Large language model driven unit commitment optimization method

Abstract

A large language model-driven unit combination optimization method, relating to the field of power system technology, includes: acquiring basic economic operation data of the power system; constructing and solving a unit combination model with integer relaxed security constraints based on the basic economic operation data; obtaining and solving the unit combination model; constructing variable dimensionality reduction algorithms, constraint dimensionality reduction algorithms, and infeasibility solution repair algorithms; using a large language model to perform single-algorithm evolution on the variable and constraint dimensionality reduction algorithms; using a large language model to perform single-algorithm evolution on the infeasibility solution repair algorithm; solving and calculating the unit combination model with integer relaxed security constraints; calculating the variable constraint dimensionality reduction security constraint unit combination model and obtaining the optimal solution; and using the optimal infeasibility solution repair algorithm to repair the optimal solution of the variable constraint dimensionality reduction security constraint unit combination model, obtaining the result. This method addresses the shortcomings of existing model dimensionality reduction methods in terms of generality, stability, and interpretability.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and more specifically to a unit combination optimization method driven by a large language model. Background Technology

[0002] Security-constrained unit combination (SCUC) is the mathematical foundation for problems in power system dispatching and operation, market clearing, and system planning. This problem is typically modeled as a mixed-integer programming (MIP) problem. With the increasing proportion of new energy sources in the power system and the continuous expansion of the interconnected grid, system decisions are becoming increasingly complex in both time and space dimensions. This leads to a sharp increase in the number of constraints and the degree of coupling in the unit combination problem, significantly increasing the difficulty of solving it. Among these, the large number of binary variables and complex coupling constraints are the main reasons for the difficulty in solving this problem.

[0003] Existing methods for dimensionality reduction of unit combination models can be mainly divided into two categories: heuristic methods and data-driven methods.

[0004] Heuristic algorithms are typically based on empirical observations of problem structure and mainly fall into three categories:

[0005] Ineffective constraint identification method: By using methods such as auxiliary problem optimization, extreme scenario evaluation or security check iteration, redundant line security constraints that have no impact on the optimal solution are identified and eliminated, thereby reducing the constraint size.

[0006] Diving algorithm: Starting from the current optimal solution or the root relaxed solution, it reduces the number of integer variables that need to be optimized by fixing the values ​​of some binary variables, thereby achieving variable reduction.

[0007] Model approximation method: By clustering similar units or aggregating time periods, the size of variables and constraints can be reduced simultaneously.

[0008] While these heuristics can deliver significant speedups in specific scenarios, their design heavily relies on expert domain experience and suffers from insufficient performance stability across different operating conditions. This strong dependence on manual design and scenario-specific adjustments limits their widespread application in complex and variable power systems.

[0009] Data-driven methods are another important research direction. They leverage the similarity of unit states across different operating days to train models (such as graph neural networks and long short-term memory networks) based on historical data to perform neural diving or constraint screening, thereby reducing variables or constraints. However, these methods also face application bottlenecks: First, most data-driven models are "black box" characteristics, lacking interpretability in the decision-making process and failing to meet the stringent requirements of power systems for transparency in operational decisions; second, when faced with new scenarios beyond the distribution of training data, the model's generalization ability is often poor, and performance degrades significantly; finally, constructing large-scale, high-quality training datasets requires extensive feature engineering and scene annotation, which imposes a heavy computational burden on the preprocessing stage.

[0010] Therefore, we propose a unit combination optimization method that can adaptively evolve, reduce reliance on human experience, and maintain high performance in various scenarios. Summary of the Invention

[0011] The purpose of this invention is to provide a large language model-driven unit combination optimization method, which addresses the shortcomings of existing model dimensionality reduction methods in terms of universality, stability, and interpretability.

[0012] This invention is achieved through the following technical solution:

[0013] Large language model-driven unit combination optimization methods specifically include:

[0014] To obtain basic data on the economic operation of the power system;

[0015] A unit combination model with integer relaxation safety constraints is constructed based on basic economic operation data, and then solved.

[0016] The unit combination model is obtained based on the unit combination model with integer relaxed safety constraints, and then solved.

[0017] Construct seed algorithms that include variable dimensionality reduction algorithm, constraint dimensionality reduction algorithm, and infeasible solution repair algorithm;

[0018] A large language model is used to perform single-algorithm evolution on the variable dimensionality reduction algorithm and the constraint dimensionality reduction algorithm in turn to obtain the optimal variable dimensionality reduction algorithm and the optimal constraint dimensionality reduction algorithm.

[0019] Based on the optimal constraint dimensionality reduction algorithm, a large language model is used to perform single-algorithm evolution on the infeasible solution repair algorithm to obtain the optimal infeasible solution repair algorithm.

[0020] Based on the optimal variable dimensionality reduction algorithm and the optimal constraint dimensionality reduction algorithm, the unit combination model after integer relaxed safety constraints is solved and calculated to obtain the variable constraint dimensionality reduction safety constraint unit combination model.

[0021] A mixed-integer linear programming solver is used to calculate the variable-constrained, dimension-reduced safety-constrained unit combination model, and the optimal solution of the variable-constrained, dimension-reduced safety-constrained unit combination model is obtained.

[0022] Based on the optimal infeasible solution repair algorithm, the optimal solution of the safety constraint unit combination model with variable constraint dimensionality reduction is repaired to obtain the solution that satisfies the safety constraint unit combination model, which is then used as the result of the safety constraint unit combination model.

[0023] Furthermore, the basic economic operation data of the power system includes: upper limit of unit output, lower limit of unit output, upper limit of unit ramp-up / ramp rate, minimum continuous start-up / shutdown time of units, unit operating cost function, unit start-up / shutdown cost function, power flow transfer distribution factor, line transmission capacity, and system positive and negative reserve rate.

[0024] Furthermore, the specific steps for constructing the unit combination model with integer relaxation safety constraints based on basic economic operation data are as follows:

[0025] Determine the objective function of the unit combination model:

[0026]

[0027] in, For the unit The start-stop cost function, For the unit The running cost function, For generator sets The initial start-stop state, For generator sets During the period The output variable, For generator sets During the period The start and stop slack variables, and The value can be a continuous value between 0 and 1. This refers to the number of time periods for unit combination. This represents the total number of generator sets.

[0028] Establish corresponding constraints based on the objective function:

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035]

[0036]

[0037] In the formula, For time period Net load of internal system , For the system's positive and negative reserve rates, , Time periods The maximum and minimum net load of the internal system. , Generator sets The upper and lower limits of output, , Generator sets Increase or decrease the upper limit of output; For generator sets Maximum output at startup; For generator sets Maximum output before shutdown , Generator sets Minimum continuous start-up and shutdown time; For the line The biggest trend For nodes For the line Power flow transfer distribution factor For the unit The node in question For nodes .

[0038] Furthermore, the specific steps for obtaining the unit combination model based on the unit combination model with integer relaxed safety constraints are as follows:

[0039] The decision variables of the unit combination model with integer relaxed safety constraints are adjusted to 0-1 variables, and the last power flow constraint is removed to obtain the unit combination model.

[0040] Furthermore, the construction steps of the variable dimensionality reduction algorithm are as follows:

[0041] From the results of the integer relaxation safety constraint unit combination, the largest and smallest 5% integer variables are identified, specifically:

[0042]

[0043]

[0044] In the formula, For unit start-up and shutdown variables, It is an intermediate variable for argmax(·) or argmin(·). The total number of unit combinations with relaxed safety constraints, representing an integer.

[0045] The variables with the largest values ​​are fixed at 1, and the variables with the smallest values ​​are fixed at 0. Specifically:

[0046]

[0047] .

[0048] Furthermore, the construction steps of the constrained dimensionality reduction algorithm are as follows:

[0049] Obtain the line power flow results of the unit combination after integer relaxation of safety constraints;

[0050] Based on weight We perform weighted calculations on the power flow results of the two types of routes to obtain a constrained dimensionality reduction algorithm.

[0051] Furthermore, the specific steps for single-algorithm evolution using a large language model to perform variable dimensionality reduction, constraint dimensionality reduction, and infeasible solution repair algorithms are as follows:

[0052] The code corresponding to the seed algorithm is generated using a large language model, and then run in a preset evaluation environment to obtain the output results;

[0053] Construct the evaluation functions corresponding to the seed algorithms respectively;

[0054] The code output of the corresponding seed algorithm is evaluated using any evaluation function. If the code fails the evaluation, the large language model regenerates the code and evaluates it again.

[0055] If the evaluation is passed, the code is determined to be a valid algorithm, and all valid algorithms are sorted to form an algorithm library;

[0056] Based on the algorithm library, we analyze the common characteristics of high-performance algorithms and the shortcomings of low-performance algorithms, and generate improvement suggestions and a summary of ideas.

[0057] Based on the improvement suggestions and the summary of ideas, new effective algorithms were generated and evaluated in turn;

[0058] If the algorithm fails the evaluation, a new valid algorithm is generated and evaluated again. If the algorithm passes the evaluation, the code is determined to be the optimal algorithm.

[0059] Furthermore, for the aforementioned variable dimensionality reduction algorithm, the fixed variables of the variable dimensionality reduction algorithm are compared with the solution of the safety constraint unit combination model to determine the evaluation function, specifically as follows:

[0060]

[0061]

[0062] In the formula, For variable dimensionality reduction algorithms, the number of variables is fixed. The number of variables correctly fixed by the variable dimensionality reduction algorithm (i.e., the number of variables that are consistent with the optimal solution result of the safety-constrained unit combination). The number of variables whose errors in the variable dimensionality reduction algorithm are fixed (i.e., the number of variables whose results are inconsistent with the optimal solution of the safety-constrained unit combination). It is an algorithm The indicators This is the number of test cases used to evaluate the algorithm's performance.

[0063] Furthermore, for the constraint dimensionality reduction algorithm, the fixed variables of the variable dimensionality reduction algorithm are compared with the solution of the safety constraint unit combination model to determine the evaluation function, specifically:

[0064]

[0065]

[0066]

[0067]

[0068] In the formula, Example of constrained dimensionality reduction algorithm prediction The accuracy of the constraints Example of constrained dimensionality reduction algorithm prediction Constrained recall and Examples of correct predictions using the constrained dimensionality reduction algorithm. The number of redundant and non-redundant constraints, and Examples of error prediction using constrained dimensionality reduction algorithms The number of non-redundant and redundant constraints, It is an algorithm The indicators This is the number of test cases used to evaluate the algorithm's performance.

[0069] Furthermore, for the aforementioned infeasibility solution repair algorithm, the infeasibility repair seed algorithm and the infeasibility repair algorithm are used to repair the infeasibility solutions of the unit combination model, respectively. The repair time and success rate of the two methods are compared with the objective function values ​​corresponding to the repaired feasible solutions to determine the evaluation function, specifically:

[0070]

[0071] In the formula, Repair example for infeasible repair seed algorithm The time consumed Repair examples for infeasible repair algorithms The time consumed Examples obtained for the solver The optimal value, For the repaired example The objective function value corresponding to the feasible solution. It is an algorithm The indicators This is the number of test cases used to evaluate the algorithm's performance.

[0072] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0073] This invention discloses a large language model-driven unit combination optimization method, realizing a fully automated and adaptive intelligent acceleration solution paradigm for unit combination. By using a large language model, the traditional manual algorithm design process is transformed into a data-driven automated process, and the model simplification, rapid solution and solution repair are seamlessly connected to form a closed loop. Finally, based on a large number of diverse calculation examples, evaluation and evolution are carried out, so that the final generated optimal algorithm can capture universal laws, thereby maintaining stable and excellent performance under various operating scenarios.

[0074] Furthermore, the "data" required for the evolution process of this method is readily available basic data of the power system and benchmark answers automatically generated by the solver, which greatly simplifies the preprocessing process and reduces the overall computational cost and complexity of the method implementation.

[0075] In addition, the seed algorithm effectively guides the evolution of the large language model, avoids the inefficiency of completely random search, and ensures that the automatically generated initial algorithm population has good performance, thereby significantly accelerating the convergence of the entire evolution process.

[0076] Furthermore, the evaluation function provides a clear and objective standard for the survival of the fittest in algorithm evolution, ensuring that large language models can be effectively guided to generate algorithm code that is not only syntactically correct but also performs exceptionally well. Attached Figure Description

[0077] Figure 1 This is a schematic diagram of a method flow of the present invention;

[0078] Figure 2 The flowchart shows the method for single-algorithm evolution of the three seed algorithms of this invention;

[0079] Figure 3 This is a schematic diagram of the electronic device in this invention. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0081] Example 1

[0082] like Figure 1 The large language model-driven unit combination optimization method shown specifically includes:

[0083] To obtain basic data on the economic operation of the power system;

[0084] The basic data for the economic operation of the power system include: upper limit of unit output, lower limit of unit output, upper limit of unit ramp-up / ramp rate, minimum continuous start-up / shutdown time of units, unit operating cost function, unit start-up / shutdown cost function, power flow transfer distribution factor, line transmission capacity and system positive and negative reserve rate.

[0085] The upper and lower limits of generator output refer to the power range within which a generator set can operate stably. This parameter is the most basic physical constraint of the generator set, which requires the optimization model to ensure that the output of each generator set is within this range, otherwise the equipment will be damaged or the operation will be unstable.

[0086] The upper limit of the ramp rate of the generator unit refers to the maximum rate at which the generator unit increases or decreases its output power between two adjacent time periods. It simulates the flexibility of the unit in responding to dispatch instructions. For example, a thermal power unit cannot generate electricity at full load from 0 instantaneously. This constraint ensures that the dispatch plan is feasible in time and will not exceed the actual adjustment capacity of the unit.

[0087] The minimum continuous start-up and shutdown time of a generator unit refers to the shortest time that a generator unit must run continuously once it starts up, and the shortest time that it must remain shut down once it stops. This is to prevent frequent start-ups and shutdowns of the unit, thereby reducing equipment wear and additional start-up and shutdown costs, and meeting the unit's thermodynamic requirements.

[0088] The unit operating cost function describes the mathematical relationship between the variable costs, such as fuel, consumed by the generator unit when generating electricity and its output. It is usually approximated as a quadratic function or a piecewise linear function and is the core of the economic objective of the optimization model.

[0089] The generator start-up and shutdown cost function refers to the additional costs incurred by a generator set to perform a single start-up or shutdown operation, including fuel, manpower, and equipment wear and tear. Including start-up and shutdown costs in the optimization model can prevent the creation of infeasible plans that appear to have low operating costs but involve frequent start-ups and shutdowns, making the optimization results more consistent with actual economics.

[0090] The power flow transfer distribution factor is a constant matrix that describes how much influence the injection or extraction of 1 unit of power at any node in the power grid will have on the power flowing through any line. Moreover, this factor can quickly estimate the power flow of each line under any power generation plan without directly solving complex nonlinear power flow equations, thereby determining whether the limit is exceeded.

[0091] Line transmission capacity refers to the maximum active power that each transmission line or transformer in the power grid can safely carry. It is a reflection of safety constraints. The optimization model must ensure that the calculated power generation plan will not cause the power flow of any line to exceed its capacity, otherwise it will cause safety accidents such as overheating and tripping.

[0092] The positive and negative reserve ratios of a system refer to the buffer space that the total power generation capacity of the system needs to reserve. Positive reserve is the ability to cope with a sudden increase in load (such as unit tripping), while negative reserve is the ability to cope with a sudden decrease in load. Together, they ensure that the power system has the ability to adjust to emergencies (such as load fluctuations and unit failures), and are key reliability indicators to ensure the safe and stable operation of the system.

[0093] The specific steps for constructing a unit combination model with integer relaxation safety constraints based on basic economic operation data are as follows:

[0094] Determine the objective function of the unit combination model, namely, minimize the total operating cost and start-up / shutdown cost for each time period:

[0095] (1)

[0096] in, For the unit The start-stop cost function, For the unit The running cost function, For generator sets The initial start-stop state, For generator sets During the period The output variable, For generator sets During the period The start and stop slack variables, and The value can be a continuous value between 0 and 1. This refers to the number of time periods for unit combination. This represents the total number of generator sets.

[0097] Establish corresponding constraints based on the objective function:

[0098] (2)

[0099] (3)

[0100] (4)

[0101] (5)

[0102] (6)

[0103] (7)

[0104] (8) (9) (10) (11)

[0105] (12)

[0106] In the formula, For time period Net load of internal system , For the system's positive and negative reserve rates, , Time periods The maximum and minimum net load of the internal system. , Generator sets The upper and lower limits of output, , Generator sets Increase or decrease the upper limit of output; For generator sets Maximum output at startup; For generator sets Maximum output before shutdown , Generator sets Minimum continuous start-up and shutdown time; For the line The biggest trend For nodes For the line Power flow transfer distribution factor For the unit The node in question For nodes ;

[0107] Formula (2) is the system power balance constraint, formulas (3) and (4) are the system positive and negative standby constraints, formulas (5) and (6) are the system positive and negative ramp standby constraints, formula (7) is the generator output range constraint, formulas (8) and (9) are the generator positive and negative ramp constraints, formulas (10) and (11) are the generator minimum continuous start-up and shutdown time constraints; in addition, (5)~(11) are constraints specific to thermal power units. For other possible power sources, such as hydropower units, wind power units, etc., they also need to meet their corresponding physical characteristic constraints; formula (12) is the line safety constraint.

[0108] The solution process for the unit combination model with integer relaxed safety constraints is as follows: A mixed-integer linear programming solver is used to solve the defined integer relaxed safety constraint unit combination model, yielding the unit output based on the relaxed state. and continuous unit start-up and shutdown status .

[0109] The unit combination model is obtained based on the unit combination model with integer relaxed safety constraints. The specific steps are as follows:

[0110] By adjusting the decision variables of the unit combination model with integer relaxed safety constraints to 0-1 variables and deleting the last power flow constraint, the unit combination model is obtained. Furthermore, by using a mixed-integer linear programming solver to solve the unit combination model, the results of the unit combination model can be obtained: unit active power output. That is, the specific power generation value of each operating unit in each time period, and the start-up and shutdown status of the unit. In other words, it is a precise, binary schedule for the start-up and shutdown of the unit.

[0111] A seed algorithm is constructed, comprising a variable dimensionality reduction algorithm, a constraint dimensionality reduction algorithm, and an infeasible solution repair algorithm. This seed algorithm plays a dual role as both an "initial population" and a "guidebook" in the subsequent evolutionary framework based on a large language model. Its main functions are as follows:

[0112] Providing a high-quality starting point: These are algorithms designed by human experts and possessing basic effectiveness. This avoids having large language models "guess randomly" from scratch, greatly accelerating the evolutionary process and ensuring that the initial population has a certain level of performance.

[0113] Injecting domain knowledge: The seed algorithm encodes and solidifies the experience and intuition of power system experts on how to simplify unit combination models (e.g., "variables with extreme values ​​in integer relaxation solutions are more worth fixing"), providing the correct direction for subsequent evolution.

[0114] Defining Algorithm Paradigms: The three seed algorithms clarify the three types of algorithms that the large language model needs to evolve and their input / output specifications. It tells the large language model: "What you need to generate is code that has this type of functionality, receives this type of data, and outputs this type of result."

[0115] Reducing evolutionary uncertainty: Without a seed algorithm, large language models may generate a large amount of completely invalid or logically chaotic code. The seed algorithm provides a reliable basic template, making subsequent evolutionary operations such as crossover and mutation more likely to produce meaningful improvements.

[0116] Specifically, variable dimensionality reduction algorithms are used to identify units that are "almost certain" to be in a certain state (on or off) in the optimal solution, and fix their binary variables in advance, thereby significantly reducing the number of integer variables that need to be optimized.

[0117] Constraint dimensionality reduction algorithms are used to identify line safety constraints that will definitely not be activated under the optimal solution (i.e., the line power flow is far from reaching the capacity limit) and remove them from the model.

[0118] As for infeasible solution repair algorithms, since variable dimensionality reduction algorithms and constraint dimensionality reduction algorithms are heuristic and may be error-prone, they may cause the simplified model to have no solution or the obtained solution to violate the constraints of the original model. Therefore, this repair algorithm is a "repair tool" that repairs infeasible solutions into feasible solutions.

[0119] A large language model is used to perform single-algorithm evolution on the variable dimensionality reduction algorithm and the constraint dimensionality reduction algorithm in turn to obtain the optimal variable dimensionality reduction algorithm and the optimal constraint dimensionality reduction algorithm.

[0120] Based on the optimal constraint dimensionality reduction algorithm, a large language model is used to perform single-algorithm evolution on the infeasible solution repair algorithm to obtain the optimal infeasible solution repair algorithm.

[0121] Single-algorithm evolution is the foundation of multi-algorithm evolution, and it can optimize the performance of a single algorithm, improving its performance in specific evaluation scenarios. Specifically, single-algorithm evolution of variable dimensionality reduction algorithms utilizes the optimal variable dimensionality reduction algorithm obtained through evolution to analyze the integer relaxation solutions of the current problem. It intelligently identifies a batch of unit start-up and shutdown variables that can be safely fixed (fixed to 0 or 1), which directly reduces the number of integer variables in the mixed integer programming problem, fundamentally reducing the combinatorial complexity of the problem.

[0122] Constraint dimensionality reduction algorithms utilize the optimal constraint dimensionality reduction algorithm derived through evolution to analyze power flow, identify a large number of redundant line safety constraints, and remove them. This reduces the computational burden on the solver, freeing it from searching for the optimal solution among countless irrelevant constraints.

[0123] The optimal variable dimensionality reduction algorithm and the optimal constraint dimensionality reduction algorithm are used to solve the unit combination model after relaxing the integer safety constraints, and obtain the variable constraint dimensionality reduction safety constraint unit combination model. This model is a simplified model optimized by the algorithm, which effectively reduces the amount of computation in subsequent calculations and improves the computational efficiency.

[0124] A mixed-integer linear programming solver is used to calculate the variable-constrained, dimension-reduced safety-constrained unit combination model, and the optimal solution of the variable-constrained, dimension-reduced safety-constrained unit combination model is obtained.

[0125] Because the model size has been greatly reduced through algorithm evolution, mixed integer programming problems that were originally difficult to solve directly are now very easy for commercial solvers to handle. Therefore, the solver can efficiently find the optimal solution on this simplified model and quickly obtain an optimal solution, which is the true optimal solution under the simplified model.

[0126] Based on the optimal infeasible solution repair algorithm, the optimal solution of the safety constraint unit combination model with variable constraint dimensionality reduction is repaired to obtain the solution that satisfies the safety constraint unit combination model, which is used as the result of the safety constraint unit combination model.

[0127] The algorithm utilizes an optimal infeasible solution repair algorithm derived from evolution. This algorithm is trained to efficiently diagnose the reasons for the infeasibility of a solution (such as which line exceeds the power flow limit) and generates a feasible solution that satisfies all original constraints by strategically adjusting the unit state (e.g., reopening the fixed state of certain key units), thus balancing the quality (objective function value) and feasibility of the solution.

[0128] Example 2

[0129] As one embodiment, the variable dimensionality reduction algorithm is constructed based on human expert experience, and the specific steps are as follows:

[0130] From the results of the integer relaxation safety constraint unit combination, the largest and smallest 5% integer variables are identified, specifically:

[0131]

[0132]

[0133] In the formula, For unit start-up and shutdown variables, It is an intermediate variable for argmax(·) or argmin(·). The total number of unit combinations with relaxed safety constraints, representing an integer.

[0134] The variables with the largest values ​​are fixed at 1, and the variables with the smallest values ​​are fixed at 0. Specifically:

[0135]

[0136] .

[0137] Furthermore, the constrained dimensionality reduction algorithm is constructed based on the experience of human experts, and the specific steps are as follows:

[0138] Obtain the line power flow results of the unit combination after integer relaxation of safety constraints;

[0139] Based on weight We perform a weighted calculation on the power flow results of the two routes to obtain a constrained dimensionality reduction algorithm, the specific formula of which is:

[0140]

[0141]

[0142] in The line power flow results for the unit combination after integer relaxation of safety constraints are given. The line power flow results for unit combination, This is a weighted average of the power flow results for the two routes. The line power flow result is a safety constraint for the unit combination; if the estimated result does not exceed the corresponding line power flow upper limit, the corresponding line safety constraint is considered a redundant constraint.

[0143] As needed, the infeasible solution repair algorithm is designed based on human expert experience. Specifically, based on the power transfer factor (PTDF) matrix of the violated line constraints, the units that contribute to the violation of line power flow are sorted from largest to smallest, and the corresponding unit start-up and shutdown variables are relaxed in turn. The partially relaxed unit combination problem is then solved again. If the solution still cannot satisfy all line safety constraints, the above operation is repeated until a feasible solution is obtained.

[0144] Example 3

[0145] As one example, such as Figure 2 As shown, the specific steps for single-algorithm evolution using a large language model to perform variable dimensionality reduction, constraint dimensionality reduction, and infeasible solution repair algorithms are as follows:

[0146] The code corresponding to the seed algorithm is generated using a large language model, and then run in a preset evaluation environment to obtain the output results;

[0147] The process of generating code using a large language model involves using pre-defined prompt words and seed algorithm code to generate a batch of code corresponding to the algorithm using a large code generation model.

[0148] Construct the evaluation functions corresponding to the seed algorithms respectively;

[0149] The code output of the corresponding seed algorithm is evaluated using any evaluation function. Each seed algorithm corresponds to a batch of initial code and an evaluation function. During evaluation, the output of the initial code corresponding to a certain seed algorithm is input into the evaluation function of that seed algorithm. The evaluation function calculates the evaluation index and then determines whether the multiple evaluation indexes can pass.

[0150] If the evaluation fails, the code will be regenerated by the large language model and re-evaluated.

[0151] If the evaluation is passed, the code is determined to be a valid algorithm, and all valid algorithms are sorted to form an algorithm library;

[0152] Based on the algorithm library, we analyze the common characteristics of high-performance algorithms and the shortcomings of low-performance algorithms, and generate improvement suggestions and a summary of ideas.

[0153] Based on the improvement suggestions and the summary of ideas, new effective algorithms were generated and evaluated in turn;

[0154] The generation of improvement suggestions involves randomly selecting a subset of algorithm pairs from the algorithm library. Based on each algorithm pair and pre-defined prompts, a large model is generated using the prompts to process each pair. The characteristics of the best-performing algorithms are then compared and summarized to arrive at the final result. The summary of ideas, on the other hand, is based on pre-defined prompts to summarize the new and effective algorithms and improvement suggestions generated from the improvement suggestions, thereby summarizing the ideas for algorithm improvement that lead to enhanced algorithm performance.

[0155] In other words, the generation of new effective algorithms based on improvement suggestions and the summary of ideas is a progressive process. First, new effective algorithms are generated based on improvement suggestions, then evaluated, then the ideas are summarized, and finally, new effective algorithms are generated again based on the summarized ideas and evaluated. This process yields the optimal effective algorithm, which specifically includes the optimal variable dimensionality reduction algorithm, the optimal constraint dimensionality reduction algorithm, and the optimal infeasible solution repair algorithm.

[0156] If the algorithm fails the evaluation, a new valid algorithm is generated and evaluated again. If the algorithm passes the evaluation, the code is determined to be the optimal algorithm.

[0157] Furthermore, for the aforementioned variable dimensionality reduction algorithm, the fixed variables of the variable dimensionality reduction algorithm are compared with the solutions of the safety-constrained unit combination model to determine the evaluation function, specifically as follows:

[0158]

[0159]

[0160] In the formula, For variable dimensionality reduction algorithms, the number of variables is fixed. The number of variables correctly fixed by the variable dimensionality reduction algorithm (i.e., the number of variables that are consistent with the optimal solution result of the safety-constrained unit combination). The number of variables whose errors in the variable dimensionality reduction algorithm are fixed (i.e., the number of variables whose results are inconsistent with the optimal solution of the safety-constrained unit combination). It is an algorithm The indicators This is the number of test cases used to evaluate the algorithm's performance.

[0161] For the constraint dimensionality reduction algorithm, the fixed variables of the variable dimensionality reduction algorithm are compared with the solution of the safety constraint unit combination model to determine the evaluation function, specifically:

[0162]

[0163]

[0164]

[0165]

[0166] In the formula, Example of constrained dimensionality reduction algorithm prediction The accuracy of the constraints Example of constrained dimensionality reduction algorithm prediction Constrained recall and Examples of correct predictions using the constrained dimensionality reduction algorithm. The number of redundant and non-redundant constraints, and Examples of error prediction using constrained dimensionality reduction algorithms The number of non-redundant and redundant constraints, It is an algorithm The indicators This is the number of test cases used to evaluate the algorithm's performance.

[0167] For the aforementioned infeasibility solution repair algorithm, the infeasibility repair seed algorithm and the infeasibility repair algorithm are used to repair infeasible solutions of the unit combination model, respectively. The repair time and success rate of the two methods are compared with the objective function value corresponding to the repaired feasible solution to determine the evaluation function, specifically as follows:

[0168]

[0169] In the formula, Repair example for infeasible repair seed algorithm The time consumed Repair examples for infeasible repair algorithms The time consumed Examples obtained for the solver The optimal value, For the repaired example The objective function value corresponding to the feasible solution. It is an algorithm The indicators This is the number of computational examples used to evaluate the algorithm's performance;

[0170] In addition, for the three evaluation functions, the lower the evaluation index in the evaluation results, the better the algorithm performance. In other words, the standard for passing the evaluation is that the evaluation index calculated by the evaluation function is lower than the set threshold.

[0171] Example 4

[0172] As attached Figure 3 An electronic device shown includes:

[0173] Processor, memory, communication interface;

[0174] The memory is used to store the executable instructions of the processor;

[0175] The processor is configured to execute the above-described large language model-driven unit combination optimization method by executing the executable instructions.

[0176] A readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described large language model-driven unit combination optimization method.

[0177] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A large language model-driven unit combination optimization method, characterized in that, Specifically, it includes: To obtain basic data on the economic operation of the power system; A unit combination model with integer relaxation safety constraints is constructed based on basic economic operation data, and then solved. The unit combination model is obtained based on the unit combination model with integer relaxed safety constraints, and then solved. Construct seed algorithms that include variable dimensionality reduction algorithm, constraint dimensionality reduction algorithm, and infeasible solution repair algorithm; A large language model is used to perform single-algorithm evolution on the variable dimensionality reduction algorithm and the constraint dimensionality reduction algorithm in turn to obtain the optimal variable dimensionality reduction algorithm and the optimal constraint dimensionality reduction algorithm. Based on the optimal constraint dimensionality reduction algorithm, a large language model is used to perform single-algorithm evolution on the infeasible solution repair algorithm to obtain the optimal infeasible solution repair algorithm. Based on the optimal variable dimensionality reduction algorithm and the optimal constraint dimensionality reduction algorithm, the unit combination model after integer relaxed safety constraints is solved and calculated to obtain the variable constraint dimensionality reduction safety constraint unit combination model. A mixed-integer linear programming solver is used to calculate the variable-constrained, dimension-reduced safety-constrained unit combination model, and the optimal solution of the variable-constrained, dimension-reduced safety-constrained unit combination model is obtained. Based on the optimal infeasible solution repair algorithm, the optimal solution of the safety constraint unit combination model with variable constraint dimensionality reduction is repaired to obtain the solution that satisfies the safety constraint unit combination model, which is then used as the result of the safety constraint unit combination model.

2. The unit combination optimization method driven by a large language model according to claim 1, characterized in that: The basic economic operation data of the power system includes: upper limit of unit output, lower limit of unit output, upper limit of unit ramp-up / ramp rate, minimum continuous start-up / shutdown time of units, unit operating cost function, unit start-up / shutdown cost function, power flow transfer distribution factor, line transmission capacity, and system positive and negative reserve rate.

3. The unit combination optimization method driven by a large language model according to claim 2, characterized in that: The specific steps for constructing the unit combination model with integer relaxation safety constraints based on basic economic operation data are as follows: Determine the objective function of the unit combination model: in, For the unit The start-stop cost function, For the unit The running cost function, For generator sets The initial start-stop state, For generator sets During the period The output variable, For generator sets During the period The start and stop slack variables, and The value can be a continuous value between 0 and 1. This refers to the number of time periods for unit combination. This represents the total number of generator sets. Establish corresponding constraints based on the objective function: In the formula, For time period Net load of internal system , For the system's positive and negative reserve rates, , Time periods The maximum and minimum net load of the internal system. , Generator sets The upper and lower limits of output, , Generator sets Increase or decrease the upper limit of output; For generator sets Maximum output at startup; For generator sets Maximum output before shutdown , Generator sets Minimum continuous start-up and shutdown time; For the line The biggest trend For nodes For the line Power flow transfer distribution factor For the unit The node in question For nodes .

4. The unit combination optimization method driven by a large language model according to claim 3, characterized in that: The specific steps for obtaining the unit combination model based on the unit combination model with integer relaxed safety constraints are as follows: The decision variables of the unit combination model with integer relaxed safety constraints are adjusted to 0-1 variables, and the last power flow constraint is removed to obtain the unit combination model.

5. The unit combination optimization method driven by a large language model according to claim 1, characterized in that: The construction steps of the variable dimensionality reduction algorithm are as follows: From the results of the integer relaxation safety constraint unit combination, the largest and smallest 5% integer variables are identified, specifically: In the formula, For unit start-up and shutdown variables, It is an intermediate variable for argmax(·) or argmin(·). The total number of unit combinations with relaxed safety constraints, representing an integer. The variables with the largest values ​​are fixed at 1, and the variables with the smallest values ​​are fixed at 0. Specifically: 。 6. The unit combination optimization method driven by a large language model according to claim 1, characterized in that: The construction steps of the constrained dimensionality reduction algorithm are as follows: Obtain the line power flow results of the unit combination after integer relaxation of safety constraints; Based on weight We perform weighted calculations on the power flow results of the two types of routes to obtain a constrained dimensionality reduction algorithm.

7. The unit combination optimization method driven by a large language model according to claim 1, characterized in that: The specific steps for single-algorithm evolution using a large language model for variable dimensionality reduction, constraint dimensionality reduction, and infeasible solution repair algorithms are as follows: The code corresponding to the seed algorithm is generated using a large language model, and then run in a preset evaluation environment to obtain the output results; Construct the evaluation functions corresponding to the seed algorithms respectively; The code output of the corresponding seed algorithm is evaluated using any evaluation function. If the code fails the evaluation, the large language model regenerates the code and evaluates it again. If the evaluation is passed, the code is determined to be a valid algorithm, and all valid algorithms are sorted to form an algorithm library; Based on the algorithm library, we analyze the common characteristics of high-performance algorithms and the shortcomings of low-performance algorithms, and generate improvement suggestions and a summary of ideas. Based on the improvement suggestions and the summary of ideas, new effective algorithms were generated and evaluated in turn; If the algorithm fails the evaluation, a new valid algorithm is generated and evaluated again. If the algorithm passes the evaluation, the code is determined to be the optimal algorithm.

8. The unit combination optimization method driven by a large language model according to claim 7, characterized in that: For the aforementioned variable dimensionality reduction algorithm, the fixed variables of the variable dimensionality reduction algorithm are compared with the solution of the safety constraint unit combination model to determine the evaluation function, specifically as follows: In the formula, For variable dimensionality reduction algorithms, the number of variables is fixed. The number of variables that are correctly fixed for the dimensionality reduction algorithm. The number of variables is fixed to prevent errors in the dimensionality reduction algorithm. It is an algorithm The indicators This is the number of test cases used to evaluate the algorithm's performance.

9. The unit combination optimization method driven by a large language model according to claim 7, characterized in that: For the constraint dimensionality reduction algorithm, the fixed variables of the variable dimensionality reduction algorithm are compared with the solution of the safety constraint unit combination model to determine the evaluation function, specifically: In the formula, Example of constrained dimensionality reduction algorithm prediction The accuracy of the constraints Example of constrained dimensionality reduction algorithm prediction Constrained recall and Examples of correct predictions using the constrained dimensionality reduction algorithm. The number of redundant and non-redundant constraints, and Examples of error prediction using constrained dimensionality reduction algorithms The number of non-redundant and redundant constraints, It is an algorithm The indicators This is the number of test cases used to evaluate the algorithm's performance.

10. The unit combination optimization method driven by a large language model according to claim 7, characterized in that: For the aforementioned infeasibility solution repair algorithm, the infeasibility repair seed algorithm and the infeasibility repair algorithm are used to repair infeasible solutions of the unit combination model, respectively. The repair time and success rate of the two methods are compared with the objective function value corresponding to the repaired feasible solution to determine the evaluation function, specifically as follows: In the formula, Repair example for infeasible repair seed algorithm The time consumed Repair examples for infeasible repair algorithms The time consumed Examples obtained for the solver The optimal value, For the repaired example The objective function value corresponding to the feasible solution. It is an algorithm The indicators This is the number of test cases used to evaluate the algorithm's performance.

Images