Solver training method and solving method for power system security constrained unit commitment model

Abstract

The present application belongs to the technical field of power system dispatching, and discloses a solver training method and a solving method for a power system security constrained unit commitment model. The training method comprises: obtaining multiple unit commitment models and generating multiple sets of near-optimal solutions for each unit commitment model by a near-optimal solution generation method; constructing a training sample set, each training sample comprising parameter features of a unit commitment model and a near-optimal solution; training a graph neural network to obtain a solver; the near-optimal solution generation method comprises: adding a near-optimal constraint to the unit commitment model to obtain a near-optimal solution generation model; randomly perturbing the coefficients of the objective function and adding a diversity constraint to obtain an updated model and solve it, updating the latest solution to the current obtained solution and adding the current obtained solution to the solution set; repeating the above process until the solution set has a preset number of near-optimal solutions. By inputting the parameter features of the unit commitment model to be solved into the solver, the value of the power resource dispatching variable can be quickly predicted.

Description

Technical Field

[0001] This invention belongs to the technical field of power system dispatching, and more specifically, relates to a solver training method and a solution method for a power system safety constraint unit combination model. Background Technology

[0002] With the increasing scale of power systems, independent system operators (ISOs) have a more pressing need for efficiency in solving large-scale security-constrained unit commitment (SCUC) problems. The goal of ISOs in solving SCUC problems is to optimize the allocation of generation resources, determine the operating status and output levels of these resources, and minimize the total operating cost of the entire system, while satisfying operating conditions such as system balance constraints, reserve constraints, and network constraints.

[0003] Currently, large-scale SCUC problems are typically modeled as mixed-integer linear programming problems, using branch and bound methods with integer programming solvers. However, due to the NP-hard nature of mixed-integer programming problems, it is difficult to search for the optimal solution to large-scale SCUC problems, resulting in enormous computational costs. Independent system operators typically allocate only 1-3 hours for solving large-scale SCUC problems. Therefore, finding a high-quality approximate optimal solution in a shorter time is crucial for system operators to reduce system operating costs. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a solver training method and a solution method for a power system safety-constrained unit combination model, the purpose of which is to find a high-quality approximate optimal solution for the power system safety-constrained unit combination model in a short time.

[0005] To achieve the above objectives, the present invention provides a solver training method for a power system safety-constrained unit combination model, comprising:

[0006] Multiple safety-constrained unit combination models of historical power systems are obtained. For each unit combination model, multiple near-optimal solutions for the power resource scheduling variables in the corresponding unit combination model are generated through a near-optimal solution generation method.

[0007] Construct a training sample set, where the input of each training sample is the parameter features of the unit combination model and the label is the near-optimal solution of the corresponding unit combination model;

[0008] Using the training sample set, a graph neural network is trained to obtain a solver that predicts the values ​​of power resource scheduling variables in the unit combination model based on the parameter features of the unit combination model.

[0009] The near-optimal solution generation method includes:

[0010] Step S11: Obtain the optimal solution X of the unit combination model. * and the objective function value z corresponding to the optimal solution * Set the number of near-optimal solutions N. sol Given a tolerance of q and a value greater than 1, near-optimal constraints are added to the unit combination model to obtain a near-optimal solution generation model. The near-optimal constraints are that the objective function of the unit combination model takes a value no greater than (1+q)z. * Initialize the solution set S = {X} *}, initialize the latest solution X new =X * ;

[0011] Step S12: Randomize the coefficients in the objective function of the near-optimal solution generation model and add diversity constraints to the near-optimal solution generation model to obtain an updated model. The perturbation amplitude does not exceed 10% of the original coefficient values. The diversity constraints are used to ensure that the solution to be generated is different from the latest solution X. new ;

[0012] Step S13: Solve and update the model, and apply the latest solution X new Update the solution to the current solution and add the current solution to the solution set S;

[0013] Step S14: Determine whether the number of solutions in the solution set S reaches N. sol If not, proceed to step S12; if yes, output the solution set S as multiple near-optimal solutions for the corresponding unit combination model.

[0014] Preferably, the variable is an integer variable.

[0015] Preferably, the variable is a binary variable;

[0016] In step S12, the diversity constraint is:

[0017]

[0018] In the formula, i is the variable number, B is the set of variables, and x i Let i be the i-th variable to be decided. The latest solution for X new The solution corresponding to the i-th variable. x represents i Let X be the i-th variable in the variable set B and its latest solution X new The value in is 0. x represents i Let X be the i-th variable in the variable set B and its latest solution Xnew The value in is 1.

[0019] Preferably, the variables include unit start-up variables, shutdown variables, and status variables.

[0020] Preferably, the tolerance q does not exceed 10%.

[0021] This invention also provides a method for solving a power system safety-constrained unit combination model, which includes:

[0022] A power system unit combination model with security constraints to be solved is determined, wherein the unit combination model has power resource scheduling variables to be decided;

[0023] Extract the parameter features of the unit combination model and input them into the solver, then use the solver to predict the values ​​of the power resource scheduling variables;

[0024] The solver is trained using the training method described above.

[0025] Preferably, after predicting the values ​​of the power resource scheduling variables using the solver, the method further includes correcting the prediction results:

[0026] Variables whose predicted values ​​reach a preset confidence level are classified as reliable variables.

[0027] A deviation term is added to the objective function of the unit combination model to obtain a modified model. The deviation term = W * cumulative deviation, where the cumulative deviation is the accumulation of the absolute deviations between all the confidence variables to be decided and the predicted values ​​of the corresponding confidence variables, and W is a preset weighting coefficient.

[0028] The modified model is solved to obtain the values ​​of the power resource scheduling variables.

[0029] Preferably, after the correction, a solution quality assessment step is also included:

[0030] The decision bias of the corresponding device is obtained by calculating the sum of the deviations between the solution results of all variables corresponding to any device in the power system in the modified model and their predicted values ​​in the solver.

[0031] Based on decision bias, equipment in the power system is classified into stable equipment and unstable equipment;

[0032] The solution results for the variables corresponding to the stable equipment are directly substituted into the unit combination model, and the unit combination model is solved again to obtain the values ​​of the variables corresponding to the unstable equipment.

[0033] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0034] The present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the method described in any of the preceding claims.

[0035] In summary, compared with the prior art, the technical solutions conceived in this invention have the following main advantages:

[0036] 1. This invention utilizes a trained neural network as a solver. By inputting the parameter features of the unit combination model into the solver, the values ​​of power resource scheduling variables can be quickly predicted, thus enabling the solution of the unit combination model in a short time. Since training the neural network requires a large amount of training data, and historical data is relatively limited, and a single optimal solution exists for a safety-constrained unit combination model in historical data, insufficient training data leads to poor training performance. Therefore, this invention also mentions a near-optimal solution generation method. By adding near-optimal constraints to the unit combination model, the solution obtained can be near-optimal, i.e., close to the optimal solution. By adding perturbations and diversity constraints, it can be ensured that each solution obtained is different from existing solutions, avoiding repeated acquisition of the same near-optimal solution. In this way, multiple near-optimal solutions to the unit combination model can be quickly obtained, thereby expanding the training data, improving the training effect of the neural network, and thus improving the accuracy of the solver. By inputting the model parameters of the power system safety-constrained unit combination model to be solved into the solver, the values ​​of the power resource scheduling variables in the corresponding unit combination model can be obtained. This enables rapid solution of the fast model.

[0037] 2. Preferably, the method proposed in this invention is used to predict the value of integer variables, which has higher prediction accuracy and a simpler process.

[0038] 3. Preferably, after predicting the value of the variable using the solver, the present invention also performs a correction step to improve the accuracy of the obtained solution.

[0039] 4. Preferably, after performing the correction step, the present invention further performs a solution quality evaluation step to further improve the accuracy of the obtained solution. Attached Figure Description

[0040] Figure 1 This is a flowchart of the solver training method in one embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram illustrating the effect of a near-optimal solution generation method in one embodiment of the present invention;

[0042] Figure 3 This is a flowchart illustrating the steps of a method for solving a power system safety constraint unit combination model according to an embodiment of the present invention.

[0043] Figure 4 This is a flowchart of the steps for correcting the prediction results in one embodiment of the present invention;

[0044] Figure 5 This is a flowchart of the steps for evaluating the quality of a solution in one embodiment of the present invention;

[0045] Figure 6 This is a comparison chart of the time efficiency of using the method of this invention and the traditional method. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be 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 illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0047] Example 1

[0048] like Figure 1 The diagram shows the steps of a solver training method in one embodiment of the present invention, which mainly includes steps S1 to S3. The key steps are described below.

[0049] Step S1: Obtain multiple security-constrained unit combination models of the historical power system. For each unit combination model, generate multiple near-optimal solutions for the power resource scheduling variables in the corresponding unit combination model using a near-optimal solution generation method.

[0050] Since training neural networks requires a large amount of training data, while the amount of historical data is relatively limited, and a safety-constrained unit combination model only has one optimal solution in the historical data, the lack of training data will lead to poor training results for the neural network. Therefore, this step mainly aims to augment the training data by generating multiple near-optimal solutions for each safety-constrained unit combination model, thereby expanding the training data.

[0051] The process involves generating multiple near-optimal solutions for the power resource scheduling variables in the corresponding unit combination model using a near-optimal solution generation method, including steps S11 to S14. The key steps are described below.

[0052] Step S11: Obtain the optimal solution X of the unit combination model. * and the objective function value z corresponding to the optimal solution * Set the number of near-optimal solutions N. solGiven a tolerance of q and a value greater than 1, near-optimal constraints are added to the unit combination model to obtain a near-optimal solution generation model. The near-optimal constraint is that the objective function of the unit combination model takes a value no greater than (1+q)z. * Initialize the solution set S = {X} *}, initialize the latest solution X new =X * .

[0053] The optimal solution X of the unit combination model can be directly obtained from historical data. * It can also obtain or calculate the function value z of the objective function when the unit combination model takes the optimal solution. * .

[0054] The following presents a conventional model for unit combination, but it is not limited to this:

[0055]

[0056] In the formula, u g,t ,s g,t ,d g,t Let p represent the state variables, start-up variables, and shutdown variables of unit g at time t, respectively. g,t Let g represent the output level of unit g at time t, g(·) be the system-level constraint of the unit combination model, h(·) be the single-unit constraint of the unit combination model, and f(·) be the objective function.

[0057] Construct the solution set S and the latest solution X new During initialization, the solution set S contains only the optimal solution X. * And with the optimal solution X * As the latest solution X at the initial moment new .

[0058] Set the number of near-optimal solutions N as needed. sol For example, N sol You can take 10.

[0059] The tolerance q is set as needed, reflecting the degree of deviation of the augmented data. A larger tolerance q results in a faster generation speed but a lower quality near-optimal solution, while a smaller tolerance q results in a slower generation speed but a higher quality near-optimal solution. In this embodiment, q = 0.5%.

[0060] Adding near-optimal constraints to the unit combination model yields a near-optimal solution generation model. The near-optimal constraint is that the objective function of the unit combination model takes a value no greater than (1+q)z. * The quality of the extended near-optimal solution is guaranteed by near-optimal constraints. In a specific implementation, the near-optimal constraints can be expressed as:

[0061] f(ug,t ,s g,t ,d g,t ,p g,t )≤(1+q)z *

[0062] Assuming the original constraint set of the unit combination model is set C, after adding near-optimal constraints, the constraint set of the near-optimal solution generation model can be expressed as:

[0063] C∪{f(u g,t ,s g,t ,d g,t ,p g,t )≤(1+q)z *}

[0064] Step S12: Randomize the coefficients in the objective function of the near-optimal solution generation model and add diversity constraints to the near-optimal solution generation model to obtain an updated model. The perturbation amplitude shall not exceed 10% of the original coefficient values. The diversity constraints are used to ensure that the solution to be generated is different from the current latest solution X. new .

[0065] Among these methods, the coefficients of the objective function of the near-optimal solution generation model are randomized and perturbed. For example, the objective function f(u g,t ,s g,t ,d g,t ,p g,t Each variable in the equation has a corresponding coefficient. If the objective function is a cost function, the coefficient is actually the unit price. The coefficients are randomly perturbed, with the perturbation amplitude not exceeding 10% of the original coefficient value. In this embodiment, the perturbation range can be 0.95 to 1.05 times the original coefficient value. The objective function is perturbed in each iteration to change the optimization direction and improve the diversity of the generated near-optimal solutions.

[0066] Add diversity constraints to the near-optimal solution generation model. These constraints are used to ensure that the solution to be generated is different from the current latest solution X. new This avoids generating the same solution repeatedly.

[0067] In one embodiment, the method targets integer variables for power resource scheduling. Compared to continuous-time variables, this method is more suitable for integer variables and provides more accurate predictions for them. Specifically, these integer variables can be, for example, generator start-up variables, shutdown variables, and status variables. Within each prediction period, the predicted value of the integer variable remains constant.

[0068] Furthermore, this integer variable is a binary variable, with only two possible values: 0 and 1. For example, for the power-on variable, a value of 1 indicates that the power-on action has been performed, while a value of 0 indicates that the power-on action has not been performed. In this case, the multi-style constraint can be represented as:

[0069]

[0070] In the formula, i is the variable number, B is the set of variables, and x i Let i be the i-th variable to be decided. The latest solution for X new The solution corresponding to the i-th variable. x represents i Let X be the i-th variable in the variable set B and its latest solution X new The value in is 0. x represents i Let X be the i-th variable in the variable set B and its latest solution X new The value in is 1.

[0071] Step S13: Solve and update the model, and apply the latest solution X new Update to the current solution and add the current solution to the solution set S.

[0072] Specifically, in solving and updating the model, once a solution is found, it is stored as a near-optimal solution in the solution set S. At this time, the number of solutions stored in the solution set S is incremented by 1, and the latest solution X is also added. new Update to the current solution.

[0073] Step S14: Determine whether the number of solutions in solution set S reaches N. sol If not, proceed to step S12; if yes, output the solution set S as multiple near-optimal solutions for the corresponding unit combination model.

[0074] Specifically, repeat steps S12 to S13 until the number of solutions in solution set S meets the requirement. At this point, a unit combination model expands from a corresponding set of solutions to multiple corresponding sets of solutions, which can be specifically represented as:

[0075]

[0076] In the formula, s q For the q-th unit combination model, This is the p-th near-optimal solution of the q-th unit combination model.

[0077] like Figure 2 The diagram shown illustrates the effect of a near-optimal solution generation method provided in an embodiment of the present invention. For each unit combination model, 25 near-optimal solutions can be generated quickly in a short time using the above-mentioned near-optimal solution generation method, which greatly reduces the time spent compared to finding the optimal solution.

[0078] Step S2: Construct a training sample set. The input of each training sample is the parameter features of the unit combination model, and the label is the near-optimal solution of the corresponding unit combination model.

[0079] Specifically, a unit combination model and one of its solutions can be used to construct a training sample. Multiple training samples can be constructed from a single unit combination model. Parameter features of the unit combination model are extracted. For example, when the predicted variable is the unit's start-up and shutdown state, the input parameter features are preferably parameters closely related to the unit's start-up and shutdown state, such as parameters related to the maximum / minimum technical output of the thermal power unit and the load power curve. Using these parameter features as input and the near-optimal solution of the unit combination model as a label, a training sample can be constructed. This method is used to build a training sample set.

[0080] Step S3: Train a graph neural network using the training sample set to obtain a solver for predicting the values ​​of power resource scheduling variables in the unit combination model based on the parameter features of the unit combination model.

[0081] Specifically, graph neural networks can choose existing neural networks, such as graph convolutional neural networks.

[0082] The parameter features of the unit combination model in the training samples are input into the graph neural network. The deviation between the prediction result and the label is calculated through the loss function and the parameters of the graph neural network are adjusted through backpropagation until the loss function converges and the training ends. The result is a solver that predicts the values ​​of power resource scheduling variables in the unit combination model based on the parameter features of the unit combination model.

[0083] Taking binary variables as an example, the loss function can be designed as follows:

[0084]

[0085] in:

[0086]

[0087] U1={x i |x i ∈X,x i =0}

[0088] U2={x i |x i ∈X,x i =1}

[0089] In the formula, S is the solution set of a unit combination model, X is a set of solutions in the solution set, and x i Let P be the solution for the i-th variable in X, where U1 represents the set of variables in the solution set that take the value 0, U2 represents the set of variables in the solution set that take the value 1, and P is the solution for the i-th variable in X. i For the variable x predicted by the graph neural network i The value of is determined to be in the range [0,1] after passing through the sigmoid activation function.

[0090] Using the above training methods, a solver for solving power system security-constrained unit combination models can be trained. By inputting the model parameters of the power system security-constrained unit combination model to be solved into the solver, the values ​​of the power resource scheduling variables in the corresponding unit combination model can be obtained. This enables rapid solution of the fast model.

[0091] Example 2

[0092] This invention also proposes a solution method for a power system safety-constrained unit combination model, such as... Figure 3 As shown, the method includes:

[0093] Step S4: Determine the unit combination model of the power system to be solved for security constraints, which includes power resource scheduling variables;

[0094] Step S5: Extract the parameter features of the safety-constrained unit combination model and input them into the solver. Use the solver to predict the values ​​of power resource scheduling variables.

[0095] The solver is trained using the training method described in Example 1.

[0096] Based on steps S4 and S5, the unit combination model can be solved, and the values ​​of the power resource scheduling variables can be obtained. Based on these values, the scheduling of power resources can be realized.

[0097] Considering that the solver's prediction results may be biased, in one embodiment, after the solver continues to make predictions, the prediction results are further corrected; that is, after step S5, the following is also performed:

[0098] Step S6: Correct the prediction results.

[0099] like Figure 4 As shown, the correction includes the following steps.

[0100] Step S61: Classify variables whose predicted values ​​reach a preset confidence level as reliable variables.

[0101] The solver is actually a trained neural network. Understandably, the output of the neural network has multiple channels. The i-th channel corresponds to the predicted value and confidence level of the i-th variable. The confidence level can be used to evaluate the reliability of the predicted value of the variable.

[0102] For example, for binary variables, a unified confidence standard can be established, where the confidence level of all channel outputs is set to 1. If the confidence level of 1 is very low, it indicates that the predicted value is 0. More specifically, a first confidence threshold α1 and a second confidence threshold α2 can be set. The first confidence threshold α1 is greater than or equal to 0.8, and the second confidence threshold α2 is less than or equal to 0.2. For example, the first confidence threshold α1 can be set to 0.9, and the second confidence threshold α2 can be set to 0.1. Variables with a predicted value confidence level greater than or equal to the first confidence threshold α1 can be categorized into the first set; variables with a predicted value confidence level less than or equal to the second confidence threshold α2 can be categorized into the second set; and the remaining variables can be categorized into the third set, as shown below:

[0103]

[0104] In the formula, J is the set of indices for all variables, and h i For variable x i With a confidence level of 1, J1 represents the set of variable numbers whose predicted value is 1 and whose prediction is reliable, J0 represents the set of variable numbers whose predicted value is 0 and whose prediction is reliable, J represents the set of variable numbers whose prediction is unreliable, and J0∪J1 represents the set of variable numbers whose prediction is reliable (i.e., reliable variables).

[0105] Step S62: Add a deviation term to the objective function in the unit combination model to obtain a modified model. The deviation term = W * cumulative deviation, where the cumulative deviation is the sum of the absolute deviations between all the confidence variables to be decided and the predicted values ​​of the corresponding confidence variables, and W is a preset weight coefficient.

[0106] Continuing with the example above, the deviation term can now be expressed as:

[0107]

[0108] Adjusting the objective function in the original unit combination model yields the following modified model:

[0109]

[0110] Step S63: Solve the modified model to obtain the values ​​of the power resource scheduling variables.

[0111] It should be noted that the weighting coefficient W can be adjusted flexibly. For example, W can be a large number so that the bias term has a greater weight than the original objective function.

[0112] In some embodiments, considering that the obtained solution may still be of low quality, in order to further improve the quality of the solution, after step S6, the following is further performed:

[0113] Step S7: Evaluate the quality of the solution.

[0114] like Figure 5 As shown, evaluating the quality of the solution includes the following steps.

[0115] Step S71: Calculate the sum of the deviations between the solution results of all variables corresponding to any device in the power system in the modified model and their predicted values ​​in the solver, and obtain the decision deviation of the corresponding device.

[0116] For example, for a generator set, which includes start-up variables, shutdown variables, and state variables, the solution results of each variable in the modified model and the predicted value obtained by the solver can be obtained using the method described above. The deviation between the solution results and the predicted values ​​of each variable of the equipment can be calculated and accumulated to obtain the decision bias of the equipment.

[0117] For example, the decision bias of device g can be specifically expressed as:

[0118]

[0119] In the formula, x j Let j be the solution result obtained by solving the modified model for the j-th variable of device g. Let be the predicted value of the j-th variable of device g in the solver.

[0120] Step S72: Classify the equipment in the power system into stable equipment and unstable equipment based on the decision bias.

[0121] Step S73: Substitute the solution results of the variables corresponding to the stable equipment directly into the unit combination model, and solve the unit combination model again to obtain the values ​​of the variables corresponding to the unstable equipment.

[0122] Specifically, the larger m(g) is, the more unstable the value of the variable corresponding to the device is, and the more it needs to be re-optimized. In this case, the device with a decision deviation greater than the threshold can be re-solved. The solution process is as described above and will not be repeated here.

[0123] like Figure 6 The figure shows a time comparison between the method of this invention and the traditional method. The numbers outside the circle represent the unit combination model number, and the divergence length along the radius from the center of the circle is the solution time. As can be seen from the figure, the solution time using the solution method proposed in this invention is greatly reduced. Figure 6 The selected model is a relatively simple unit combination model. The unit combination models used in actual engineering are very large. Using this invention to solve the problem has a more obvious advantage in terms of timeliness.

[0124] Example 3

[0125] The present invention also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0126] Specifically, the memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0127] Example 4

[0128] This invention provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method described in the above embodiments of this invention.

[0129] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" are intended to illustrate the present invention and are not intended to limit the present invention.

[0130] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A solver training method for a power system safety-constrained unit combination model, characterized in that, include: Multiple safety-constrained unit combination models of historical power systems are obtained. For each unit combination model, multiple near-optimal solutions for the power resource scheduling variables in the corresponding unit combination model are generated through a near-optimal solution generation method. Construct a training sample set, where the input of each training sample is the parameter features of the unit combination model and the label is the near-optimal solution of the corresponding unit combination model; Using the training sample set, a graph neural network is trained to obtain a solver that predicts the values ​​of power resource scheduling variables in the unit combination model based on the parameter features of the unit combination model. The near-optimal solution generation method includes: Step S11: Obtain the optimal solution of the unit combination model. X and the objective function value corresponding to the optimal solution Set the number of near-optimal solutions and tolerance q Near-optimal constraints are added to the unit combination model to obtain a near-optimal solution generation model. The near-optimal constraints are that the objective function value of the unit combination model does not exceed a certain value. Initialize the solution set Initialize the latest solution ; Step S12: Randomize the coefficients in the objective function of the near-optimal solution generation model and add diversity constraints to the near-optimal solution generation model to obtain an updated model. The perturbation amplitude does not exceed 10% of the original coefficient values. The diversity constraints are used to ensure that the solution to be generated is different from the latest solution. ; Step S13: Solve and update the model, and apply the latest solution. Update to the currently obtained solution and add the currently obtained solution to the solution set. S ; Step S14: Determine the solution set S Has the number of solutions reached the target? If not, proceed to step S12; if yes, output the solution set. S This serves as multiple near-optimal solutions for the corresponding unit combination model.

2. The solver training method for a power system safety-constrained unit combination model as described in claim 1, characterized in that, The variable is an integer variable.

3. The solver training method for a power system safety-constrained unit combination model as described in claim 1, characterized in that, The variable is a binary variable; In step S12, the diversity constraint is: ； In the formula, i Number the variables. B For a set of variables, The first to be decided i One variable, The latest solution The corresponding number in the middle i Solution for each variable express For a set of variables B The first in i There are variables and they are in the latest solution. The value in is 0. express For a set of variables B The first in i There are variables and they are in the latest solution. The value in is 1.

4. The solver training method for a power system security-constrained unit combination model as described in claim 2 or 3, characterized in that, The variables include the unit's start-up variables, shutdown variables, and status variables.

5. The solver training method for a power system safety-constrained unit combination model as described in claim 1, characterized in that, The tolerance q No more than 10%.

6. A method for solving a power system safety-constrained unit combination model, characterized in that, include: A power system unit combination model with security constraints to be solved is determined, wherein the unit combination model has power resource scheduling variables to be decided; Extract the parameter features of the unit combination model and input them into the solver, then use the solver to predict the values ​​of the power resource scheduling variables; The solver is trained by the training method described in any one of claims 1 to 5.

7. The solution method for the power system security constraint unit combination model as described in claim 6, characterized in that, After using the solver to predict the values ​​of the power resource scheduling variables, the method also includes correcting the prediction results: Variables whose predicted values ​​reach a preset confidence level are classified as reliable variables. A deviation term is added to the objective function of the unit combination model to obtain a modified model, wherein the deviation term is equal to a preset weighting coefficient. Cumulative bias, which is the sum of the absolute deviations between all the confidence variables to be decided and the predicted values ​​of the corresponding confidence variables; The modified model is solved to obtain the values ​​of the power resource scheduling variables.

8. The solution method for the power system security constraint unit combination model as described in claim 7, characterized in that, After making corrections, a solution quality assessment step is also included: The decision bias of the corresponding device is obtained by calculating the sum of the deviations between the solution results of all variables corresponding to any device in the power system in the modified model and their predicted values ​​in the solver. Based on decision bias, equipment in the power system is classified into stable equipment and unstable equipment; The solution results for the variables corresponding to the stable equipment are directly substituted into the unit combination model, and the unit combination model is solved again to obtain the values ​​of the variables corresponding to the unstable equipment.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method as described in any one of claims 1 to 8.

Images