Machine learning enhanced unit commitment problem optimization solving method and system and storage medium
By constructing input feature space and output tag feature space, using machine learning algorithms to train classifiers and perform probability fusion, optimize the unit combination problem model, the problem of insufficient feature space mining in UC problems is solved, and prediction accuracy and computing efficiency are improved.
Patent Information
- Application Number
- CN202510656675.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-21
- Publication Date
- 2025-09-05
AI Technical Summary
The existing technology has insufficient space mining of artificial intelligence feature in the unit combination problem, difficult to improve prediction accuracy, and difficult to ensure the accuracy of variable prediction, resulting in computing efficiency and safety issues.
By constructing the input feature space and the output tag-like feature space, using machine learning algorithms to train the classifier, combining the probability fusion method to predict the effectiveness of constraints and the value of decision variables, the machine learning enhancement method is used to optimize the unit combination problem model.
It improves the prediction accuracy and calculation efficiency of unit combination problems, ensures feasibility and optimization of understanding, reduces redundant constraint calculations, and improves the efficiency and robustness of solving UC problems.
Smart Images

Figure CN120597126A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of automated intelligent dispatching of power systems, and in particular relates to a method, system and storage medium for optimizing and solving unit combination problems. Background Art
[0002] In power system operation, the unit commitment (UC) problem, as a core optimization task, faces numerous technical challenges that need to be addressed. With the continuous advancement of technology, the application of artificial intelligence (AI) in solving UC problems is becoming increasingly widespread. Currently, AI applications in UC problems face multiple challenges. The construction of AI feature spaces is insufficient in mining and utilizing multi-source heterogeneous data (load, topology, unit parameters, and safety constraints), making it difficult to improve prediction accuracy. UC problems and their optimal solutions lack explicit structured features and require complex feature engineering before they can be used for AI analysis and training. Furthermore, variable prediction accuracy is difficult to guarantee, which may lead to computational efficiency and safety issues.
[0003] Therefore, a better solution method is urgently needed to improve the prediction accuracy of artificial intelligence on UC problems and provide an optimal solution to UC problems with explicit structured characteristics. Summary of the Invention
[0004] Purpose of the invention: This application discloses a machine learning enhanced unit combination problem optimization solution method, system and equipment, aiming to improve the accuracy of UC problem solution by using artificial intelligence technology.
[0005] Technical solution:
[0006] In a first aspect, the present invention discloses a method for optimizing and solving unit commitment problems enhanced by machine learning, comprising:
[0007] Analyze the unit commitment problem model of the power system, including the objective function and constraints of the unit commitment problem; collect power system load data, substitute it into the unit commitment problem model and solve it;
[0008] Constructing an input feature space based on the power system load data; selecting constraints and decision variables to construct an output label feature space, wherein the decision variables include the start and stop status of the unit, the start action, and the shutdown action;
[0009] Taking the input feature space and the output label class feature space as a sample set, training the sample set based on a machine learning algorithm to obtain a classifier corresponding to the machine learning algorithm;
[0010] Extracting the load data of the target problem to be solved, inputting it into the classifier, and performing probabilistic fusion on the output results of the classifier to obtain the validity of the constraint conditions and the values of the decision variables;
[0011] The validity of the constraint conditions and the values of the decision variables are written into the unit commitment problem model to solve the target problem.
[0012] Furthermore, the objective function of the unit commitment problem model is expressed as:
[0013]
[0014] Where T is the total number of time periods in the dispatch cycle, N is the number of generator sets, S i,t is the startup cost of the i-th generator set in the t-th period, C i,t is the downtime cost of the i-th generator set in the t-th period, F i,t is the power generation cost of the i-th generator set in the t-th period;
[0015] The constraints of the unit commitment problem include:
[0016] 1) System power balance constraints:
[0017]
[0018] Among them, P i,t is the output of the i-th generator set in the t-th period, P Load,t is the system load in period t;
[0019] 2) Startup and shutdown costs:
[0020]
[0021] Among them, I i,t is the start and stop status of the i-th generator set in the t-th period, 1 means start, 0 means stop, K i represents the fixed startup cost of the i-th generator, J i represents the fixed downtime cost of the i-th generator;
[0022] 3) Upper and lower limits of unit output:
[0023]
[0024] Among them, P i,min is the minimum output of the i-th generator set, P i,max is the maximum output of the i-th generator set;
[0025] 4) Climbing and sliding constraints:
[0026]
[0027] r u,i is the ramp rate of the i-th generator set, rd,i is the landslide rate of the i-th generator set;
[0028] 5) Minimum boot time constraint:
[0029] I i,t =1,t∈[1,U i ],U i =min{T,T on,i I i,0}
[0030]
[0031]
[0032] Among them, T on,i is the minimum startup time of the i-th generator set;
[0033] 6) Minimum downtime constraint:
[0034] I i,t =0,t∈[1,D i ],D i =min{T,T off,i (1-I i,0 )}
[0035]
[0036] Among them, T off,i is the minimum downtime of the i-th generator set;
[0037] 7) Spinning reserve constraints:
[0038]
[0039] Among them, P i,t is the output of the i-th new energy unit in the t-th period; P i,min is the minimum output of the i-th new energy unit, P i,max is the maximum output of the i-th new energy unit; Δt is the spinning standby response time of the generator set, S u,i,t is the positive spinning reserve capacity that generator set i can provide in period t, S d,i,t is the negative spinning reserve capacity that generator set i can provide in time period t; Δu is the upper deviation of the load forecast, and Δd is the lower deviation of the load forecast.
[0040] Furthermore, the power system load data includes load data, new energy time series data, unit operating parameters, and grid safety constraints, wherein the load data includes the total active power demand in each time period and its statistical characteristics.
[0041] Furthermore, constraints and decision variables are selected to construct an output label class feature space. The validity sequences corresponding to the selected constraints are extracted and labeled from the power system load data, unit commitment problem model and its solution to obtain an output label class set. One label class corresponds to a combination of decision variable values and constraint validity values of a UC problem.
[0042] Furthermore, the machine learning algorithm includes a K-nearest neighbor algorithm, a random forest algorithm, and a support vector machine algorithm, and the classifier obtained after training includes a K-nearest neighbor sub-classifier, a random forest sub-classifier, and a support vector machine sub-classifier.
[0043] Furthermore, the target problem is substituted into the classifier, and the target problem is processed by each sub-classifier respectively, and the probability value of each label obtained by each sub-classifier is output; the probability fusion process includes selecting several labels with the highest probability values output by each sub-classifier, and merging the selected labels into a comprehensive probability vector based on the probability fusion mechanism as the predicted mean corresponding to the label, and judging the validity of the constraint conditions based on the predicted mean.
[0044] Furthermore, let the predicted mean be p, use the merge method to capture all necessary constraints, set thresholds P1 and P2, and P1 <P2:
[0045] If the mean value p≥P2, the start-stop state variable of the unit is fixed to 1;
[0046] If the mean value p≤P1, the start-stop state variable of the unit is fixed to 0;
[0047] If the mean value p>P1 and p<P2, the start-stop state variables of the unit are not fixed.
[0048] In a second aspect, the present invention discloses a machine learning-enhanced unit commitment problem optimization solution system, comprising:
[0049] The analysis module is used to analyze the unit commitment problem model of the power system, including the objective function and constraints of the unit commitment problem; collect power system load data, substitute it into the unit commitment problem model and solve it;
[0050] A modeling module is configured to construct an input feature space based on the power system load data; select constraints and decision variables to construct an output label feature space, wherein the decision variables include the start and stop status of the unit, the start action, and the shutdown action;
[0051] A training module is used to use the input feature space and the output label class feature space as a sample set, train the sample set based on a machine learning algorithm, and obtain a classifier corresponding to the machine learning algorithm;
[0052] A prediction module is used to extract the load data of the target problem to be solved, input it into the classifier, and perform probabilistic fusion on the output results of the classifier to obtain the validity of the constraint conditions and the values of the decision variables;
[0053] The solution module is used to write the validity of the constraint conditions and the values of the decision variables into the unit commitment problem model to solve the target problem.
[0054] Furthermore, the objective function of the unit commitment problem model is expressed as:
[0055]
[0056] Where T is the total number of time periods in the dispatch cycle, N is the number of generator sets, S i,t is the startup cost of the i-th generator set in the t-th period, C i,t is the downtime cost of the i-th generator set in the t-th period, F i,t is the power generation cost of the i-th generator set in the t-th period;
[0057] The constraints of the unit commitment problem include:
[0058] 1) System power balance constraints:
[0059]
[0060] Among them, P i,t is the output of the i-th generator set in the t-th period, P Load,t is the system load in period t;
[0061] 2) Startup and shutdown costs:
[0062]
[0063] Among them, I i,t is the start and stop status of the i-th generator set in the t-th period, 1 means start, 0 means stop, K i represents the fixed startup cost of the i-th generator, J i represents the fixed downtime cost of the i-th generator;
[0064] 3) Upper and lower limits of unit output:
[0065]
[0066] Among them, P i,min is the minimum output of the i-th generator set, P i,max is the maximum output of the i-th generator set;
[0067] 4) Climbing and sliding constraints:
[0068]
[0069] r u,i is the ramp rate of the i-th generator set, r d,i is the landslide rate of the i-th generator set;
[0070] 5) Minimum boot time constraint:
[0071] I i,t =1,t∈[1,U i ],U i =min{T,T on,i I i,0}
[0072]
[0073] Among them, T on,i is the minimum startup time of the i-th generator set;
[0074] 6) Minimum downtime constraint:
[0075] I i,t =0,t∈[1,D i ],D i =min{T,T off,i (1-I i,0 )}
[0076]
[0077] Among them, T off,i is the minimum downtime of the i-th generator set;
[0078] 7) Spinning reserve constraints:
[0079]
[0080] Among them, P i,t is the output of the i-th new energy unit in the t-th period; P i,min is the minimum output of the i-th new energy unit, P i,max is the maximum output of the i-th new energy unit; Δt is the spinning standby response time of the generator set, S u,i,t is the positive spinning reserve capacity that generator set i can provide in period t, S d,i,t is the negative spinning reserve capacity that generator set i can provide in time period t; Δu is the upper deviation of the load forecast, and Δd is the lower deviation of the load forecast.
[0081] Furthermore, the power system load data includes load data, new energy time series data, unit operating parameters, and grid safety constraints, wherein the load data includes the total active power demand in each time period and its statistical characteristics.
[0082] Furthermore, constraints and decision variables are selected to construct an output label class feature space. The validity sequences corresponding to the selected constraints are extracted and labeled from the power system load data, the unit commitment problem model and its solution to obtain an output label class set. One label class corresponds to a combination of decision variable values and constraint validity values of a unit commitment problem.
[0083] Furthermore, the machine learning algorithm includes a K-nearest neighbor algorithm, a random forest algorithm, and a support vector machine algorithm, and the classifier obtained after training includes a K-nearest neighbor sub-classifier, a random forest sub-classifier, and a support vector machine sub-classifier.
[0084] Furthermore, the target problem is substituted into the classifier, and the target problem is processed by each sub-classifier respectively, and the probability value of each label obtained by each sub-classifier is output; the probability fusion process includes selecting several labels with the highest probability values output by each sub-classifier, and merging the selected labels into a comprehensive probability vector based on the probability fusion mechanism as the predicted mean corresponding to the label, and judging the validity of the constraint conditions based on the predicted mean.
[0085] Furthermore, let the predicted mean be p, use the merge method to capture all necessary constraints, set thresholds P1 and P2, and P1 <P2:
[0086] If the mean value p≥P2, the start-stop state variable of the unit is fixed to 1;
[0087] If the mean value p≤P1, the start-stop state variable of the unit is fixed to 0;
[0088] If the mean value p>P1 and p<P2, the start-stop state variables of the unit are not fixed.
[0089] In a third aspect, the present invention discloses a computer-readable storage medium having a computer program stored thereon, which implements the steps of the aforementioned method when executed by a processor.
[0090] Beneficial effects:
[0091] This paper proposes a machine learning-enhanced optimization method for solving unit commitment problems. It first uses machine learning to rapidly generate high-confidence initial variable values and key constraint predictions, and then uses mathematical programming to refine these predictions. This method maintains computational efficiency while strictly ensuring the feasibility and optimality of the solution. Compared to existing technologies, this method offers significant advantages, including:
[0092] 1) Intelligent prediction method for the activation status of key constraints: This invention analyzes and models the UC problem, summarizes the objective function and constraints of the UC problem from the inherent properties of the unit, constraint rules, and economic measurement function of the scheduling plan, and then selects the most critical constraints through similarity analysis. The activation status of the constraints is accurately identified and predicted, which reduces redundant constraint calculations and improves the efficiency of solving the UC problem.
[0093] 2) Multi-classifier collaborative prediction method based on probability fusion: Compared with the single classifier prediction method used in the existing technology and the simple proportional threshold method to determine the integer variable, this invention integrates multiple classifiers by constructing a probabilistic fusion model, improving the prediction robustness and being able to more accurately predict the activation status of key constraints;
[0094] 3) Top-N high-confidence integer variable fixing optimization method: Design a variable screening mechanism based on confidence sorting, prioritize fixing high-confidence integer variables, effectively reduce the variable scale of the target problem, and improve solution efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0095] Figure 1 Flowchart of the UC accelerated solution model added by machine learning;
[0096] Figure 2 Schematic diagram of the UC acceleration solution module based on probability fusion. DETAILED DESCRIPTION
[0097] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.
[0098] The present invention discloses a method for optimizing and solving unit commitment problems enhanced by machine learning, the process of which is as follows: Figure 1 As shown, including:
[0099] Analyzing the unit commitment problem model of the power system to obtain the objective function and constraints of the unit commitment problem; collecting power system load data, substituting it into the unit commitment problem model and solving it to obtain an original data set, wherein the original data set includes the power system load data, the unit commitment problem model and its optimal solution;
[0100] An input feature space is constructed based on the power system load data of the original dataset to characterize the system power consumption pattern. Constraints are selected as the feature space of the output label class. The validity sequences corresponding to the selected constraints are extracted from the original dataset and labeled to obtain the output label class set.
[0101] The input feature space and the output label class set are used as a sample set, and the sample set is trained based on a machine learning algorithm to obtain a corresponding classifier;
[0102] Extract the feature space of the target problem to be solved, substitute the feature space into the classifier to obtain the probability value of each label, and obtain the final classification result through the Top-N candidate fusion to determine the prediction value and effective constraints corresponding to the label;
[0103] The predicted values and effective constraints are written into the unit commitment problem model to solve the target problem.
[0104] The technical solution of the present invention will be further described below in conjunction with a preferred embodiment.
[0105] The first step is to construct the unit commitment (UC) problem model of the power system.
[0106] The unit commitment problem can be expressed as the following mixed integer linear programming model:
[0107] Objective function:
[0108]
[0109] Constraints:
[0110] System power balance constraints:
[0111]
[0112] Startup and shutdown costs:
[0113]
[0114] Upper and lower limit constraints of unit output:
[0115]
[0116] Climbing and Sliding Constraints:
[0117]
[0118] Minimum startup time constraint:
[0119] I i,t =1,t∈[1,U i ],U i =min{T,T on,i I i,0}
[0120]
[0121]
[0122] Minimum downtime constraints:
[0123] I i,t =0,t∈[1,D i ],D i =min{T,T off,i (1-I i,0 )}
[0124]
[0125] Spinning reserve constraints:
[0126]
[0127] Where: T is the total number of time periods in the dispatch cycle; N is the number of generator sets; S i,t and C i,t are the startup and shutdown costs of the i-th generator set in the t-th period respectively; F i,t is the power generation cost of the i-th generator set in the t-th period; I i,t is the start / stop status of the i-th generator set in the t-th period (1 means start, 0 means stop); P i,t is the output of the i-th generator set in the t-th period; P Load,t is the system load in period t; P i,min and P i,max are the minimum and maximum output of the i-th generator set respectively; r u,i and r d,i are the climbing rate and sliding rate of the i-th generator set respectively; T on,i and T off,i are the minimum start-up time and minimum shutdown time of the i-th generator set respectively; I j,t is the start / stop status of the jth new energy unit in the tth period (1 means start, 0 means stop); P j,t is the output of the jth new energy unit in the tth period; P j,min and P j,max are the minimum and maximum outputs of the jth renewable energy unit respectively; Δt is the spinning standby response time of the generator set, which is the length of a period (e.g. 15 minutes); s u,i,t and s d,i,t are the positive and negative spinning reserve capacities that generator set i can provide in time period t; Δu and Δd are the upper and lower deviations of load forecast, respectively.
[0128] The second step is to build a sample set for the machine learning enhancement model.
[0129] The construction and feature engineering of historical samples are the foundation for accelerating the solution of UC problems enhanced by machine learning. Therefore, designing the feature space and solution space of the machine learning algorithm and establishing the corresponding dataset are prerequisites for designing and training the accelerated solution model. The steps for dataset construction are as follows:
[0130] (1) Constructing the original dataset
[0131] Following the UC problem model specifications from the first step, we substituted load and renewable energy time series data, unit operating parameters, and grid security constraints into the model. This was then solved using a commercial solver to obtain the corresponding optimal solution. The entire set of UC problem models and their corresponding optimal solutions constituted the original dataset.
[0132] (2) Construct the input feature space X
[0133] The feature space X encompasses multidimensional information about power system operation, primarily including load and renewable energy time series data from the original dataset, unit operating parameters, and grid security constraints. Load data typically includes total active power demand and its statistical characteristics (such as maximum and minimum load, average load, load standard deviation, peak-to-valley difference, and volatility) for each time period, and is used to characterize system power consumption patterns. Renewable energy output, unit operating parameters, and grid security constraints are constructed by extracting parameters from the original dataset.
[0134] (3) Construct output label class Y
[0135] The integer decision variables of the UC problem primarily include unit start / stop states, startup actions, and shutdown actions. Considering the exponential growth of the number of random combinations and time periods, this embodiment selects all unit start / stop state variables and the minimum start / stop time constraint as the feature space of the label class Y. A label class corresponds to a combination of integer decision variable values and the minimum start / stop time constraint validity value for a particular UC problem. All sequences of unit start / stop state variable values and minimum start / stop time constraint validity values from the optimal solution in the original dataset are extracted, merged into a binary variable sequence, and labeled. The resulting labels constitute the label class Y set.
[0136] (4) Sample balance and data enhancement
[0137] In actual systems, some operating scenarios (such as unit combinations under extreme weather conditions) have few samples, which can easily lead to model overfitting. In this embodiment, the model is optimized using the following methods:
[0138] Oversampling / undersampling: adjust the sampling weight of minority class samples;
[0139] Scenario generation: Based on Monte Carlo simulation, it adds diverse data such as renewable energy fluctuations and load forecast errors.
[0140] By fusing multi-source data with physical constraints, the present invention constructs input features with high information density, and combines labeling strategies to convert multi-variable and multi-classification predictions into single-label and multi-classification predictions. For example, cluster design not only retains the physical essence of the UC problem, but also provides a reliable data foundation for subsequent probabilistic fusion models.
[0141] The third step is to build a UC acceleration solution module based on probability fusion.
[0142] Compared to traditional UC solving methods, the core concept of this invention is to use historical solution data to learn the characteristic patterns of UC problems, predicting integer variables whose values can be determined in advance and the necessary constraints that have a substantial impact on the optimal solution. These variables are then input into commercial solvers (such as Gurobi or CPLEX), thereby constructing a simplified UC problem, significantly reducing the solution difficulty and ultimately obtaining the optimal unit commitment solution. The UC accelerated solution module based on probabilistic fusion includes two key components: a multi-classifier group and a top-N candidate probabilistic fusion.
[0143] The multi-classifier group consists of random forests, decision trees, and K-nearest neighbors, and improves the robustness and generalization of the model through differentiated modeling strategies. Top-N candidate probability fusion predicts the most likely top-N candidate classifications for each classifier and derives the predicted value of the target variable based on probability-weighted fusion.
[0144] In terms of process, the model first gives the probability estimate of the target problem belonging to each category in each classifier, and screens out the top N classification values with the highest probability values in each classifier. Then, based on the probability fusion mechanism, the probability estimates of all candidate classifications are merged into a comprehensive probability vector to obtain the final classification result. This not only retains the advantageous output of each classifier, but also realizes decision bias correction through probability reweighting, significantly improving the classification robustness in complex data scenarios.
[0145] like Figure 2 As shown in the figure, the construction process of the UC acceleration solution module based on probability fusion is as follows:
[0146] (1) According to the sample set (η1,h1),...,(η n ,h m ), where η i =(η i,1 ,...η i,t ,...η i,T ) represents the input feature parameter of the i-th sample, h j Represents the label class value corresponding to the j-th sample, which is trained based on the K-nearest neighbor, random forest, and support vector machine algorithms, and obtains the corresponding three classifiers.
[0147] (2) Extract the feature space corresponding to the new problem and substitute it into the three classifiers respectively to obtain the probability value of each classification in each classifier.
[0148] (3) Select the top K most probable classification labels in each classifier. Obtain K Y labels ranked from high to low in probability. Each Y label itself represents the 0 / 1 value of each binary variable and the validity of each constraint. Weighted scores are obtained for each binary variable and each constraint. In this embodiment, the mean of the predicted values of the start / stop state variable and the minimum start / stop time constraint validity variable for each unit is obtained based on the sequence of values of the unit start / stop state variable and the minimum start / stop time constraint validity corresponding to each classification label.
[0149] (4) Furthermore, the thresholds P1 and P2 are used to determine whether the variable can be fixed, to what extent it can be fixed, and whether the constraint is valid. Based on the predicted mean value p of the variable, the method of applying the double threshold P1 and P2 decision rule to fix the integer variable is as follows:
[0150] If the mean value p≥P2, the start-stop state variable of the unit is fixed to 1;
[0151] If the mean value p≤P1, the start-stop state variable of the unit is fixed to 0;
[0152] If the mean value p>P1 and p<P2, the start-stop state variables of the unit are not fixed.
[0153] Final values of the unit start / stop state variables: Fixing incorrect binary integer variable values can exclude the optimal solution from the solution space, or even lead to an infeasible solution. Therefore, only when the model has a very high confidence level in predicting the variable values should the variable be fixed. In this example, P1 = 0.1 and P2 = 0.9 are set.
[0154] Final value of the minimum start-up and downtime constraint validity: Considering the importance of the minimum start-up and downtime constraint, in order to avoid missing necessary constraints, a merging approach should be adopted to capture all necessary constraints as much as possible. Therefore, when p>0, it is a valid minimum start-up and downtime constraint.
[0155] According to the MILP modeling specification, the values of all unit start-up and shutdown state variables and the effective minimum start-up and shutdown time constraints predicted in (3) are added to the problem model, and then the commercial solver is called to achieve accelerated solution of the UC problem.
[0156] In this example, the mixed integer programming algorithms of Gurobi V11 and CPLEX V22 commercial solvers were used for solving the problem. The server configuration was two Intel Xeon 4216 2.1 GHz 16-core CPUs, 256 GB of memory, and an NVIDIA Tesla V100 GPU computing card. An SG-126 system was constructed for the experiment. The system consists of 126 buses and 54 units. As the main test system, the system originated from the State Grid Dispatching AI Innovation Competition. The data set includes 200,000 AC flow sections with a period of 5 minutes and 2 years. The flow characteristics are consistent with the actual grid operation characteristics. The maximum fluctuation amplitude of the total output of renewable energy and the total load can reach 20-50%. Two day-ahead unit combination model scenarios are set, with a time granularity of 15 minutes and 5 minutes, corresponding to 96 time periods and 288 continuous time periods in 24 hours respectively. During the test, 200 test samples under different load conditions are randomly generated and tested on two commercial solvers, Gurobi and CPLEX, respectively. The benchmark method directly uses the commercial solver to solve the UC problem, and the performance value of the benchmark method is taken as 1. The ratio of the proposed method to the benchmark method is recorded to evaluate the impact of the proposed method on the solution acceleration performance.
[0157] Tables 1 and 2 show the solving performance of Gurobi and CPLEX solvers in the 96-period scenario of the SG126 system, respectively.
[0158] Table 1 Gurobi’s solution performance in the 96-hour scenario of the SG126 system
[0159]
[0160] Table 2. CPLEX solution performance in the 96-period scenario of the SG126 system.
[0161]
[0162] As can be seen from Tables 1 and 2, the performance of the present invention is significantly better than the benchmark method, because the integer variable scale is the biggest factor affecting the solution time, and reducing the integer variable scale is the most effective way to accelerate. The relative error of the target value of the present invention remains at an extremely low level, proving that the model of the present invention can ensure the accuracy of the solution. The acceleration effect on both Gurobi and CPLEX solvers is obvious, indicating that the method of this paper has good versatility. At the same time, this paper tested the solution performance in the 288-hour scenario of the SG126 system, and the results are shown in Tables 3 and 4:
[0163] Table 3 Gurobi’s solution performance in the 288-hour scenario of the SG126 system
[0164]
[0165] Table 4. CPLEX solution performance in the 288-hour scenario of the SG126 system.
[0166]
[0167] It can be seen from Tables 3 and 4 that, compared with the results of the 96-period scenario, the acceleration effect of the method of the present invention in the 288-period scenario is more significant.
[0168] The machine learning-enhanced UC problem optimization model designed in this paper achieves acceleration effects of approximately 6 times and 11 times in the 96-period and 288-period scenarios of the SG-126 system, respectively, without sacrificing or even slightly improving the quality of the optimal solution. This shows that the UC problem optimization solution method disclosed in this invention significantly reduces the solution time compared to traditional methods, providing a reliable technical path for real-time scheduling under a high proportion of new energy access, and has good versatility, stability, and scalability.
Claims
1. A machine learning-enhanced method for optimizing and solving unit commitment problems, characterized in that: include: Analyze the unit commitment problem model of the power system, including the objective function and constraints of the unit commitment problem; Collect power system load data, substitute it into the unit commitment problem model and solve it; constructing an input feature space based on the power system load data; Selecting constraints and decision variables to construct an output label feature space, wherein the decision variables include the start and stop status of the unit, the start action, and the shutdown action; Taking the input feature space and the output label class feature space as a sample set, training the sample set based on a machine learning algorithm to obtain a classifier corresponding to the machine learning algorithm; Extracting the load data of the target problem to be solved, inputting it into the classifier, and performing probabilistic fusion on the output results of the classifier to obtain the validity of the constraint conditions and the values of the decision variables; The validity of the constraint conditions and the values of the decision variables are written into the unit commitment problem model to solve the target problem.
2. The method for optimizing and solving the unit commitment problem according to claim 1, characterized in that: The objective function of the unit commitment problem model is expressed as: Where T is the total number of time periods in the dispatch cycle, N is the number of generator sets, S i,t is the startup cost of the i-th generator set in the t-th period, C i,t is the downtime cost of the i-th generator set in the t-th period, F i,t is the power generation cost of the i-th generator set in the t-th period; The constraints of the unit commitment problem include: 1) System power balance constraints: Among them, P i,t is the output of the i-th generator set in the t-th period, P Load,t is the system load in period t; 2) Startup and shutdown costs: Among them, I i,t is the start and stop status of the i-th generator set in the t-th period, 1 means start, 0 means stop, K i represents the fixed startup cost of the i-th generator, J i represents the fixed downtime cost of the i-th generator; 3) Upper and lower limits of unit output: Among them, P i,min is the minimum output of the i-th generator set, P i,max is the maximum output of the i-th generator set; 4) Climbing and sliding constraints: r u,i is the ramp rate of the i-th generator set, r d,i is the landslide rate of the i-th generator set; 5) Minimum boot time constraint: I i,t =1,t∈[1,U i ],U i =min{T,T on,i I i,0 } Among them, T on,i is the minimum startup time of the i-th generator set; 6) Minimum downtime constraint: I i,t =0,t∈[1,D i ],D i =min{T,T off,i (1-I i,0 )} Among them, T off,i is the minimum downtime of the i-th generator set; 7) Spinning reserve constraints: Among them, P i,t is the output of the i-th new energy unit in the t-th period; P i,min is the minimum output of the i-th new energy unit, P i,max is the maximum output of the i-th new energy unit; Δt is the spinning standby response time of the generator set, S u,i,t is the positive spinning reserve capacity that generator set i can provide in period t, S d,i,t is the negative spinning reserve capacity that generator set i can provide in time period t; Δu is the upper deviation of the load forecast, and Δd is the lower deviation of the load forecast.
3. The method for optimizing and solving the unit commitment problem according to claim 2, characterized in that: The power system load data includes load data, new energy time series data, unit operating parameters, and grid security constraints, wherein the load data includes the total active power demand in each time period and its statistical characteristics.
4. The method for optimizing and solving the unit commitment problem according to claim 3, characterized in that: Constraints and decision variables are selected to construct an output label class feature space. The validity sequences corresponding to the selected constraints are extracted and labeled from the power system load data, the unit commitment problem model and its solution to obtain an output label class set. One label class corresponds to a combination of decision variable values and constraint validity values of a unit commitment problem.
5. The method for optimizing and solving the unit commitment problem according to claim 4, characterized in that: The machine learning algorithms include K-nearest neighbor algorithm, random forest algorithm, and support vector machine algorithm. The classifiers obtained after training include K-nearest neighbor sub-classifier, random forest sub-classifier, and support vector machine sub-classifier.
6. The method for optimizing and solving the unit commitment problem according to claim 5, characterized in that: The target problem is substituted into the classifier, and the target problem is processed by each sub-classifier respectively, and the probability value of each label obtained by each sub-classifier is output; the probability fusion process includes selecting several labels with the highest probability values output by each sub-classifier, and merging the selected labels into a comprehensive probability vector based on the probability fusion mechanism as the predicted mean corresponding to the label, and judging the validity of the constraint conditions based on the predicted mean.
7. The method for optimizing and solving the unit commitment problem according to claim 6, characterized in that: The predicted mean is p, and all necessary constraints are captured by merging, and thresholds P1 and P2 are set. <P2: If the mean p ≥ P2, the decision variable takes the value of 1; If the mean p≤P1, the decision variable takes the value of 0; If the mean p>P1 and p<P2, the decision variable is not fixed.
8. A machine learning-enhanced unit commitment problem optimization solution system, characterized in that: include: Analysis module, used to analyze the unit commitment problem model of the power system, including the objective function and constraints of the unit commitment problem; Collect power system load data, substitute it into the unit commitment problem model and solve it; A modeling module, configured to construct an input feature space based on the power system load data; Selecting constraints and decision variables to construct an output label feature space, wherein the decision variables include the start and stop status of the unit, the start action, and the shutdown action; A training module is used to use the input feature space and the output label class feature space as a sample set, train the sample set based on a machine learning algorithm, and obtain a classifier corresponding to the machine learning algorithm; A prediction module is used to extract the load data of the target problem to be solved, input it into the classifier, and perform probabilistic fusion on the output results of the classifier to obtain the validity of the constraint conditions and the values of the decision variables; The solution module is used to write the validity of the constraint conditions and the values of the decision variables into the unit commitment problem model to solve the target problem.
9. The unit commitment problem optimization solution system according to claim 8, characterized in that: The objective function of the unit commitment problem model is expressed as: Where T is the total number of time periods in the dispatch cycle, N is the number of generator sets, S i,t is the startup cost of the i-th generator set in the t-th period, C i,t is the downtime cost of the i-th generator set in the t-th period, F i,t is the power generation cost of the i-th generator set in the t-th period; The constraints of the unit commitment problem include: 1) System power balance constraints: Among them, P i,t is the output of the i-th generator set in the t-th period, P Load,t is the system load in period t; 2) Startup and shutdown costs: Among them, I i,t is the start and stop status of the i-th generator set in the t-th period, 1 means start, 0 means stop, K i represents the fixed startup cost of the i-th generator, J i represents the fixed downtime cost of the i-th generator; 3) Upper and lower limits of unit output: Among them, P i,min is the minimum output of the i-th generator set, P i,max is the maximum output of the i-th generator set; 4) Climbing and sliding constraints: r u,i is the ramp rate of the i-th generator set, r d,i is the landslide rate of the i-th generator set; 5) Minimum boot time constraint: I i,t =1,t∈[1,U i ],U i =min{T,T on,i I i,0 } Among them, T on,i is the minimum startup time of the i-th generator set; 6) Minimum downtime constraint: I i,t =0,t∈[1,D i ],D i =min{T,T off,i (1-I i,0 )} Among them, T off,i is the minimum downtime of the i-th generator set; 7) Spinning reserve constraints: Among them, P i,t is the output of the i-th new energy unit in the t-th period; P i,min is the minimum output of the i-th new energy unit, P i,max is the maximum output of the i-th new energy unit; Δt is the spinning standby response time of the generator set, S u,i,t is the positive spinning reserve capacity that generator set i can provide in period t, S d,i,t is the negative spinning reserve capacity that generator set i can provide in time period t; Δu is the upper deviation of the load forecast, and Δd is the lower deviation of the load forecast.
10. The unit commitment problem optimization solution system according to claim 9, characterized in that: The power system load data includes load data, new energy time series data, unit operating parameters, and grid security constraints, wherein the load data includes the total active power demand in each time period and its statistical characteristics.
11. The unit commitment problem optimization solution system according to claim 10, characterized in that: Constraints and decision variables are selected to construct an output label class feature space. The validity sequences corresponding to the selected constraints are extracted and labeled from the power system load data, the unit commitment problem model and its solution to obtain an output label class set. One label class corresponds to a combination of decision variable values and constraint validity values of a unit commitment problem.
12. The unit commitment problem optimization solution system according to claim 11, characterized in that: The machine learning algorithms include K-nearest neighbor algorithm, random forest algorithm, and support vector machine algorithm. The classifiers obtained after training include K-nearest neighbor sub-classifier, random forest sub-classifier, and support vector machine sub-classifier.
13. The unit commitment problem optimization solution system according to claim 12, characterized in that: The target problem is substituted into the classifier, and the target problem is processed by each sub-classifier respectively, and the probability value of each label obtained by each sub-classifier is output; the probability fusion process includes selecting several labels with the highest probability values output by each sub-classifier, and merging the selected labels into a comprehensive probability vector based on the probability fusion mechanism as the predicted mean corresponding to the label, and judging the validity of the constraint conditions based on the predicted mean.
14. The unit commitment problem optimization solution system according to claim 13, characterized in that The predicted mean is p, and all necessary constraints are captured by merging, and thresholds P1 and P2 are set. <P2: If the mean value p≥P2, the start-stop state variable of the unit is fixed to 1; If the mean value p≤P1, the start-stop state variable of the unit is fixed to 0; If the mean value p>P1 and p<P2, the start-stop state variables of the unit are not fixed.
15. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.
Citation Information
Patent Citations
Mixed integer programming joint optimization method, system and equipment for security constraint unit
CN116011664A
Short-term unit commitment calculation method and system, medium and equipment
CN116542463A
Unit commitment rapid solving method based on adjoint model dynamic interaction
CN117638859A
Power system unit combination acceleration method
CN117674089A
Deep learning unit commitment problem solving method considering confidence interval
CN118735309A
Cited By
Active power distribution network integrated optimization scheduling method and system based on SVM-L2O
CN121906426A
