t under the base state; T is the power transfer distribution factor; g∈G(b) is the set of units at node b; F l max is the long-term allowable current-carrying power of branch l.

[0144] Furthermore, in a possible implementation manner, step S2 of the present invention calls the Gurobi solver to solve the unit commitment problem constructed every day and record the optimal unit power scheduling.

[0145] In one possible implementation, the present invention uses a sliding window to construct a new data set in step S3. Specifically, the width of the sliding window is defined as F, and the step length is 1. Then, the present invention starts from the first moment of each day and moves one time period at a time until the end of the data set is reached. For each time period t (from 1 to T-F+1, with a step length of s), the present invention takes the load of the F time periods within the window [D b,t ,D b,t+1 ,...,D b,t+F-1 ] and unit output [P g,t ,P g,t+1 ,...,P g,t+F-1 ] as a new sample. The sliding window length can be 6, 12, or 24.

[0146] like Figure 1 As shown, in a possible embodiment, step S4 of the present invention uses the divided bus load as input data and the corresponding optimal unit power as a data label, and the steps of training a pre-established spatiotemporal graph neural network model include:

[0147] 1) Based on the expanded training dataset, the graph neural connectivity is obtained from the real power grid topology;

[0148] 2) The load D of all nodes in the power grid over a period of time F is input into a spatiotemporal graph convolutional network without a linear constraint layer. The graph convolutional layer of the spatiotemporal graph convolutional network extracts the input spatial features. The graph convolutional layer extracts and aggregates the load features of each node, learns the spatial features between the data, and expands the time series length of the data.

[0149] 3) The graph data processed by the graph convolution layer is converted into time series data. Causal dilation time convolution is used to extract time series features for each node. Subsequently, a fully connected neural network layer is used to aggregate the time series features between different nodes. The data format is converted and output as the predicted power P of the thermal power unit within a period of time F. The neural network completes the training of the spatiotemporal graph neural network model by repeatedly performing steps 2) and 3) by updating the loss function between the predicted power and the actual unit power scheduling data.

[0150] Further, such as Figure 3 As shown, in a possible implementation, step S4 of the embodiment of the present invention adds a linear constraint layer to the end of the spatiotemporal graph convolution, uses the linear constraint layer to linearly constrain the output result of the fully connected layer, and projects the predicted power of the neural network into a feasible region that meets the specified linear constraints. The steps include:

[0151] make Represents the input value of the constrained activation layer, that is, the predicted power value of the spatiotemporal graph neural network model; let Represents the output of the constrained activation layer, that is, the power value that meets the linear constraint; where N G Represents the number of units, F represents the time series length; Construct a convex feasible set for all linear constraints in the unit combination problem The convex feasible set Ω is the intersection of a set of constraints in the form of m inequality constraints, and the mathematical expression is as follows:

[0152] Ω={x|h i (x)≤0,i=1,...,m}

[0153] Ensure that each constraint h i (x)≤0 is convex, that is, it conforms to the following mathematical expression:

[0154]

[0155] Determine a fixed point p in the convex feasible set Ω, that is, p∈Ω, then any point x in the set Ω is expressed as x=p+α·r, and calculate the scaling factor α≥0; is a vector from point p; for any p and r, the scaling factor α has an upper limit The mathematical expression is as follows:

[0156]

[0157] At this time, the line segment Belongs to the set Ω, upper limit The meaning of is the intersection of ray r and a constraint;

[0158] For the point z outside the constrained feasible region in the spatiotemporal graph neural network, by specifying the ray r(z) = zp, the scaling factor The calculation is performed through the following mathematical expression, that is, projecting the point z outside the feasible domain along the ray from z to p to the point g inside the feasible domain:

[0159]

[0160] In one possible implementation, step S5 limits the predicted power output by the spatiotemporal graph neural network model, including filtering the predicted results that are less than the minimum power and adjusting the predicted results that do not meet the start-stop requirements. The steps of ensuring the availability of the predicted power include:

[0161] Set a threshold δ, first check and adjust the output power of the spatiotemporal graph neural network model for each unit, set the prediction result whose difference between the minimum power and the prediction result is less than the threshold δ as the minimum power of the corresponding unit, and set the prediction result whose difference between the minimum power and the prediction result is greater than the threshold δ to 0;

[0162] Check and adjust the minimum start and stop time of each unit in the prediction results;

[0163] The prediction results are re-passed through the linear constraint activation layer with branch active power flow security constraints added to ensure that the output power meets the branch active power flow security constraints.

[0164] Furthermore, the step of checking and adjusting the minimum start and stop time of each unit in the prediction result includes:

[0165] Record the duration of each start and stop of each unit;

[0166] Check the duration of each start and stop of each unit. For any start and stop that violates the minimum start and stop time constraints, compare the start (stop) time with the stop (start) time before and after it in chronological order.

[0167] When the start and stop time of a constraint violation is shorter than the start and stop time on both sides, adjust the corresponding start and stop.

[0168] The unit commitment scheduling method based on a constrained spatiotemporal graph neural network in this embodiment of the present invention can quickly and accurately solve unit commitment problems and automatically verify and adjust the feasibility and accuracy of the solution. This invention not only significantly reduces the solution time for power system scheduling, but also greatly improves the level of intelligence in the field of power system scheduling.

[0169] Another embodiment of the present invention further provides a unit commitment scheduling system based on a constrained spatiotemporal graph neural network, comprising:

[0170] The data collection module is used to collect bus load data and power system load data of the power grid at various times of the day to establish a training data set;

[0171] The optimal unit power solution module is used to construct the unit combination problem on a daily basis and solve the corresponding optimal unit power under the constraints;

[0172] The training data set expansion module is used to use a sliding window to divide the bus load and the corresponding optimal unit power according to time periods to expand the training data set;

[0173] A spatiotemporal graph neural network model training module is used to train a pre-established spatiotemporal graph neural network model based on the expanded training data set, using the divided bus load as input data and the corresponding optimal unit power as the data label. The spatiotemporal graph neural network model adds a linear constraint layer to the end of the spatiotemporal graph convolution, uses the linear constraint layer to linearly constrain the output of the fully connected layer, and projects the predicted power of the neural network into a feasible region that meets the specified linear constraints.

[0174] The predicted power post-processing output module is used to divide the bus load data that needs to be predicted according to time periods, input the trained space-time graph neural network model, and limit the predicted power output by the space-time graph neural network model, including filtering the prediction results that are less than the minimum power and adjusting the prediction results that do not meet the start-stop requirements, to ensure the availability of the predicted power and obtain the unit combination scheduling plan.

[0175] In one possible implementation, the optimal unit power solving module constructs the unit commitment problem on a daily basis, including:

[0176] Assume there are N nodes, N B busbar nodes, N G The total number of decision-making periods per day in the power system is N. T ; The objective function of the unit commitment problem is constructed as follows:

[0177]

[0178] In the formula, g, t represent the sequence number of the unit and time period respectively; N G and N T Represent the total number of units and time periods respectively; p g,t is the output of unit g in time period t; a g , b g , cg are the coefficients of the quadratic term, linear term and constant term of the operating cost function of unit g; SU g is the startup cost of unit g; y g,t It is an integer variable indicating whether the unit g has started in time period t. A value of 1 indicates that the unit g has started in time period t, and a value of 0 indicates that the unit g has not started.

[0179] In a possible implementation, when the optimal unit power solving module solves the corresponding optimal unit power under constraints, the constraints include:

[0180] Power balance constraints:

[0181]

[0182] Where b is the node number; N B is the total number of busbar nodes; D b,t is the load of node b in time period t;

[0183] Upper and lower limit constraints of unit output:

[0184]

[0185] Where, and are the minimum and maximum output of unit g respectively; u g,t is an integer variable indicating the on / off status of unit g in time period t. A value of 1 indicates that unit g is on in time period t, and a value of 0 indicates that unit g is off in time period t.

[0186] Unit climbing constraints:

[0187] -RD g ≤p g,t -p g,t-1 ≤RU g ,g=1,…,N G ,T=2,…,N T

[0188] Where, RD g and RU g They are the ramp rate limits for the unit’s g output decrease and increase, respectively;

[0189] Minimum start and stop time constraints:

[0190]

[0191]

[0192] y g,t +z g,t≤1,g=1,…,N G ,t=1,…,N T

[0193] Where TU g and TD g are the minimum start and stop time of unit g; y g,t , z g,t is an integer variable indicating whether the unit g has started or stopped in time period t. A value of 1 indicates that the unit g has started or stopped in time period t, and a value of 0 indicates that the unit g has not started or stopped.

[0194] Base-state branch active power flow security constraints:

[0195]

[0196] Where, l is the branch number; N L is the total number of branches; is the active power flow of branch l in time period t under the base state; T is the power transfer distribution factor; g∈G(b) is the set of units at node b; F l max is the long-term allowable current-carrying power of branch l.

[0197] In one possible implementation, the spatiotemporal graph neural network model training module uses the divided bus load as input data and the corresponding optimal unit power as a data label. The steps of training the spatiotemporal graph neural network model include:

[0198] Based on the expanded training dataset, the graph neural connection relationship is obtained from the real power grid topology;

[0199] The load D of all nodes in the power grid within a period of time F is input into the spatiotemporal graph convolutional network without a linear constraint layer. The graph convolution layer of the spatiotemporal graph convolutional network extracts the input spatial features. The graph convolution layer extracts and aggregates the load features of each node, learns the spatial features between the data, and expands the time series length of the data.

[0200] The graph data processed by the graph convolution layer is converted into time series data, and then causal dilation time convolution is used to extract the time series features of each node. Subsequently, a fully connected neural network layer is used to aggregate the time series features between different nodes, and the data format is converted to output the predicted power P of the thermal power unit within a period of time F. The neural network repeatedly executes the above steps by updating the loss function between the predicted power and the actual unit power scheduling data to complete the training of the spatiotemporal graph neural network model.

[0201] In one possible implementation, the step of the spatiotemporal graph neural network model training module projecting the predicted power of the neural network into a feasible region that meets the specified linear constraints includes:

[0202] make Represents the input value of the constrained activation layer, that is, the predicted power value of the spatiotemporal graph neural network model; let Represents the output of the constrained activation layer, that is, the power value that meets the linear constraint; where N G Represents the number of units, F represents the time series length; Construct a convex feasible set for all linear constraints in the unit combination problem The convex feasible set Ω is the intersection of a set of constraints in the form of m inequality constraints, and the mathematical expression is as follows:

[0203] Ω={x|h i (x)≤0,i=1,...,m}

[0204] Ensure that each constraint h i (x)≤0 is convex, that is, it conforms to the following mathematical expression:

[0205]

[0206] Determine a fixed point p in the convex feasible set Ω, that is, p∈Ω, then any point x in the set Ω is expressed as x=p+α·r, and calculate the scaling factor α≥0; is a vector from point p; for any p and r, the scaling factor α has an upper limit The mathematical expression is as follows:

[0207]

[0208] At this time, the line segment Belongs to the set Ω, upper limit The meaning of is the intersection of ray r and a constraint;

[0209] For the point z outside the constrained feasible region in the spatiotemporal graph neural network, by specifying the ray r(z) = zp, the scaling factor The calculation is performed through the following mathematical expression, that is, projecting the point z outside the feasible domain along the ray from z to p to the point g inside the feasible domain:

[0210]

[0211] In one possible implementation, the predicted power post-processing output module limits the predicted power output by the spatiotemporal graph neural network model, including filtering the predicted results that are less than the minimum power and adjusting the predicted results that do not meet the start-stop requirements. To ensure the availability of the predicted power, a threshold δ is set. First, the output power of the spatiotemporal graph neural network model is checked and adjusted for each unit. The predicted result whose difference between the minimum power and the predicted result is less than the threshold δ is set as the minimum power of the corresponding unit, and the predicted result whose difference between the minimum power and the predicted result is greater than the threshold δ is set to 0.

[0212] Check and adjust the minimum start and stop time of each unit in the prediction results;

[0213] The prediction results are re-passed through the linear constraint activation layer with branch active power flow security constraints added to ensure that the output power meets the branch active power flow security constraints.

[0214] Furthermore, the steps of the predicted power post-processing output module checking and adjusting the minimum start and stop time of each unit in the prediction result include:

[0215] Record the duration of each start and stop of each unit;

[0216] Check the duration of each start and stop of each unit. For any start or stop that violates the minimum start or stop time constraint, compare the start or stop time with the previous and subsequent stop or start times in chronological order.

[0217] When the start and stop time of a constraint violation is shorter than the start and stop time on both sides, adjust the corresponding start and stop.

[0218] Another embodiment of the present invention further proposes an electronic device, including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the unit combination scheduling method based on the constrained spatiotemporal graph neural network.

[0219] Another embodiment of the present invention further proposes a computer-readable storage medium, which stores at least one instruction. When the at least one instruction is executed by a processor, it implements the unit combination scheduling method based on the constrained spatiotemporal graph neural network.

[0220] The computer program includes computer program code, which may be in source code form, object code form, executable file or some intermediate form. The computer-readable storage medium may include: any entity or device, medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electric carrier signal, telecommunication signal and software distribution medium that can carry the computer program code. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals. For ease of explanation, the above content only shows the part related to the embodiment of the present invention. For specific technical details not disclosed, please refer to the method part of the embodiment of the present invention. The computer-readable storage medium is non-transitory and can be stored in a storage device formed by various electronic devices, and can implement the execution process recorded in the method of the embodiment of the present invention.

[0221] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0222] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0223] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0224] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A unit commitment scheduling method based on constrained spatiotemporal graph neural network, characterized in that: include: Collect bus load data and power system load data at all times of the day to establish a training data set; Construct the unit combination problem on a daily basis and solve it under the constraints to obtain the corresponding optimal unit power; Use a sliding window to divide the bus load and the corresponding optimal unit power into time periods to expand the training data set; Based on the expanded training data set, the divided bus load is used as input data and the corresponding optimal unit power is used as the data label to train a pre-established spatiotemporal graph neural network model. The spatiotemporal graph neural network model adds a linear constraint layer to the end of the spatiotemporal graph convolution, uses the linear constraint layer to linearly constrain the output of the fully connected layer, and projects the predicted power of the neural network into a feasible region that meets the specified linear constraints. The bus load data to be predicted is divided into time periods and input into a trained spatiotemporal graph neural network model. The predicted power output by the spatiotemporal graph neural network model is restricted, including filtering prediction results that are less than the minimum power and adjusting prediction results that do not meet the start-stop requirements, to ensure the availability of the predicted power and obtain a unit combination scheduling plan. The steps of training a pre-established spatiotemporal graph neural network model based on the expanded training data set, using the divided bus load as input data and the corresponding optimal unit power as a data label include: Based on the expanded training dataset, the graph neural connection relationship is obtained from the real power grid topology; The load D of all nodes in the power grid within a period of time F is input into the spatiotemporal graph convolutional network without a linear constraint layer. The graph convolution layer of the spatiotemporal graph convolutional network extracts the input spatial features. The graph convolution layer extracts and aggregates the load features of each node, learns the spatial features between the data, and expands the time series length of the data. The graph data processed by the graph convolution layer is converted into time series data. Causal dilation time convolution is then used to extract time series features from each node. A fully connected neural network layer is then used to aggregate the time series features between different nodes. The data format is then converted to output as the predicted power P of the thermal power unit within a period F. The neural network repeatedly performs the above steps by updating the loss function between the predicted power and the actual unit power scheduling data to complete the training of the spatiotemporal graph neural network model. The step of projecting the predicted power of the neural network into a feasible region that meets the specified linear constraints includes: make Represents the input value of the constrained activation layer, that is, the predicted power value of the spatiotemporal graph neural network model; let Represents the output of the constrained activation layer, that is, the power value that meets the linear constraint; where N G Represents the number of units, F represents the time series length; Construct a convex feasible set for all linear constraints in the unit combination problem The convex feasible set Ω is the intersection of a set of constraints in the form of m inequality constraints, and the mathematical expression is as follows: Ω={x|h i (x)≤0,i=1,...,m} Ensure that each constraint h i (x)≤0 is convex, that is, it conforms to the following mathematical expression: Determine a fixed point p in the convex feasible set Ω, that is, p∈Ω, then any point x in the set Ω is expressed as x=p+α·r, and calculate the scaling factor α≥0; is a vector from point p; for any p and r, the scaling factor α has an upper limit The mathematical expression is as follows: At this time, the line segment Belongs to the set Ω, upper limit The meaning of is the intersection of ray r and a constraint; For the point z outside the constrained feasible region in the spatiotemporal graph neural network, by specifying the ray r(z) = zp, the scaling factor The calculation is performed through the following mathematical expression, that is, projecting the point z outside the feasible domain along the ray from z to p to the point g inside the feasible domain:

2. The unit commitment scheduling method based on constrained spatiotemporal graph neural network according to claim 1 is characterized in that: The steps of constructing the unit commitment problem on a daily basis include: Assume there are N nodes, N B busbar nodes, N G The total number of decision-making periods per day in the power system is N. T ; The objective function of the unit commitment problem is constructed as follows: In the formula, g, t represent the sequence number of the unit and time period respectively; N G and N T Represent the total number of units and time periods respectively; p g,t is the output of unit g in time period t; a g , b g , c g are the coefficients of the quadratic term, linear term and constant term of the operating cost function of unit g; SU g is the startup cost of unit g; y g,t It is an integer variable indicating whether the unit g has started in time period t. A value of 1 indicates that the unit g has started in time period t, and a value of 0 indicates that the unit g has not started.

3. The unit commitment scheduling method based on constrained spatiotemporal graph neural network according to claim 2 is characterized in that: In the step of solving the corresponding optimal unit power under the constraint conditions, the constraint conditions include: Power balance constraints: Where b is the node number; N B is the total number of busbar nodes; D b,t is the load of node b in time period t; Upper and lower limit constraints of unit output: Where, and are the minimum and maximum output of unit g respectively; u g,t is an integer variable indicating the on / off status of unit g in time period t. A value of 1 indicates that unit g is on in time period t, and a value of 0 indicates that unit g is off in time period t. Unit climbing constraints: -RD g ≤p g,t -p g,t-1 ≤RU g ,g=1,…,N G ,T=2,…,N T Where, RD g and RU g They are the ramp rate limits for the unit’s g output decrease and increase, respectively; Minimum start and stop time constraints: y g,t +z g,t ≤1,g=1,…,N G ,t=1,…,N T Where TU g and TD g are the minimum start and stop time of unit g; y g,t , z g,t is an integer variable indicating whether the unit g has started or stopped in time period t. A value of 1 indicates that the unit g has started or stopped in time period t, and a value of 0 indicates that the unit g has not started or stopped. Base-state branch active power flow security constraints: Where, l is the branch number; N L is the total number of branches; is the active power flow of branch l in time period t under the base state; T is the power transfer distribution factor; g∈G(b) is the set of units at node b; F l max is the long-term allowable current-carrying power of branch l.

4. The unit commitment scheduling method based on constrained spatiotemporal graph neural network according to claim 1 is characterized in that: The steps of limiting the predicted power output by the spatiotemporal graph neural network model, including filtering the predicted results that are less than the minimum power and adjusting the predicted results that do not meet the start-stop requirements, to ensure the availability of the predicted power include: Set a threshold δ, first check and adjust the output power of the spatiotemporal graph neural network model for each unit, set the prediction result whose difference between the minimum power and the prediction result is less than the threshold δ as the minimum power of the corresponding unit, and set the prediction result whose difference between the minimum power and the prediction result is greater than the threshold δ to 0; Check and adjust the minimum start and stop time of each unit in the prediction results; The prediction results are re-passed through the linear constraint activation layer with branch active power flow security constraints added to ensure that the output power meets the branch active power flow security constraints.

5. The unit commitment scheduling method based on constrained spatiotemporal graph neural network according to claim 4 is characterized in that: The step of checking and adjusting the minimum start and stop time of each unit in the prediction result includes: Record the duration of each start and stop of each unit; Check the duration of each start and stop of each unit. For any start or stop that violates the minimum start or stop time constraint, compare the start or stop time with the previous and subsequent stop or start times in chronological order. When the start and stop time of a constraint violation is shorter than the start and stop time on both sides, adjust the corresponding start and stop.

6. A unit commitment scheduling system based on constrained spatiotemporal graph neural network, characterized in that: include: The data collection module is used to collect bus load data and power system load data of the power grid at various times of the day to establish a training data set; The optimal unit power solution module is used to construct the unit combination problem on a daily basis and solve the corresponding optimal unit power under the constraints; The training data set expansion module is used to use a sliding window to divide the bus load and the corresponding optimal unit power according to time periods to expand the training data set; A spatiotemporal graph neural network model training module is used to train a pre-established spatiotemporal graph neural network model based on the expanded training data set, using the divided bus load as input data and the corresponding optimal unit power as the data label. The spatiotemporal graph neural network model adds a linear constraint layer to the end of the spatiotemporal graph convolution, uses the linear constraint layer to linearly constrain the output of the fully connected layer, and projects the predicted power of the neural network into a feasible region that meets the specified linear constraints. The predicted power post-processing output module is used to divide the bus load data to be predicted into time periods, input the trained spatiotemporal graph neural network model, and limit the predicted power output by the spatiotemporal graph neural network model. This includes filtering the predicted results that are less than the minimum power and adjusting the predicted results that do not meet the start-stop requirements, ensuring the availability of the predicted power and obtaining the unit combination scheduling plan. The spatiotemporal graph neural network model training module uses the divided bus load as input data and the corresponding optimal unit power as a data label. The steps of training the pre-established spatiotemporal graph neural network model include: Based on the expanded training dataset, the graph neural connection relationship is obtained from the real power grid topology; The load D of all nodes in the power grid within a period of time F is input into the spatiotemporal graph convolutional network without a linear constraint layer. The graph convolution layer of the spatiotemporal graph convolutional network extracts the input spatial features. The graph convolution layer extracts and aggregates the load features of each node, learns the spatial features between the data, and expands the time series length of the data. The graph data processed by the graph convolution layer is converted into time series data. Causal dilation time convolution is then used to extract time series features from each node. A fully connected neural network layer is then used to aggregate the time series features between different nodes. The data format is then converted to output as the predicted power P of the thermal power unit within a period F. The neural network repeatedly performs the above steps by updating the loss function between the predicted power and the actual unit power scheduling data to complete the training of the spatiotemporal graph neural network model. The step of the spatiotemporal graph neural network model training module projecting the predicted power of the neural network into a feasible domain that meets the specified linear constraints includes: make Represents the input value of the constrained activation layer, that is, the predicted power value of the spatiotemporal graph neural network model; let Represents the output of the constrained activation layer, that is, the power value that meets the linear constraint; where N G Represents the number of units, F represents the time series length; Construct a convex feasible set for all linear constraints in the unit combination problem The convex feasible set Ω is the intersection of a set of constraints in the form of m inequality constraints, and the mathematical expression is as follows: Ω={x|h i (x)≤0,i=1,...,m} Ensure that each constraint h i (x)≤0 is convex, that is, it conforms to the following mathematical expression: Determine a fixed point p in the convex feasible set Ω, that is, p∈Ω, then any point x in the set Ω is expressed as x=p+α·r, and calculate the scaling factor α≥0; is a vector from point p; for any p and r, the scaling factor α has an upper limit The mathematical expression is as follows: At this time, the line segment Belongs to the set Ω, upper limit The meaning of is the intersection of ray r and a constraint; For the point z outside the constrained feasible region in the spatiotemporal graph neural network, by specifying the ray r(z) = zp, the scaling factor The calculation is performed through the following mathematical expression, that is, projecting the point z outside the feasible domain along the ray from z to p to the point g inside the feasible domain:

7. The unit commitment scheduling system based on constrained spatiotemporal graph neural network according to claim 6 is characterized in that: The optimal unit power solving module constructs the unit commitment problem on a daily basis, including the following steps: Assume there are N nodes, N B busbar nodes, N G The total number of decision-making periods per day in the power system is N. T ; The objective function of the unit commitment problem is constructed as follows: In the formula, g, t represent the sequence number of the unit and time period respectively; N G and N T Represent the total number of units and time periods respectively; p g,t is the output of unit g in time period t; a g , b g , c g are the coefficients of the quadratic term, linear term and constant term of the operating cost function of unit g; SU g is the startup cost of unit g; y g,t It is an integer variable indicating whether the unit g has started in time period t. A value of 1 indicates that the unit g has started in time period t, and a value of 0 indicates that the unit g has not started.

8. The unit commitment scheduling system based on constrained spatiotemporal graph neural network according to claim 7 is characterized in that: When the optimal unit power solving module solves the corresponding optimal unit power under the constraint conditions, the constraint conditions include: Power balance constraints: Where b is the node number; N B is the total number of busbar nodes; D b,t is the load of node b in time period t; Upper and lower limit constraints of unit output: Where, and are the minimum and maximum output of unit g respectively; u g,t is an integer variable indicating the on / off status of unit g in time period t. A value of 1 indicates that unit g is on in time period t, and a value of 0 indicates that unit g is off in time period t. Unit climbing constraints: -RD g ≤p g,t -p g,t-1 ≤RU g ,g=1,…,N G ,T=2,…,N T Where, RD g and RU g They are the ramp rate limits for the unit’s g output decrease and increase, respectively; Minimum start and stop time constraints: y g,t +z g,t ≤1,g=1,…,N G ,t=1,…,N T Where TU g and TD g are the minimum start and stop time of unit g; y g,t , z g,t is an integer variable indicating whether the unit g has started or stopped in time period t. A value of 1 indicates that the unit g has started or stopped in time period t, and a value of 0 indicates that the unit g has not started or stopped. Base-state branch active power flow security constraints: Where, l is the branch number; N L is the total number of branches; is the active power flow of branch l in time period t under the base state; T is the power transfer distribution factor; g∈G(b) is the set of units at node b; F l max is the long-term allowable current-carrying power of branch l.

9. The unit commitment scheduling system based on constrained spatiotemporal graph neural network according to claim 6 is characterized in that: The predicted power post-processing output module limits the predicted power output by the spatiotemporal graph neural network model, including filtering the predicted results that are less than the minimum power and adjusting the predicted results that do not meet the start-stop requirements. When ensuring the availability of the predicted power, a threshold δ is set. First, the output power of the spatiotemporal graph neural network model is checked and adjusted for each unit. The predicted result whose difference between the minimum power and the predicted result is less than the threshold δ is set as the minimum power of the corresponding unit, and the predicted result whose difference between the minimum power and the predicted result is greater than the threshold δ is set to 0; Check and adjust the minimum start and stop time of each unit in the prediction results; The prediction results are re-passed through the linear constraint activation layer with branch active power flow security constraints added to ensure that the output power meets the branch active power flow security constraints.

10. The unit commitment scheduling system based on constrained spatiotemporal graph neural network according to claim 9, characterized in that: The step of the predicted power post-processing output module checking and adjusting the minimum start and stop time of each unit in the prediction result includes: Record the duration of each start and stop of each unit; Check the duration of each start and stop of each unit. For any start or stop that violates the minimum start or stop time constraint, compare the start or stop time with the previous and subsequent stop or start times in chronological order. When the start and stop time of a constraint violation is shorter than the start and stop time on both sides, adjust the corresponding start and stop.

11. An electronic device, characterized in that: It comprises a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the unit combination scheduling method based on a constrained spatiotemporal graph neural network as described in any one of claims 1 to 5.

12. A computer-readable storage medium, characterized in that The computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by the processor, it implements the unit combination scheduling method based on the constrained spatiotemporal graph neural network as described in any one of claims 1 to 5.

Images