Topology-guided Security-constrained Unit Commitment Solving Method, Device and Medium

Abstract

The present invention discloses a topology-guided method, device, and medium for solving safety-constrained unit commitment, belonging to the field of power system optimization. The method comprises: training a graph convolutional neural network using power grid diagram data as input and unit start / stop scheduling plans as labels; calculating the error between the predicted start / stop states generated by the graph convolutional neural network for a generator and its target start / stop states, and determining whether the generator is a high-confidence generator based on the error and an error threshold δ; calculating a threshold α based on the predicted start / stop states of all high-confidence generators; calculating the average of the probabilities corresponding to the predicted start / stop states with correct and incorrect trends, respectively, and using the larger average as a threshold η; and in an application phase, using the trained graph convolutional neural network, α, and η to solve the safety-constrained unit commitment problem to obtain a unit start / stop scheduling strategy. A high-quality feasible solution to the safety-constrained unit commitment problem can be found in a relatively short period of time.

Description

Technical Field

[0001] The present invention belongs to the field of power system optimization, and more specifically, relates to a topology-guided safety-constrained unit combination solving method, device and medium. Background Art

[0002] Safety-constrained unit commitment involves developing multi-period unit start-up and shutdown plans that meet power system safety constraints and optimize the system's electricity purchase cost. This is a crucial issue in power system operation and scheduling. Existing approaches primarily fall into two categories: model-based and learning-based.

[0003] Model-based approaches focus on the ingenious construction of mathematical models and solution algorithms, resulting in relatively high-quality solutions. However, these approaches ignore the inherent historical similarities of safety-constrained unit commitment problems. Each solution requires significant computational resources, especially when dealing with large-scale systems. This can lead to the potential "curse of dimensionality," significantly increasing solution time or even preventing a feasible solution from being found within the specified timeframe. Learning-based approaches can leverage information from historical solutions to aid in solving new safety-constrained unit commitment problems, thereby accelerating solution times. However, these approaches require reasonable input and output data to learn the correct mapping relationship, and the quality of the solutions obtained using these approaches is difficult to guarantee, and may even be infeasible. Consequently, existing methods struggle to find a high-quality feasible solution in a short period of time.

[0004] The safety-constrained unit commitment problem is essentially a mixed-integer linear programming (MILP) problem. Its complexity increases exponentially with system size, making it increasingly difficult to solve. Finding a high-quality, feasible solution to the safety-constrained unit commitment problem in a relatively short timeframe is a pressing issue. Summary of the Invention

[0005] In response to the defects of the existing technology and the need for improvement, the present invention provides a topology-guided safety-constrained unit combination solution method, device and medium, which aims to solve the problem that the existing solution methods are unable to find a high-quality feasible solution to the safety-constrained unit combination problem in a relatively short time.

[0006] To achieve the above objectives, according to one aspect of the present invention, a topology-guided method for solving safety-constrained unit commitment is provided, comprising: establishing a graph convolutional neural network, generating a training set using historical power grid data, the training set including power grid graph data and unit start / stop scheduling plans; training the graph convolutional neural network using the training set with the power grid graph data as input and the unit start / stop scheduling plans as labels; for each generator in the unit: calculating the error between the predicted start / stop state generated by the graph convolutional neural network for the generator and the target start / stop state of the generator, and determining whether the generator is a high-confidence generator based on the relationship between the error and an error threshold δ; calculating a first threshold α for fixing the generator state based on the predicted start / stop states of all high-confidence generators in the unit; calculating the average value of the corresponding probabilities of the predicted start / stop states with a correct trend and the average value of the corresponding probabilities of the predicted start / stop states with an incorrect trend, and using the larger average value as a second threshold η; and in an application phase, solving the safety-constrained unit commitment problem using the trained graph convolutional neural network, the first threshold α, and the second threshold η to obtain a unit start / stop scheduling strategy.

[0007] Furthermore, the power grid map data includes: an adjacency matrix A describing the topological structure of the power grid, and a normalized node feature matrix describing the information contained in each node of the power grid.

[0008] Furthermore, the error is:

[0009]

[0010] Among them, d i is the error between the predicted start-stop state and the target start-stop state of generator i, u i,t,s are the predicted start / stop state and target start / stop state of generator i in sample s in period t, respectively. S is the training set, N T is the set of time periods, N G is the set of generators in the unit, n S N S The number of samples in , n T N T The number of time periods.

[0011] Furthermore, the first threshold α is:

[0012]

[0013] Among them, hc i Used to indicate whether generator i is a high-confidence generator. When it is 1, generator i is a high-confidence generator, and when it is 0, generator i is not a high-confidence generator; is the predicted start and stop status of generator i in sample s at time period t, is the number of high confidence generators in the unit, N S is the training set, N T is the set of time periods, N G is the set of generators in the unit, n S N S The number of samples in , n T N T The number of time periods.

[0014] Furthermore, the second threshold η is:

[0015]

[0016] Among them, u f is the fth predicted start-stop state with an error trend, n f is the number of predicted start and stop states with an error trend, u r is the rth predicted start and stop state with the correct trend, n r The number of predicted start and stop states with the correct trend.

[0017] Furthermore, for each predicted start-stop state: if its value is greater than the set value and the label is 1, or, if its value is not greater than the set value and the label is 0, the predicted start-stop state is a predicted start-stop state with a correct trend; if its value is greater than the set value and the label is 0, or, if its value is not greater than the set value and the label is 1, the predicted start-stop state is a predicted start-stop state with an incorrect trend.

[0018] Furthermore, the application stage specifically includes: inputting the current power grid diagram data into the trained graph convolutional neural network to obtain the start and stop state prediction value of each generator in the unit; fixing the start and stop state prediction values within [0,1-α] and [α,1] to 0 and 1 respectively, combining the already fixed start and stop state prediction values, and fixing the fixable part of the remaining start and stop state prediction values to 0-1 according to the start and stop constraints, and using it as the start and stop plan of the corresponding generator; under the premise of satisfying the constraints of the safety constraint unit combination, with the goal of minimizing the distance between the feasible unit start and stop scheduling plan and the unit's start and stop state prediction value, calculating the neighborhood lower limit r min Determine the maximum distance that the start / stop state prediction value of each generator needs to be adjusted according to the second threshold η, and take the sum of all maximum distances as the neighborhood upper limit r max ; For the unfixed start-stop state prediction value: with itself as the center and r min and r maxWithin the range formed by the radius, a corresponding feasible solution is determined as a start-stop plan for the corresponding generator; and the start-stop state of the corresponding generator is controlled according to the start-stop plan.

[0019] Furthermore, the neighborhood upper limit r max for:

[0020]

[0021]

[0022] in, for The maximum distance that needs to be adjusted, is the predicted value of the start and stop status of generator i in period t, N T is the set of time periods, N G is the collection of generators in the unit.

[0023] According to another aspect of the present invention, an electronic device is provided, comprising: a processor; and a memory storing a computer executable program, wherein when the program is executed by the processor, the processor executes the topology-guided safety constraint unit combination solution method as described above.

[0024] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that when the program is executed by a processor, the topology-guided safety-constrained unit combination solving method as described above is implemented.

[0025] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0026] (1) A topology-guided method for solving safety-constrained unit commitment is proposed. Graph convolutional neural networks are used to learn the historical similarities in safety-constrained unit commitment problems. The trained graph convolutional neural networks are used to predict the unit start / stop scheduling schemes in safety-constrained unit commitment problems and provide the start / stop probabilities of each unit in each time period. In addition, the adaptive thresholds α and η calculated based on the training data can reduce the size of the original problem and help improve the subsequent solution speed.

[0027] (2) Combining the topological structure of the power system and the node characteristics that describe the information contained in each node of the power grid to form power grid graph data can fully reflect the characteristics of the problem of safety constraint unit combination, which is conducive to improving the prediction ability of graph convolutional neural network;

[0028] (3) In the practical application stage, the predicted unit start-up and shutdown probabilities are fixed based on the threshold α, and the neighborhood upper and lower limits of the predicted unit start-up and shutdown probabilities are obtained based on the threshold η. For the unfixed unit start-up and shutdown probabilities, the feasibility of the solution can be restored in a relatively short time through neighborhood search, while also ensuring the high quality of the solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A flowchart of a topology-guided safety-constrained unit combination solving method provided by an embodiment of the present invention;

[0030] Figure 2 A schematic diagram showing changes in the loss function value during the training of a graph neural network corresponding to the IEEE-300 node system provided in an embodiment of the present invention;

[0031] Figure 3 The confusion matrix of the prediction results corresponding to the IEEE-300 node system provided in the embodiment of the present invention;

[0032] Figure 4 The IEEE-300 node system provided in the embodiment of the present invention corresponds to the number of fixed binary variables for each sample;

[0033] Figure 5 Neighborhood data calculated for each sample corresponding to the IEEE-300 node system provided in an embodiment of the present invention;

[0034] Figure 6 Comparison of the solution time of each sample corresponding to the IEEE-300 node system provided by the embodiment of the present invention using the proposed method and the MILP solver;

[0035] Figure 7 The quality comparison between the solutions obtained by the proposed method and the initial solutions for each sample corresponding to the IEEE-300 node system provided by the embodiment of the present invention;

[0036] Figure 8 Comparison of the number of nodes explored in the solution process of each sample corresponding to the IEEE-300 node system provided by the embodiment of the present invention by the proposed method and the MILP solver;

[0037] Figure 9 The comparison of the solution performance of each sample corresponding to the IEEE-300 node system provided by the embodiment of the present invention under different neighborhood upper limit conditions using the proposed method. DETAILED DESCRIPTION

[0038] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0039] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0040] Example 1

[0041] Figure 1 This is a flowchart of a topology-guided safety-constrained unit combination solution method provided by an embodiment of the present invention. Figure 1 , combined with Figure 2-Figure 9 , a topology-guided safety constraint unit combination solving method in this embodiment is described in detail, and the method includes operations S1 to S5.

[0042] Operation S1 establishes a graph convolutional neural network and uses historical power grid data to generate a training set. The training set includes power grid diagram data and unit start-stop scheduling plans. The graph convolutional neural network is trained using the training set, using the power grid diagram data as input and the unit start-stop scheduling plans as labels.

[0043] In this embodiment, the IEEE-300 node system is used as an example to demonstrate the solving capability of the topology-guided safety-constrained unit commitment solution method. The system parameters are: 300 nodes, 69 generators, and 411 lines.

[0044] The unit start-stop scheduling scheme in the training set is obtained by solving the original safety-constrained unit commitment problem. The objective function of the original safety-constrained unit commitment problem is to minimize the operating cost and startup cost of the units in the system, which is in the following form:

[0045]

[0046] Where f(·) represents the generator operation cost calculation function, P i,t is the output of generator i in period t, C i,t is the start-up and shutdown cost of generator i in period t.

[0047] The constraints of the original safety-constrained unit commitment problem include system balance constraints, system spinning reserve constraints, ramping constraints, minimum start-stop constraints, and branch constraints.

[0048] The system equilibrium constraint is:

[0049]

[0050] Among them, P t L is the total load demand of the system in period t.

[0051] The system spinning reserve constraints are:

[0052]

[0053] in, are the maximum technical output and minimum technical output of generator i, are the upper and lower backup demands of the system in time period t respectively.

[0054] The climbing constraint is:

[0055]

[0056] in, are respectively the up-climbing power limit and down-climbing power limit of generator i, are the maximum starting power and maximum shutdown power of generator i respectively.

[0057] The minimum start-stop constraint is:

[0058]

[0059] in, are the minimum start-up time and minimum shutdown time of generator i respectively.

[0060] The branch constraints are:

[0061] -f L ≤|TG·P-TL·L|≤f L

[0062] Among them, TG and TL are the power transmission transfer distribution factor matrices of the generator and the load respectively, P and L are the power matrices of the generator and the load respectively, and f L is the line transmission capacity matrix.

[0063] Preferably, the power grid map data includes: an adjacency matrix A describing the topological structure of the power grid, and a normalized node feature matrix describing the information contained in each node of the power grid.

[0064] The information of the power grid can be represented by graph data, which contains two important pieces of information: the adjacency matrix A that describes the topological structure of the power grid and the characteristic matrix X that describes the information contained in each node of the power grid. The node types in the power grid can be divided into two categories according to whether there is a generator. Each node in the power grid is characterized by load data for T time periods. For nodes with generators, their characteristics also include the technical parameters of the generator: maximum technical output, minimum technical output, up-ramp power limit, down-ramp power limit, maximum starting power, maximum shutdown power, minimum starting time, minimum shutdown time, and parameters related to operating costs. In addition to the basic information on these nodes, the constraints related to line transmission capacity between nodes also need to be taken into account to form the characteristic matrix X, whose structure is as follows:

[0065]

[0066]

[0067] Among them, x j is the eigenvector of node j, is the eigenvector of the constraints related to line transmission capacity of node j, is the load information of node j in period t, is the generator characteristic of node j, is the line transmission constraint between node j and node k. If there is no line between the two nodes, The corresponding value is 0, otherwise, Corresponding to 1. g j It is a binary variable used to indicate whether node j has a generator. If there is a generator, the value is 1, and if not, the value is 0.

[0068] Different features have different dimensions and need to be normalized. The operation can be shown as follows:

[0069]

[0070] Among them, x min 、x max are the maximum and minimum values in each column of feature data, is the feature after normalization. The node feature matrix after normalization is recorded as

[0071] In this embodiment, the graph convolutional neural network is trained to learn the potential relationship between the input data set and the label set. The input of the graph convolutional neural network is the normalized node feature matrix and the adjacency matrix A, which includes several layers of graph convolutional layers and the final fully connected layer. The graph convolutional layer process mainly includes the aggregation and linear transformation of node features, and its process can be shown as follows:

[0072]

[0073] Among them, H l is the output feature matrix of the lth graph convolutional layer, W l is the training parameter of the l graph convolutional layers, and σ(·) is the activation function. If A is used directly for calculation, the numerical range of the feature will be inaccurate, so A needs to be normalized. The operation is as follows:

[0074]

[0075] Where I is the degree matrix of the graph, D is the node degree matrix of the graph, is the normalized adjacency matrix. Therefore, the graph convolution process can be expressed as follows:

[0076]

[0077] After the feature transformation of the graph convolution layer and the fully connected layer, a T-dimensional feature matrix is obtained. The feature vector of the node corresponding to the generator node is extracted as the corresponding unit start-up and shutdown probability. The corresponding process can be expressed by the following formula:

[0078]

[0079] Among them, G(·) represents the processing process of graph convolutional neural network, F(·) represents the process of extracting the start-stop probability vector of the corresponding unit, and U p It is a predictive unit start-up and shutdown scheduling plan.

[0080] The learning task of graph neural networks can be viewed as a multi-label binary classification task. The binary cross entropy loss function can be used to measure the gap between the predicted value and the target value:

[0081]

[0082] Among them, Loss is the gap between the predicted value and the target value. is the target unit start-stop scheduling plan, are the predicted unit state and target unit state of generator i in time period t, respectively.

[0083] Operation S2, for each generator in the unit: calculate the error between the predicted start / stop state generated by the graph convolutional neural network for the generator and the target start / stop state of the generator, and determine whether the generator is a high-confidence generator based on the relationship between the error and the error threshold δ.

[0084] Preferably, for any generator i, after training is completed, the error between the predicted start / stop state generated by the graph convolutional neural network for generator i and the target start / stop state of generator i is:

[0085]

[0086] Among them, d i is the error between the predicted start-stop state and the target start-stop state of generator i, u i,t,s are the predicted start / stop state and target start / stop state of generator i in sample s in period t, respectively. S is the training set, N T is the set of time periods, N G is the set of generators in the unit, n S N S The number of samples in , n T N T The closer the predicted value of the graph convolutional neural network is to the target value, the higher the credibility of the corresponding generator prediction result.

[0087] For any generator i, using hc i To indicate whether it is a high confidence generator, if d i Less than or equal to the error threshold δ, hc i is 1, indicating that generator i is a high confidence generator. If d i Greater than the error threshold δ, hc i is 0, indicating that generator i is not a high-confidence generator.

[0088] Operation S3 : calculating a first threshold α for fixing the generator state according to the predicted start / stop states of all high-confidence generators in the generator set.

[0089] In determining After a high confidence generator is determined, a first threshold α for fixing the generator state is calculated based on the performance of the determined high confidence generator in the training sample. Preferably, the first threshold α is:

[0090]

[0091] Among them, hc i Used to indicate whether generator i is a high-confidence generator. When it is 1, generator i is a high-confidence generator, and when it is 0, generator i is not a high-confidence generator; is the number of high-confidence generators in the unit, and max(·) indicates the maximum value is selected.

[0092] In operation S4, an average value of the predicted start / stop state corresponding probabilities with a correct trend and an average value of the predicted start / stop state corresponding probabilities with an incorrect trend are calculated, and the larger average value is used as the second threshold η.

[0093] According to an embodiment of the present invention, for each predicted start / stop state: if its value is greater than a set value and the label is 1, or if its value is not greater than the set value and the label is 0, the predicted start / stop state is a predicted start / stop state with a correct trend; if its value is greater than the set value and the label is 0, or if its value is not greater than the set value and the label is 1, the predicted start / stop state is a predicted start / stop state with an incorrect trend. The set value is, for example, 0.5.

[0094] Preferably, the second threshold η is:

[0095]

[0096] Among them, u f is the fth predicted start-stop state with an error trend, n f is the number of predicted start and stop states with an error trend, u r is the rth predicted start and stop state with the correct trend, n r The number of predicted start and stop states with the correct trend.

[0097] Operation S5, the application phase, uses the trained graph convolutional neural network, the first threshold α and the second threshold η to solve the safety constrained unit commitment problem to obtain the unit start and stop scheduling strategy.

[0098] Preferably, according to an embodiment of the present invention, operation S5 specifically includes sub-operations S51 to S56.

[0099] In sub-operation S51, the current power grid diagram data is input into the trained graph convolutional neural network to obtain the start and stop state prediction value of each generator in the unit.

[0100] In sub-operation S52, the start-stop state prediction values within the range of [0, 1-α] and [α, 1] are fixed to 0 and 1 respectively. Combined with the already fixed start-stop state prediction values, the fixable parts of the remaining start-stop state prediction values are fixed to 0-1 according to the start-stop constraints, and used as the start-stop plan of the corresponding generator.

[0101] For all generator start / stop state predictions (probability values between 0 and 1) within the entire dispatch cycle, if the value is in the range [α, 1], the corresponding generator state is fixed to start; if the value is in the range [0, 1-α], the corresponding generator state is fixed to shut down. Start / stop state predictions within the range (1-α, α) are considered to have insufficient confidence, and directly fixing them carries the risk of making the final unit dispatch plan infeasible.

[0102] Based on the start-stop constraints, the fixed portion of the remaining start-stop state predictions is fixed to 0-1. The principle is as follows. After the threshold is fixed, there may be a few unfixed generator states between two determined generator states. Suppose that after the first threshold α is fixed, the partial start-stop states of a generator are ...0000011x111..., where x represents the unfixed unit state. The fifth and sixth states in the displayed state transition from 0 to 1, indicating that the generator is starting. If the minimum start time period of this generator is 5, then the generator state should be set to 1 in all five consecutive states after startup. Therefore, it can be determined that the state x is also 1, which is subject to the minimum start time period of the minimum start-stop constraint. The final start-stop state of the generator is ...00000111111...

[0103] Therefore, in order to further determine more generator states and reduce the scale of the problem, the start and stop constraints of the generators in the safety constraint unit commitment problem can be used to determine whether these states can be directly fixed:

[0104] U pf =T(U pα )

[0105] Where T(·) represents the function of fixing the unit state by using the generator start-stop constraint, U pα is the predicted unit start-stop scheduling plan after the first threshold α is fixed, U pf It is a predictive unit start-stop scheduling plan that is fixed in 0-1 state.

[0106] In sub-operation S53, under the premise of satisfying the constraints of the safety constraint unit combination, the neighborhood lower limit r is calculated with the goal of minimizing the distance between the feasible unit start-stop scheduling plan and the unit start-stop state prediction value. min . can be expressed as follows:

[0107] min[(U pf ,U nf )]

[0108]

[0109] Satisfy the above constraints. Among them, Uns is the feasible unit start-stop scheduling scheme obtained through neighborhood search, Δ(U pf ,U ns ) is U pf and U ns The Hamming distance between It's U pf The state of generator i in time period t, It's U ns The state of generator i in period t; U nf is the initial feasible unit start-stop scheduling plan, and the distance U pf The nearest feasible solution. By setting this distance as the objective function of the original safety-constrained unit commitment problem, the neighborhood lower limit r can be obtained. min .

[0110] In sub-operation S54, the maximum distance that the start / stop state prediction value of each generator needs to be adjusted is determined according to the second threshold η, and the sum of all maximum distances is used as the neighborhood upper limit r max .

[0111] Preferably, the neighborhood upper limit r max for:

[0112]

[0113]

[0114] in, for The maximum distance that needs to be adjusted, is the predicted value of the start and stop status of generator i in period t, N T is the set of time periods, N G is the collection of generators in the unit.

[0115] In sub-operation S55, for each unfixed start-stop state prediction value: with itself as the center and r min and r max Within the range formed by the radius, the corresponding feasible solution is determined as the start-stop plan of the corresponding generator.

[0116] According to the obtained r min and r max The following new constraints can be obtained, under which a feasible unit start-up and shutdown scheduling plan can be quickly obtained while ensuring its quality:

[0117] r min <(U pf ,U ns )<r max

[0118] Adding the fixed start-stop state prediction value and the above constraints into the original safety-constrained unit commitment problem can obtain a sub-safety-constrained unit commitment problem with a reduced scale and a greatly reduced feasible domain. On this basis, using the MILP solver to solve the problem can quickly obtain a high-quality feasible unit start-stop scheduling plan.

[0119] In sub-operation S56 , the start and stop states of the corresponding generators are controlled according to the start and stop plan.

[0120] Specifically, according to the start-stop scheme obtained in sub-operation S52 and the start-stop scheme obtained in sub-operation S54, the start-stop status of the corresponding generator in the unit is controlled.

[0121] In this embodiment, the scheduling period is set to 24 periods (ie N T =24) safety-constrained unit combination mathematical model, based on which 2196 sets of corresponding input data and labels were obtained. The data were divided into training, validation, and test sets in a 6:2:2 ratio. Results from 20 difficult examples in the test set were selected to demonstrate the solving capability of the proposed method.

[0122] The simulation results are as follows Figure 2 As shown, analysis Figure 2 It can be seen that with the increase in the number of training rounds, the loss function value of the graph convolutional neural network continues to decrease, and the loss function value of the training set is slightly lower than that of the validation set. This shows that the trained network is not overfitting or underfitting, and can be used to predict the unit start and stop scheduling plan for new safety-constrained unit combination problems.

[0123] Through this method, the first threshold α for fixing the generator state is obtained to be 0.964. The confusion matrix is used to analyze the accuracy of the trained graph convolutional neural network in predicting trends with a boundary of 0.5. The results are as follows: Figure 3 Analysis Figure 3 It can be seen that the accuracy of the predicted trend has reached more than 90%. According to the prediction results of the training set, the second threshold η is calculated to be 0.911.

[0124] According to the state of the fixed unit α, the number of fixed binary variables is as follows Figure 4 As shown in , an average of 774 binary variables can be fixed, accounting for 46.74% of the total number of binary variables, which greatly helps to reduce the scale of the original problem. The calculated neighborhood parameters are as follows Figure 5 As shown in Figure 2, the size of the neighborhood can reflect the accuracy and reliability of the predicted unit start-up and shutdown scheduling plan. min The smaller the value, the closer the current prediction unit start-stop scheduling plan with fixed state is to a feasible solution. max The smaller the value, the closer the predicted unit start-stop scheduling plan with fixed state is to the optimal solution. minand r max The smaller the difference, the smaller the feasible domain that needs to be searched.

[0125] First, use this method to solve the sample and get the value of the objective function. Then set this value as the target value of the MILP solver and solve the corresponding sample. Comparing the solution time of the two methods, the results are as follows: Figure 6 As shown in the figure, the proposed method takes significantly less time than the MILP solver to find a unit start / stop scheduling solution of equivalent quality. In the samples presented, the proposed method saved up to 898.66 seconds of solution time in sample 6, and achieved the fastest solution speed improvement of 89.42% in sample 7, with an average solution speed improvement of 75.36%.

[0126] In sub-operation S51, an initial feasible unit start-up and shutdown scheduling plan is obtained, but its quality cannot be guaranteed. A high-quality feasible unit start-up and shutdown scheduling plan can be obtained under the obtained custom neighborhood upper and lower limit conditions. Figure 7 The difference between the two unit start-up and shutdown scheduling schemes is shown. The comparison shows that the final solution obtained by the proposed method is much higher quality than the initial feasible solution and is relatively stable, with guaranteed quality. The average gap corresponding to the solution is 0.047%, very close to the optimal solution. The gap is calculated as follows:

[0127]

[0128] Among them, COST cal is the calculated objective function value, COST opt is the optimal objective function value.

[0129] In order to further analyze the solution efficiency of this method, Figure 8 The number of nodes explored by the proposed algorithm and MILP solver during the solution process is shown. During the solution process, the interval where the optimal objective function value of the problem is located will continue to shrink as the search tree is explored. Therefore, the fewer nodes that need to be explored in the search tree, the less time will be consumed in solving the problem. Figure 8 Analysis shows that this method can greatly reduce the number of nodes that need to be explored, with an average of 45,802 fewer nodes explored, and a maximum of 101,504 fewer nodes in sample 5.

[0130] according to Figure 5 Analysis of max It is a very important parameter, and its size affects the solution time and the quality of the final solution. Taking sample 8 as an example, the current r max As the starting point, 5 as the interval, a total of 8 different r max Analyze the solution effect of the proposed algorithm, the results are as follows Figure 9As shown. According to the changing trend of the curve, as r max As , the solution time becomes longer and longer, and the quality of the solution finally obtained becomes higher and higher, and finally stabilizes at the optimal solution. The most reasonable r max In the interval from 203.34 to 213.34, the r calculated by the proposed algorithm max Relatively conservative. max The solution time becomes longer and longer as the value increases, but it is still much lower than Figure 6 The solution time of the MILP solver is shown. This is because the proposed method fixes a large number of binary variables, which greatly reduces the difficulty of solving the original problem.

[0131] The topology-guided safety-constrained unit combination solution method provided by an embodiment of the present invention uses historical data to train a graph convolutional neural network to learn the potential relationship between unit start-up and shutdown and power grid diagram data, and calculates an adaptive threshold based on the training set results for subsequent variable fixation and neighborhood search; uses the trained graph convolutional neural network to predict unit start-up and shutdown probabilities, and fixes the high-confidence unit start-up and shutdown states in the predicted unit start-up and shutdown probabilities based on the adaptive threshold to achieve scale reduction of the original problem, and uses the information contained in the predicted unit start-up and shutdown probabilities and the adaptive threshold to calculate the customized neighborhood upper and lower limits, and restores the feasibility of the prediction results through the neighborhood search method while ensuring its high quality.

[0132] Example 2

[0133] An electronic device includes: a processor; and a memory storing a computer-executable program. When the program is executed by the processor, the processor executes the above-mentioned topology-guided safety-constrained unit combination solving method. The related technical solutions are the same as those in the first embodiment and are not further described here.

[0134] Example 3

[0135] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned topology-guided safety-constrained unit combination solution method. The relevant technical solutions are the same as those in the first embodiment and will not be described in detail here.

[0136] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A topology-guided safety-constrained unit commitment solution method, characterized in that: include: Establish a graph convolutional neural network and use historical power grid data to generate a training set, which includes power grid graph data and unit start-stop scheduling plans; The graph convolutional neural network is trained using the training set, with the power grid diagram data as input and the unit start-stop scheduling plan as a label; For each generator in the unit: calculating the error between the predicted start / stop state generated by the graph convolutional neural network for the generator and the target start / stop state of the generator, and determining whether the generator is a high-confidence generator based on the relationship between the error and an error threshold δ; Calculating a first threshold α for fixing the generator state based on the predicted start / stop states of all high-confidence generators in the unit; Calculate the average value of the predicted start / stop state probabilities with a correct trend and the average value of the predicted start / stop state probabilities with an incorrect trend, and use the larger average value as the second threshold η; In the application phase, the trained graph convolutional neural network, the first threshold α, and the second threshold η are used to solve the safety-constrained unit commitment problem to obtain the unit start-stop scheduling strategy. The first threshold α is: in, Used to indicate whether generator i is a high-confidence generator. When it is 1, generator i is a high-confidence generator, and when it is 0, generator i is not a high-confidence generator; is the predicted start and stop status of generator i in sample s at time period t, is the number of high-confidence generators in the unit, is the training set, is a collection of time periods, is the set of generators in the unit, for The number of samples in for the number of time periods; The second threshold η is: in, For the A predicted start-stop state with an error trend, is the number of predicted start and stop states with an error trend, For the A predicted start and stop state with the correct trend, The number of predicted start and stop states with the correct trend.

2. The topology-guided safety-constrained unit combination solving method according to claim 1, characterized in that: The power grid map data includes: an adjacency matrix describing the topological structure of the power grid , and the normalized node feature matrix describing the information contained in each node of the power grid .

3. The topology-guided safety-constrained unit combination solving method according to claim 1, characterized in that: The error is: in, is the error between the predicted start-stop state and the target start-stop state of generator i, 、 are the predicted start / stop state and target start / stop state of generator i in sample s in period t, respectively. is the training set, is a collection of time periods, is the set of generators in the unit, for The number of samples in for The number of time periods.

4. The topology-guided safety-constrained unit combination solving method according to claim 1, characterized in that: For each predicted start and stop state: If its value is greater than the set value and the label is 1, or if its value is not greater than the set value and the label is 0, the predicted start-stop state is a predicted start-stop state with a correct trend; If its value is greater than the set value and the label is 0, or if its value is not greater than the set value and the label is 1, the predicted start-stop state is a predicted start-stop state with an error trend.

5. The topology-guided safety-constrained unit combination solving method according to any one of claims 1 to 4, characterized in that: The application stage specifically includes: The current power grid diagram data is input into the trained graph convolutional neural network to obtain the start and stop status prediction value of each generator in the unit; The value in 、 The start-stop state prediction values in the grid are fixed to 0 and 1 respectively. Combined with the fixed start-stop state prediction values, the fixable part of the remaining start-stop state prediction values is fixed to 0-1 according to the start-stop constraints, and used as the start-stop plan of the corresponding generator; Under the premise of satisfying the constraints of the safety constraint unit combination, the neighborhood lower limit is calculated with the goal of minimizing the distance between the feasible unit start-stop scheduling plan and the unit start-stop state prediction value. ; The maximum distance that the start / stop state prediction value of each generator needs to be adjusted is determined according to the second threshold η, and the sum of all maximum distances is used as the neighborhood upper limit ; For the unfixed prediction values of each start-stop state: and Within the range formed by the radius, determine the corresponding feasible solution as the start-stop plan of the corresponding generator; The start and stop states of the corresponding generators are controlled according to the start and stop scheme.

6. The topology-guided safety-constrained unit combination solving method according to claim 5, characterized in that: The neighborhood upper limit for: in, for The maximum distance that needs to be adjusted, is the predicted value of the start and stop status of generator i in period t, is a collection of time periods, is the collection of generators in the unit.

7. An electronic device, characterized in that: include: processor; A memory storing a computer executable program, wherein when the program is executed by the processor, the processor executes the topology-guided safety-constrained unit commitment solving method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the topology-guided safety-constrained unit commitment solving method according to any one of claims 1 to 6 is implemented.

Images