Power system unit combination method and system

Abstract

The invention provides a power system unit combination method and system. The method comprises the following steps: training a pre-constructed mixed integer programming model through a pre-constructed graph data sample set and a residual graph convolutional neural network to obtain a trained graph neural network, and forming a mapping relationship between input data features of the graph neural network and data labels of graph data samples in the graph data sample set; explaining the trained graph neural network through the mapping relation and a graph neural network interpreter based on mask learning to obtain an interpretable graph neural network; converting the mixed integer programming model into a linear programming model through a unit start-stop strategy output by the interpretable graph neural network; and according to a solving result obtained by solving the linear programming model, selecting and outputting an optimal solution of the linear programming model or outputting an optimal solution of the mixed integer programming model as a unit commitment scheduling result. According to the invention, the credibility of the unit commitment scheduling decision is improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of power system optimal scheduling, and in particular to a power system unit commitment method and system. BACKGROUND

[0002] The power system is one of the infrastructures supporting the operation of modern society, and the unit commitment problem is a crucial link in the operation of the power system. The unit commitment problem refers to determining the start-stop and output plan of the unit in the dispatching period under the constraints of meeting system load requirements and unit operation technical limitations, so as to realize the safety and economy of system operation. The unit commitment problem has integer variables such as unit start-stop state, and its mathematical model is a mixed integer programming with high-dimensional discrete non-convex non-linear characteristics, and it is essentially an NP-Hard problem. The traditional mixed integer programming method directly solves the unit commitment problem, which faces the challenges of "combinatorial explosion" and "dimension disaster", and it is difficult to obtain a high-quality unit scheduling scheme in a short time.

[0003] With the development of artificial intelligence technology, machine learning, neural networks and other artificial intelligence technologies provide innovative solutions for efficient solution of the unit commitment problem. However, the existing artificial intelligence methods still face many challenges. On the one hand, most of the existing artificial intelligence technologies only take structured data as input, ignoring the huge topology and complex correlation of the actual power system, and it is difficult to capture the dynamic changes of the interaction relationship between nodes; on the other hand, in power system dispatching, the transparency and explainability of decision-making is crucial. Most of the existing artificial intelligence technologies lack clear model understanding and explanation mechanism, making it difficult for decision-makers to understand the working process inside the network, thereby reducing the credibility of the dispatching decision. SUMMARY

[0004] Therefore, embodiments of the present application provide a power system unit commitment method and system to at least partially solve the above problems.

[0005] According to a first aspect of an embodiment of the present application, a power system unit commitment method is provided, comprising training a pre-constructed mixed integer programming model through a pre-constructed graph data sample set and a residual graph convolutional neural network to obtain a trained graph neural network, and forming a mapping relationship between input data features of the graph neural network and data labels of graph data samples in the graph data sample set, wherein the mixed integer programming model is used to express power system unit commitment of the graph neural network, and the graph data sample set is used for training of the graph neural network; interpreting the trained graph neural network through the mapping relationship and a graph neural network interpreter based on mask learning to obtain an interpretable graph neural network; converting the mixed integer programming model into a linear programming model through a unit start-stop strategy output by the interpretable graph neural network; and selecting an optimal solution of the linear programming model or an optimal solution of the mixed integer programming model as a unit commitment scheduling result according to a solution result obtained by solving the linear programming model.

[0006] In an implementation manner, the mixed integer programming model is constructed by the following manner:

[0007] A target function of minimizing power system operation cost is constructed:

[0008]

[0009] wherein g represents a generator unit index, G represents a generator unit set, t represents a scheduling time period index, T represents a scheduling time period set, P g,t is an active power of the unit g in the t time period, f g (P g,t ) represents a cost function of the unit, , and are respectively a start cost and a stop cost of the unit g, u g,t represents an on-off signal of the unit g at the time t;

[0010] A constraint condition of the mixed integer programming model is constructed, and the constraint condition comprises a system power balance constraint, a system reserve constraint, a network maximum transmission capacity constraint, and a unit operation constraint, wherein the unit operation constraint specifically comprises a unit operation state logic constraint, an output limit constraint, a minimum start-stop time constraint, and a maximum ramp rate constraint.

[0011] The mixed integer programming model is constructed according to the target function and the constraint condition.

[0012] In another implementation, the input data features of the graph neural network are generated by: generating a plurality of load scenarios by a Monte Carlo sampling method; obtaining generator parameters; constructing node features of a graph data sample in the graph data sample set according to the plurality of load scenarios and the generator parameters; constructing an adjacency matrix of the graph data sample according to a power grid topology structure; and generating the input data features of the graph neural network according to the node features of the graph data sample and the adjacency matrix of the graph data sample.

[0013] In another implementation, the data labels of the graph data sample in the graph data sample set are generated by: solving a mixed integer programming model under each load scenario in the plurality of load scenarios according to the input data features to obtain optimal unit start-stop integer variables; and taking the optimal unit start-stop integer variables as the data labels of the graph data sample.

[0014] In another implementation, the node features of the graph data sample and the adjacency matrix of the graph data sample are constructed by:

[0015]

[0016]

[0017] wherein X represents node features, n represents a scenario number, D n represents an nth load scenario, N represents a scenario set, Y ij represents an element value of an i-th row and a j-th column of an adjacency matrix, P g min represents a minimum output of a unit g, P g max represents a maximum output of the unit g.

[0018] In another implementation, the training of the pre-constructed mixed integer programming model by the pre-constructed graph data sample set and the residual graph convolutional neural network comprises:

[0019] standardizing the graph data samples in the graph data sample set into a test set and a validation set, and initializing weights and other parameters of the graph neural network by random initialization or pre-training;

[0020] for each training sample in the graph data sample set, performing forward propagation to calculate an output of the mixed integer programming model, and in the residual graph convolutional neural network, the process of forward propagation involves a plurality of graph convolutional layers and residual connections, and the forward propagation is realized by:

[0021]

[0022] wherein H k is the node representation of the k-th layer, Y is the result of the adjacency matrix plus the self-loop, Q is the degree matrix, W k is the weight matrix of the k-th layer, and σ is the activation function.

[0023] The loss function is calculated using the difference between the output of the mixed integer programming model and the real label:

[0024]

[0025] wherein a n represents the predicted output, U n represents the actual label.

[0026] The gradient of the loss with respect to the parameters of the mixed integer programming model is calculated by the Adam optimizer, and the parameters of the mixed integer programming model are updated according to the gradient information to minimize the loss function.

[0027] The parameters of the mixed integer programming model are updated according to the gradient by the gradient descent optimization algorithm until the stopping condition of training is reached.

[0028] The performance of the mixed integer programming model is evaluated by the validation set, and the hyperparameters are adjusted to prevent overfitting.

[0029] The generalization performance of the mixed integer programming model is evaluated by the test set.

[0030] In another implementation, the step of selecting the optimal solution of the linear programming model or the optimal solution of the mixed integer programming model as the unit commitment scheduling result according to the solution result of the linear programming model includes: solving the linear programming model to obtain the solution result; if the solution result indicates that the linear programming model is feasible, outputting the optimal solution of the linear programming model as the unit commitment scheduling result; if the solution result indicates that the linear programming model is not feasible, solving the mixed integer programming model, and outputting the optimal solution of the mixed integer programming model as the unit commitment scheduling result.

[0031] According to a second aspect of the embodiment of the present application, there is provided a power system unit commitment system, comprising a training module configured to train a pre-constructed mixed integer programming model by a pre-constructed graph data sample set and a residual graph convolutional neural network, to obtain a trained graph neural network, and to form a mapping relationship between input data features of the graph neural network and data labels of graph data samples in the graph data sample set, wherein the mixed integer programming model is configured to express power system unit commitment of the graph neural network, and the graph data sample set is configured to train the graph neural network; an interpretation module configured to interpret the trained graph neural network by the mapping relationship and a mask learning-based graph neural network interpreter, to obtain an interpretable graph neural network; a conversion module configured to convert the mixed integer programming model into a linear programming model by a unit start-stop strategy output by the interpretable graph neural network; and an output module configured to select an optimal solution of the linear programming model or an optimal solution of the mixed integer programming model as a unit commitment scheduling result according to a solution result obtained by solving the linear programming model.

[0032] In an implementation manner, the mixed integer programming model is constructed by the following manner:

[0033] a target function of minimizing power system operation cost is constructed:

[0034]

[0035] wherein g represents a generator unit index, G represents a set of generator units, t represents a scheduling period index, T represents a set of scheduling periods, P g,t represents active power of the unit g in the period t, f g (P g,t ) represents a cost function of the unit, , and represent start-up cost and shut-down cost of the unit g respectively, u g,t represents an on-off signal of the unit g at the time t;

[0036] a constraint condition of the mixed integer programming model is constructed, and the constraint condition comprises a system power balance constraint, a system reserve constraint, a network maximum transmission capacity constraint, and a unit operation constraint, wherein the unit operation constraint specifically comprises a unit operation state logic constraint, an output limit constraint, a minimum start-stop time constraint, and a maximum ramp rate constraint;

[0037] the mixed integer programming model is constructed according to the target function and the constraint condition.

[0038] In another implementation, the input data features of the graph neural network are generated by: generating a plurality of load scenarios by a Monte Carlo sampling method; obtaining generator parameters; constructing node features of a graph data sample in the graph data sample set according to the plurality of load scenarios and the generator parameters; constructing an adjacency matrix of the graph data sample according to a power grid topology structure; and generating the input data features of the graph neural network according to the node features of the graph data sample and the adjacency matrix of the graph data sample.

[0039] In another implementation, the data labels of the graph data sample in the graph data sample set are generated by: solving a mixed integer programming model under each load scenario in the plurality of load scenarios according to the input data features to obtain optimal unit start-stop integer variables; and taking the optimal unit start-stop integer variables as the data labels of the graph data sample.

[0040] In another implementation, the node features of the graph data sample and the adjacency matrix of the graph data sample are constructed by:

[0041]

[0042]

[0043] wherein X represents node features, n represents a scenario number, D n represents an nth load scenario, N represents a scenario set, Y ij represents an element value of an ith row and jth column of an adjacency matrix, P g min represents a minimum output of a unit g, P g max represents a maximum output of the unit g.

[0044] In another implementation, the training module is specifically configured to:

[0045] standardize the graph data samples in the graph data sample set and divide the graph data samples into a test set and a validation set, and initialize weights and other parameters of the graph neural network by random initialization or pre-training;

[0046] for each training sample in the graph data sample set, forward propagation is performed to calculate an output of the mixed integer programming model, and in the residual graph convolutional neural network, the process of forward propagation involves a plurality of graph convolutional layers and residual connections, and the forward propagation is implemented by:

[0047]

[0048] wherein H kis the representation of the node of the kth layer, Y is the result of the adjacency matrix plus the self-loop, Q is the degree matrix, W k is the weight matrix of the kth layer, and sigma is the activation function.

[0049] The loss function is calculated using the difference between the output of the mixed integer programming model and the real label:

[0050]

[0051] where a n represents the predicted output, U n represents the actual label.

[0052] The gradient of the loss with respect to the parameters of the mixed integer programming model is calculated by the Adam optimizer, and the parameters of the mixed integer programming model are updated according to the gradient information to minimize the loss function.

[0053] The parameters of the mixed integer programming model are updated according to the gradient by the gradient descent optimization algorithm until the stopping condition of training is reached.

[0054] The performance of the mixed integer programming model is evaluated by the validation set, and the hyperparameters are adjusted to prevent overfitting.

[0055] The generalization performance of the mixed integer programming model is evaluated by the test set.

[0056] In another implementation, the output module is specifically configured to: solve the linear programming model to obtain the solution result; if the solution result indicates that the linear programming model is feasible, output the optimal solution of the linear programming model as the unit commitment scheduling result; if the solution result indicates that the linear programming model is not feasible, solve the mixed integer programming model, and output the optimal solution of the mixed integer programming model as the unit commitment scheduling result.

[0057] According to a third aspect of the embodiments of the present application, an electronic device is provided, including a processor and a memory storing a program. The program includes instructions that, when executed by the processor, cause the processor to perform the steps performed by the method of the first aspect.

[0058] According to a fourth aspect of the embodiments of the present application, a computer storage medium is provided, which stores a computer program. The program is executed by a processor to implement the method of the first aspect.

[0059] To sum up, the scheme of the embodiment of the present application trains the mapping relationship between the load scene and the unit start-stop variable through the graph convolutional neural network, and interprets the trained graph neural network based on the mask learning-based graph neural network interpreter, converts the mixed integer programming model into a linear programming model, i.e. a linear programming problem, through the unit start-stop strategy output by the interpretable graph neural network, significantly improves the decision speed of the unit combined dispatching, overcomes the challenge of "combination explosion" and "dimension disaster" faced by directly solving the unit combined problem, and improves the solving efficiency of the power system unit combination. In addition, the method of the present application takes the graph structure data as the sample input, can consider the huge topological structure and complex correlation of the actual power system. And by fusing the graph neural network and the interpretable mechanism, the decision maker can more easily understand the working process inside the network, and the credibility of the dispatching decision is improved. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0061] Figure 1 The step flow chart of the power system unit combination method of the present application;

[0062] Figure 2 The framework diagram of the power system unit combination method corresponding to the embodiment; Figure 1 The framework diagram of the power system unit combination method corresponding to the embodiment;

[0063] Figure 3 The working principle diagram of the mask learning-based graph neural network interpreter of the present application;

[0064] Figure 4 The structure block diagram of the power system unit combination system corresponding to the embodiment; Figure 1 The structure block diagram of the power system unit combination system corresponding to the embodiment;

[0065] Figure 5 The structure diagram of an electronic device of the present application. DETAILED DESCRIPTION

[0066] In order to have a more clear understanding of the technical features, purposes and effects of the embodiments of the present application, the specific implementation ways of the embodiments of the present application will be described with reference to the drawings.

[0067] In this document, "exemplarily" means "serving as an example, an instance or an illustration", and any illustration, embodiment described as "exemplarily" in this document should not be interpreted as a more preferred or more advantageous technical solution.

[0068] In the specific implementation, the method comprises the following steps.

[0069] The specific implementation of the method is further described below with reference to the accompanying drawings of the embodiments of the present application.

[0070] Referring to Figure 1 , Figure 2 The method for unit commitment of a power system mainly comprises the following steps.

[0071] Step S1: training a pre-constructed mixed integer programming model by using a pre-constructed graph data sample set and a residual graph convolutional neural network, obtaining a trained graph neural network, and forming a mapping relationship between input data features of the graph neural network and data labels of graph data samples in the graph data sample set, wherein the mixed integer programming model is used to express unit commitment of the graph neural network, and the graph data sample set is used for training of the graph neural network.

[0072] Step S2: interpreting the trained graph neural network by using the mapping relationship and a graph neural network interpreter based on mask learning, and obtaining an interpretable graph neural network.

[0073] Step S3: converting the mixed integer programming model into a linear programming model by using a unit start-stop strategy output by the interpretable graph neural network.

[0074] Step S4: selecting an optimal solution of the linear programming model or an optimal solution of the mixed integer programming model as a unit commitment scheduling result according to a solution result obtained by solving the linear programming model.

[0075] In an implementation manner, the mixed integer programming model is constructed by the following manner.

[0076] A target function of minimizing power system operation cost is constructed.

[0077]

[0078] wherein g represents a generator unit index, G represents a set of generator units, t represents a scheduling period index, T represents a set of scheduling periods, P g,t is an active power of the unit g in the t period, f g (P g,t ) represents a cost function of the unit, , and are respectively the start-up cost and the shut-down cost of the unit g, u g,t represents the start signal of the unit g at time t;

[0079] The constraint conditions of the mixed integer programming model are constructed, and the constraint conditions include a system power balance constraint, a system reserve constraint, a network maximum transmission capacity constraint, and a unit operation constraint, wherein the unit operation constraint specifically includes a unit operation state logic constraint, an output limit constraint, a minimum start-stop time constraint, and a maximum ramp rate constraint;

[0080] The mixed integer programming model is constructed according to the objective function and the constraint conditions.

[0081] In another implementation manner, the input data features of the graph neural network are generated by: generating a plurality of load scenarios by a Monte Carlo sampling method; obtaining generator parameters; constructing node features of a graph data sample in the graph data sample set according to the plurality of load scenarios and the generator parameters; constructing an association matrix of the graph data sample according to a power grid topology structure; and generating the input data features of the graph neural network according to the node features of the graph data sample and the association matrix of the graph data sample.

[0082] In another implementation manner, the data labels of the graph data sample in the graph data sample set are generated by: solving a mixed integer programming model under each load scenario in the plurality of load scenarios according to the input data features, to obtain optimal unit start-stop integer variables; and taking the optimal unit start-stop integer variables as the data labels of the graph data sample.

[0083] In another implementation manner, the node features of the graph data sample and the association matrix of the graph data sample are constructed by:

[0084]

[0085]

[0086] wherein X represents the node features, n represents a scenario number, D n represents the nth load scenario, N represents a scenario set, Y ij represents an element value of the i-th row and the j-th column of the association matrix, P g min represents the minimum output of the unit g, P g max represents the maximum output of the unit g.

[0087] In another implementation, the training of the pre-constructed mixed integer programming model by the pre-constructed graph data sample set and the residual graph convolutional neural network comprises:

[0088] The graph data samples in the graph data sample set are divided into a test set and a validation set after standardization processing, and the weights and other parameters of the graph neural network are initialized by random initialization or pre-training;

[0089] For each training sample in the graph data sample set, forward propagation is performed to calculate the output of the mixed integer programming model, and in the residual graph convolutional neural network, the process of forward propagation involves multiple graph convolutional layers and residual connections, and the forward propagation is realized by the following formula:

[0090]

[0091] Wherein, H k is the node representation of the kth layer, Y is the result of the adjacency matrix plus the self-loop, Q is the degree matrix, W k is the weight matrix of the kth layer, and σ is the activation function.

[0092] The difference between the output of the mixed integer programming model and the true label is used to calculate the loss function:

[0093]

[0094] Wherein, a n represents the predicted output, and U n represents the actual label.

[0095] The gradient of the loss to the parameters of the mixed integer programming model is calculated by the Adam optimizer, and the parameters of the mixed integer programming model are updated according to the gradient information to minimize the loss function;

[0096] The parameters of the mixed integer programming model are updated according to the gradient by the gradient descent optimization algorithm until the stopping condition of training is reached.

[0097] The performance of the mixed integer programming model is evaluated by the validation set, and the hyperparameters are adjusted to prevent overfitting;

[0098] The generalization performance of the mixed integer programming model is evaluated by the test set.

[0099] In another implementation, according to a solution result obtained by solving the linear programming model, an optimal solution of the linear programming model or an optimal solution of the mixed integer programming model is selected as the unit commitment scheduling result, and the method comprises the following steps: solving the linear programming model to obtain the solution result; if the solution result indicates that the linear programming model is feasible, outputting the optimal solution of the linear programming model as the unit commitment scheduling result; and if the solution result indicates that the linear programming model is not feasible, solving the mixed integer programming model and outputting the optimal solution of the mixed integer programming model as the unit commitment scheduling result.

[0100] Specifically, the scheme of the application is further described according to the following embodiments:

[0101] Before step S1, a mixed integer programming model is constructed, which is used to express the power system unit commitment of the graph neural network. The objective function of the mixed integer programming model is to minimize the system operation cost, and the constraint conditions include system power balance constraints, system reserve constraints, network maximum transmission capacity constraints and unit operation constraints.

[0102] The objective function of the mixed integer programming model is to minimize the operation cost of the power system:

[0103]

[0104] Wherein: g represents the generator unit index; G represents the generator unit set; t represents the scheduling period index; T represents the scheduling period set; P g,t represents the active power of the unit g at the t period; f g (P g,t ) represents the cost function of the unit, which is usually a convex quadratic function, and in engineering, in order to improve the solving efficiency, it is usually expressed as a piecewise linear function; , and are the start-up cost and shutdown cost of the unit g respectively. u g,t represents the start signal of the unit g at time t, if the unit is in the shutdown state at t-1 and is turned on at t, then u g,t =1; otherwise, u g,t =0. Let v g,t represent the shutdown signal of the unit g at time t, if the unit is in the start state at t-1 and is turned off at t, then v g,t =1; otherwise, v g,t =0.

[0105] The constraint conditions of the mixed integer programming model include system power balance constraints, system reserve constraints, network maximum transmission capacity constraints and unit operation constraints.

[0106] 1) System power balance constraint

[0107] The power balance of power system refers to the balance between power supply and demand, which requires the total power generation and total load demand of the power system to be balanced at all times, which can be expressed as:

[0108]

[0109] where D t is the total active power of the load at time t.

[0110] 2) System reserve constraint

[0111] The amount of rotating reserve of thermal power unit refers to the reserve capacity left to balance the load fluctuation and load prediction error of the power grid. Generally, the rotating reserve is the sum of the difference between the current output of synchronous unit in the system and its maximum available output, which can be expressed as:

[0112]

[0113] where x g,t is the operating state of unit g at time t, if the unit is started, x g,t = 1; otherwise, x g,t = 0. is the maximum output of unit g; RD t is the rotating reserve capacity at time t.

[0114] 3) Network maximum transmission capacity constraint

[0115] Network is a power transmission network, and its physical boundary is the line maximum transmission capacity boundary. The main impact of renewable energy consumption is the lack of maximum transmission capacity, which causes line congestion. The maximum transmission capacity boundary is usually modeled by DC power flow, which is expressed as follows. The line maximum transmission capacity boundary refers to the power flowing through the line cannot exceed a certain threshold, which is usually modeled by DC power flow as:

[0116]

[0117] where b represents the bus number; B represents the bus set; l represents the line number; L represents the line set; represents; D b,t is the load of bus b at time t; H PTDF is the line power flow distribution factor matrix; F l max is the transmission capacity boundary.

[0118] 4) Unit operation constraint

[0119] The unit operation constraint specifically includes unit operation state logic constraint, output limit constraint, minimum start-up and shutdown time constraint, and maximum ramp rate constraint.

[0120] The unit operation state logical constraints characterize the relationship between the generator g operation state x g,t and the start-up and shut-down state u g,t and v g,t . If the generator g starts up at time t, i.e., x g,t-1 = 0 and x g,t = 1, then u g,t = 1; otherwise, v g,t = 0. Similarly, if the generator g shuts down at time t, i.e., x g,t-1 = 1 and x g,t = 0, then v g,t = 1; otherwise v g,t = 0. The above logic can be formalized as:

[0121] u g,t + v g,t ≤ 1, g G, t T (5)

[0122] x g,t - x g,t-1 = u g,t - v g,t , g G, t T (6)

[0123] The power output limit constraints require the corresponding power output upper and lower bounds and cannot operate outside the power output interval, which can be expressed as:

[0124]

[0125] where: and are the minimum power output and the maximum power output of the unit g, respectively. The maximum power output limit depends on the rated capacity of the unit design and manufacture, which is the maximum capacity that can be continuously output during the entire normal service life. The minimum power output limit, also known as the technical minimum output, refers to the minimum power output that the boiler can stably burn and the generator set can continuously and stably operate under technical condition constraints.

[0126] The minimum start-up and shut-down time constraints require that the unit continuous on / off state should not be less than a certain time limit, which is modeled as:

[0127]

[0128]

[0129] where: and The minimum on / off time of the unit g. This is because the start-stop of the thermal power unit is realized by the start-stop of the boiler and the steam turbine, and the frequent start-stop of the boiler will lead to the damage of the boiler pipe joint, and the frequent start-stop of the steam turbine will cause the bending of the shaft, which will greatly reduce the service life of the thermal power unit. In order to reduce the wear and tear of parts and mechanical fatigue, the thermal power unit cannot be started and stopped continuously in a short time.

[0130] The maximum ramp rate constraint means that the output change of the unit in two consecutive adjacent time periods cannot exceed a certain threshold when the unit is continuously running, which is modeled as:

[0131] -R g ≤P g,t -P g,t-1 ≤R g , g∈G,t∈T (10)

[0132] Where: R g represents the maximum value of the upward / downward ramp of the unit g. Generally, the ramp rate of the thermal power unit is 10% to 20% of the rated power per hour, and the ramp rate of the gas and oil unit is faster, which can reach 50% to 100% of the rated power per hour.

[0133] Before step S1, a graph data sample set is constructed, which is used for graph neural network training.

[0134] A plurality of load scenarios are generated by the Monte Carlo sampling method, node features X of the graph data samples in the graph data sample set are constructed based on the load scenarios and generator parameters, and an association matrix Y of the graph data samples is constructed based on the power grid topology structure, so as to form the input data features IN of the graph neural network.

[0135] The unit combination mixed integer programming model under each load scenario is solved to obtain the corresponding optimal unit start-stop integer variable value, which is used as the data label U of the graph data sample.

[0136] Specifically, the node features X and the association matrix Y of the graph data sample are constructed by the following formula:

[0137]

[0138]

[0139] Where: n represents the scenario number; D n represents the nth load scenario; N represents the scenario set; Y ij represents the element value of the ith row and jth column of the association matrix, P g min represents the minimum output of the unit g, P g max represents the maximum output of the unit g.

[0140] Step S1: training a pre-constructed mixed integer programming model through a pre-constructed graph data sample set and a residual graph convolutional neural network, obtaining a trained graph neural network, and forming a mapping relationship between input data features of the graph neural network and data labels of graph data samples in the graph data sample set, wherein the mixed integer programming model is used to express unit commitment of the graph neural network, and the graph data sample set is used for training of the graph neural network.

[0141] The model training based on the residual graph convolutional neural network specifically includes the following steps:

[0142] 1) Initialize network parameters: standardize the graph data samples in the graph data sample set, and then test the set and the validation set. Randomly initialize or pre-train the weights and other parameters of the graph neural network.

[0143] 2) Forward propagation: for each training sample in the graph data sample set, forward propagation is performed to calculate the output of the model. In the residual graph convolutional neural network, the process of forward propagation involves multiple graph convolutional layers and residual connections. Forward propagation is achieved by the following formula:

[0144]

[0145] wherein H k is the node representation of the kth layer, Y is the result of the adjacency matrix plus the self-loop, Q is the degree matrix, W k is the weight matrix of the kth layer, and σ is the activation function.

[0146] 3) Loss calculation: the difference between the output of the model and the true label is used to calculate the loss function:

[0147]

[0148] wherein a n represents the predicted output, and U n represents the actual label.

[0149] 4) Back propagation: the gradient of the loss with respect to the model parameters is calculated by the Adam optimizer, and the parameters of the mixed integer programming model are updated according to the gradient information to minimize the loss function.

[0150] 5) Parameter update: using gradient descent or other optimization algorithms, the parameters of the mixed integer programming model are updated according to the gradient until the stopping condition of training is reached.

[0151] 6) Verification and testing: the performance of the mixed integer programming model parameters is evaluated using the validation set, the hyperparameters are adjusted to prevent overfitting. Finally, the generalization performance of the mixed integer programming model parameters is evaluated on the test set.

[0152] Step S2: The trained graph neural network is explained by the mapping relationship and the mask learning-based graph neural network interpreter to obtain an interpretable graph neural network, and the most influential subgraph structure and node feature on the unit start-stop variable prediction result are identified, so as to realize the physical interpretation of the unit combined dispatching decision. See Figure 3 The working principle diagram of the mask learning-based graph neural network interpreter.

[0153] Step S3: The mixed integer programming model is converted into a linear programming model through the unit start-stop strategy output by the interpretable graph neural network.

[0154] Specifically, the integer variable in the mixed integer programming model is fixed as a constant through the unit start-stop strategy output by the interpretable graph neural network, so that the mixed integer programming model is converted into a linear programming model, that is, a linear programming problem.

[0155] Step S4: According to the solving result obtained by solving the linear programming model, the optimal solution of the linear programming model or the optimal solution of the mixed integer programming model is selected as the unit combined dispatching result.

[0156] Specifically, the linear programming model is solved to obtain a solving result.

[0157] If the solving result indicates that the linear programming model is feasible, the optimal solution of the linear programming model is output as the unit combined dispatching result.

[0158] If the solving result indicates that the linear programming model is not feasible, the mixed integer programming model is solved, and the optimal solution of the mixed integer programming model is output as the unit combined dispatching result.

[0159] In summary, the scheme of the embodiment of the present application trains the mapping relationship between the load scenario and the unit start-stop variable through the graph convolutional neural network, and explains the trained graph neural network based on the mask learning-based graph neural network interpreter. The mixed integer programming model is converted into a linear programming model, that is, a linear programming problem, through the unit start-stop strategy output by the interpretable graph neural network, which significantly improves the decision speed of the unit combined dispatching, overcomes the challenge of "combination explosion" and "dimension disaster" faced by directly solving the unit combined problem, and improves the efficiency of the unit combined solution of the power system. In addition, the method of the present application takes the graph structure data as the sample input, can consider the huge topology structure and complex correlation of the actual power system. And by fusing the graph neural network and the interpretable mechanism, the decision maker can more easily understand the working process inside the network, and the credibility of the dispatching decision is improved.

[0160] As another example, see Figure 4The present invention provides a power system unit combination system 400, comprising:

[0161] Training module 410 is used to train a pre-built mixed integer programming model using a pre-built graph data sample set and a residual graph convolutional neural network to obtain a trained graph neural network, and to form a mapping relationship between the input data features of the graph neural network and the data labels of the graph data samples in the graph data sample set, wherein the mixed integer programming model is used to describe the power system unit combination of the graph neural network, and the graph data sample set is used for training the graph neural network;

[0162] The interpretation module 420 is used to interpret the trained graph neural network through the mapping relationship and the graph neural network interpreter based on mask learning, so as to obtain an interpretable graph neural network.

[0163] The conversion module 430 is used to convert the mixed integer programming model into a linear programming model through the unit start-up and shutdown strategy output by the interpretable graph neural network.

[0164] The output module 440 is used to select either the optimal solution of the linear programming model or the optimal solution of the mixed integer programming model as the unit combination scheduling result based on the solution result obtained from solving the linear programming model.

[0165] In one implementation, the mixed-integer programming model is constructed as follows:

[0166] Construct an objective function that minimizes the operating cost of the power system:

[0167]

[0168] Where g represents the generator set index, G represents the generator set set, t represents the scheduling period index, T represents the scheduling period set, and P represents the generator set index. g,t f is the active power of unit g during time period t. g (P g,t ) represents the cost function of the generator set. ,and These represent the start-up cost and shutdown cost of unit g, respectively, and u. g,t This represents the start-up signal of unit g at time t;

[0169] The constraints for constructing the mixed integer programming model include system power balance constraints, system reserve constraints, network maximum transmission capacity constraints, and unit operation constraints. Specifically, the unit operation constraints include unit operation state logic constraints, output limit constraints, minimum start-up and shutdown time constraints, and maximum ramp rate constraints.

[0170] The mixed integer programming model is constructed according to the objective function and the constraint condition.

[0171] In another implementation manner, the input data features of the graph neural network are generated in the following manner: a plurality of load scenarios are generated by a Monte Carlo sampling method; generator parameters are acquired; node features of a graph data sample in the graph data sample set are constructed according to the plurality of load scenarios and the generator parameters; an association matrix of the graph data sample is constructed according to a power grid topology structure; and the input data features of the graph neural network are generated according to the node features of the graph data sample and the association matrix of the graph data sample.

[0172] In another implementation manner, data labels of the graph data sample in the graph data sample set are generated in the following manner: a mixed integer programming model under each load scenario in the plurality of load scenarios is solved according to the input data features, to obtain optimal generator start-stop integer variables; and the optimal generator start-stop integer variables are taken as the data labels of the graph data sample.

[0173] In another implementation manner, the node features of the graph data sample and the association matrix of the graph data sample are constructed in the following manner:

[0174]

[0175]

[0176] wherein X represents node features, n represents a scenario number, D n represents an nth load scenario, N represents a scenario set, Y ij represents an element value of an ith row and jth column of an association matrix, P g min represents a minimum output of a generator g, P g max represents a maximum output of the generator g.

[0177] In another implementation manner, the training module 410 is specifically configured to:

[0178] the graph data samples in the graph data sample set are divided into a test set and a validation set after being standardized, and the weights and other parameters of the graph neural network are initialized in a random initialization or pre-training manner;

[0179] for each training sample in the graph data sample set, forward propagation is performed to calculate an output of the mixed integer programming model, and in the residual graph convolutional neural network, the process of forward propagation involves a plurality of graph convolutional layers and residual connections, and the forward propagation is implemented in the following manner:

[0180]

[0181] where H k is the node representation of the k-th layer, Y is the result of the adjacency matrix plus self-loop, Q is the degree matrix, W k is the weight matrix of the k-th layer, and σ is the activation function.

[0182] The loss function is calculated using the difference between the output of the mixed integer programming model and the real label:

[0183]

[0184] where a n represents the predicted output, and U n represents the actual label.

[0185] The gradient of the loss with respect to the parameters of the mixed integer programming model is calculated by the Adam optimizer, and the parameters of the mixed integer programming model are updated according to the gradient information to minimize the loss function.

[0186] The parameters of the mixed integer programming model are updated according to the gradient by the gradient descent optimization algorithm until the stopping condition of training is reached.

[0187] The performance of the mixed integer programming model is evaluated by the validation set, and the hyperparameters are adjusted to prevent overfitting.

[0188] The generalization performance of the mixed integer programming model is evaluated by the test set.

[0189] In another implementation, the output module 440 is specifically configured to: solve the linear programming model to obtain the solving result; if the solving result indicates that the linear programming model is feasible, output the optimal solution of the linear programming model as the unit commitment scheduling result; if the solving result indicates that the linear programming model is not feasible, solve the mixed integer programming model, and output the optimal solution of the mixed integer programming model as the unit commitment scheduling result.

[0190] To sum up, the scheme of the embodiment of the present application trains the mapping relationship between the load scene and the unit start-stop variable through the graph convolutional neural network, and interprets the trained graph neural network based on the mask learning graph neural network interpreter, converts the mixed integer programming model into a linear programming model, i.e. a linear programming problem, through the unit start-stop strategy output by the interpretable graph neural network, significantly improves the decision speed of the unit commitment scheduling, overcomes the challenge of "combination explosion" and "dimension disaster" faced by directly solving the unit commitment problem, and improves the solving efficiency of the power system unit commitment. In addition, the method of the present application takes the graph structure data as the sample input, can consider the huge topology structure and complex correlation of the actual power system. And by fusing the graph neural network and the interpretable mechanism, the decision maker can more easily understand the working process inside the network, and the credibility of the scheduling decision is improved.

[0191] The power system unit commitment system of the embodiment is used to implement the corresponding method in the plurality of method embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be described here.

[0192] As another example, see Figure 5 An electronic device 500 will now be described, which can be a server or a client of the present application, and is an example of a hardware device that can be applied to aspects of the present application. The electronic device is intended to represent various forms of digital electronic computing devices such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computing devices. The electronic device can also represent various forms of mobile devices such as personal digital processing, cellular telephones, smart phones, wearable devices, and other similar computing devices. The components shown here, their connections, and their functions, are meant to be examples only, and are not intended to limit the implementations of the present application described and / or claimed herein.

[0193] The electronic device 500 can include a processor 502, a communications interface 504, a memory 506, and a communications bus 508.

[0194] The processor 502, the communications interface 504, and the memory 506 complete mutual communication through the communications bus 508. The communications interface 504 is used to communicate with other electronic devices or servers.

[0195] The processor 502 is configured to execute the program 510, and can specifically execute the related steps in the method embodiments described above.

[0196] Specifically, the program 510 can include program code, which includes computer operation instructions.

[0197] The processor 502 can be a processor CPU, or an application specific integrated circuit ASIC, or one or more integrated circuits configured to implement embodiments of the present application. The one or more processors of the smart device can be the same type of processor, such as one or more CPUs, or different types of processors, such as one or more CPUs and one or more ASICs.

[0198] The memory 506 is configured to store a program 510. The memory 506 can include a high-speed RAM memory, and can further include a non-volatile memory, such as at least one disk memory.

[0199] The program 510, when executed by the processor 502, is configured to cause the electronic device to perform a power system unit commitment method, including: training a pre-constructed mixed integer programming model through a pre-constructed graph data sample set and a residual graph convolutional neural network, obtaining a trained graph neural network, and forming a mapping relationship between input data features of the graph neural network and data labels of graph data samples in the graph data sample set, wherein the mixed integer programming model is used to express power system unit commitment of the graph neural network, and the graph data sample set is used for training of the graph neural network; interpreting the trained graph neural network through the mapping relationship and a graph neural network interpreter based on mask learning, to obtain an interpretable graph neural network; converting the mixed integer programming model into a linear programming model through a unit start-stop strategy output by the interpretable graph neural network; and selecting an optimal solution of the linear programming model or an optimal solution of the mixed integer programming model as a unit commitment scheduling result according to a solving result obtained by solving the linear programming model.

[0200] In addition, specific implementation of each step in the program 510 can refer to the corresponding description in the corresponding steps and units of the foregoing method embodiments, and will not be described herein. It can be clearly understood by those skilled in the art that, for the convenience and brevity of description, the specific working process of the device and the module described above can refer to the corresponding process description in the foregoing method embodiments, and will not be described herein.

[0201] The exemplary embodiments of the present application also provide a computer storage medium storing a computer program, wherein the computer program is executed by a processor to implement the method of the embodiments of the present application, and can refer to the corresponding process description in the foregoing method embodiments, and will not be described herein.

[0202] The above-described methods according to embodiments of the application can be implemented in hardware, firmware, or software, or any combination thereof, and can be implemented as software storable on a recording medium which is readable from a general use computer, a special processor, or programmable or special hardware (such as an ASIC or an FPGA) using such software. It is understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component (for example, RAM, ROM, flash memory, etc.) that can store or receive software or computer code when accessed and executed by the computer, the processor, or the hardware, to implement the methods described herein. Furthermore, when a general use computer accesses code for implementing the methods shown herein, the execution of the code transforms the general use computer into a special purpose computer for executing the methods shown herein.

[0203] Thus far, specific embodiments of the application have been described. Other embodiments are within the scope of the following claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve desirable results. Additionally, the processes depicted in the figures do not necessarily require the particular order shown, or sequential order, to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous.

[0204] It should be understood that, although the specification is described with respect to various embodiments, not every embodiment contains only one independent technical solution, and the specification is described in this way only for clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that those skilled in the art can understand.

[0205] Finally, it should be noted that: the above implementation is only used to illustrate the embodiments of the present application, and not to limit the embodiments of the present application, and those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present application. Therefore, all equivalent technical solutions also belong to the scope of the embodiments of the present application, and the patent protection scope of the embodiments of the present application should be defined by the claims.

Claims

1. A power system unit commitment method, characterized in that, The application relates to a method for constructing an interpretable graph neural network for unit commitment. The method comprises the following steps: training a pre-constructed mixed integer programming model through a pre-constructed graph data sample set and a residual graph convolutional neural network to obtain a trained graph neural network and form a mapping relationship between input data features of the graph neural network and data labels of graph data samples in the graph data sample set, wherein the mixed integer programming model is used for expressing unit commitment of the graph neural network, and the graph data sample set is used for training the graph neural network; interpreting the trained graph neural network through the mapping relationship and a graph neural network interpreter based on mask learning to obtain an interpretable graph neural network; transforming the mixed integer programming model into a linear programming model through a unit start-stop strategy output by the interpretable graph neural network; 2. The method of claim 1, wherein, selecting an optimal solution of the linear programming model or an optimal solution of the mixed integer programming model as a unit commitment scheduling result according to a solving result of the linear programming model. The mixed integer programming model is constructed in the following manner: where g denotes the generator unit index, G denotes the generator unit set, t denotes the dispatch period index, T denotes the dispatch period set, P g,t is the active power of unit g in period t, f g (P g,t ) denotes the cost function of the unit, , and are the start-up cost and shut-down cost of unit g, respectively, u g,t denotes the start signal of unit g at time t; a target function of minimizing power system operation cost is constructed; constraint conditions of the mixed integer programming model are constructed, the constraint conditions comprising system power balance constraints, system reserve constraints, network maximum transmission capacity constraints and unit operation constraints, wherein the unit operation constraints specifically comprise unit operation state logic constraints, output limit constraints, minimum start-stop time constraints and maximum climbing rate constraints; 3. The method of claim 1, wherein, the mixed integer programming model is constructed according to the target function and the constraint conditions. Input data features of the graph neural network are generated in the following manner: a plurality of load scenarios are generated through a Monte Carlo sampling method; generator unit parameters are acquired; node features of graph data samples in the graph data sample set are constructed according to the plurality of load scenarios and the generator unit parameters; an association matrix of the graph data samples is constructed according to a power grid topology structure; 4. The method of claim 3, wherein, input data features of the graph neural network are generated according to the node features of the graph data samples and the association matrix of the graph data samples. Data labels of the graph data samples in the graph data sample set are generated in the following manner: optimal unit start-stop integer variables are obtained by solving a mixed integer programming model under each load scenario in the plurality of load scenarios according to the input data features; 5. The method of claim 3, wherein, the optimal unit start-stop integer variables are taken as data labels of the graph data samples. wherein X represents a node feature, n represents a scene number, D n represents the nth load scene, N represents a scene set, Y ij represents an element value of the i-th row and the j-th column of the correlation matrix, P g min represents the minimum output of the unit g, P g max represents the maximum output of the unit g.

6. The method of claim 1, wherein, The node features of the graph data samples and the association matrix of the graph data samples are constructed in the following manner: The training of the pre-constructed mixed integer programming model through the pre-constructed graph data sample set and the residual graph convolutional neural network comprises the following steps: graph data samples in the graph data sample set are divided into a test set and a validation set after standardization processing, and weights and other parameters of the graph neural network are initialized in a random initialization or pre-training manner; for each training sample in the graph data sample set, forward propagation is performed to calculate an output of the mixed integer programming model, in the residual graph convolutional neural network, the process of forward propagation involves multiple graph convolutional layers and residual connections, and forward propagation is realized in the following formula: where H k is the node representation of the kth layer, Y is the result of the adjacency matrix plus self-loops, Q is the degree matrix, W k is the weight matrix of the kth layer, and σ is the activation function. using the difference between the output of the mixed integer programming model and the true label to calculate a loss function: where a n represents the predicted output, U n represents the actual label; calculating the gradient of the loss with respect to the parameters of the mixed integer programming model using an Adam optimizer, updating the parameters of the mixed integer programming model according to the gradient information to minimize the loss function; updating the parameters of the mixed integer programming model according to the gradient through a gradient descent optimization algorithm until a stopping condition of training is reached; evaluating the performance of the mixed integer programming model through the validation set, adjusting the hyperparameters to prevent overfitting; evaluating the generalization performance of the mixed integer programming model through the test set.

7. The method according to any one of claims 1 to 6, characterized in that, The method according to the solving result of the linear programming model, selecting the optimal solution of the linear programming model or the optimal solution of the mixed integer programming model as the unit commitment scheduling result, comprising: solving the linear programming model to obtain the solving result; if the solving result indicates that the linear programming model is feasible, outputting the optimal solution of the linear programming model as the unit commitment scheduling result; if the solving result indicates that the linear programming model is not feasible, solving the mixed integer programming model, and outputting the optimal solution of the mixed integer programming model as the unit commitment scheduling result.

8. A power system unit commitment system, characterized in that, comprising: a training module configured to train a pre-constructed mixed integer programming model through a pre-constructed graph data sample set and a residual graph convolutional neural network, obtain a trained graph neural network, and form a mapping relationship between input data features of the graph neural network and data labels of graph data samples in the graph data sample set, wherein the mixed integer programming model is used to express unit commitment of the graph neural network, and the graph data sample set is used for training of the graph neural network; an interpretation module configured to interpret the trained graph neural network through the mapping relationship and a mask learning-based graph neural network interpreter to obtain an interpretable graph neural network; a conversion module configured to convert the mixed integer programming model into a linear programming model through a unit start-stop strategy output by the interpretable graph neural network; an output module configured to select an optimal solution of the linear programming model or an optimal solution of the mixed integer programming model as a unit commitment scheduling result according to a solving result of the linear programming model.

9. An electronic device, comprising: comprising: a processor; a memory storing a program; wherein the program includes instructions that, when executed by the processor, cause the processor to perform the steps of the method according to any one of claims 1-7.

10. A computer storage medium, characterized in that, a computer program stored thereon, which is executed by a processor to implement the method according to any one of claims 1-7.

Images