Acceleration method, device and storage medium for unit commitment optimization considering safety constraints

Abstract

The present invention discloses a method, device and storage medium for accelerating the optimization of unit combination considering safety constraints, belonging to the field of unit combination optimization acceleration algorithms considering safety constraints in power systems. The SCUC model is further designed by characterizing the corresponding parameters. Then, samples are continuously generated, explored and pruned in GCN, and the corresponding acceleration strategy is optimized. In this acceleration strategy, only tight constraints are retained, binary variables are optimized, and optimization branches of large-scale calculations are reduced. Through the acceleration strategy, the computational efficiency of SCUC is improved, and its decision-making requirements in large-scale systems are met. Finally, the proposed optimization acceleration algorithm is verified in IEEE 30-node and 118-node systems, the computational time and optimality of the solution are evaluated, the experimental results are analyzed, and compared with traditional solvers, GCN achieves performance improvement, the average computation speed is increased by 39.9 times, and the relative gaps in predicted solutions are 0.07% and 0.16%.

Description

Technical Field

[0001] The present invention belongs to the field of unit combination optimization acceleration algorithms considering safety constraints in power systems, and specifically relates to a unit combination optimization acceleration method, device and storage medium considering safety constraints. Background Art

[0002] With the continuous development of current science and technology, the security-constrained unit commitment (SCUC) problem is a classic basic research content in power systems and an important task in power system operation. However, operations research algorithms may encounter problems such as complex models, large number of iterations, insufficient stability and inaccuracy. Considering large-scale real-time scheduling, the convergence time is often difficult to meet the requirements of the UC problem. At the same time, with the continuous development of future power systems, the next generation of SCUC decision-making tools should effectively integrate various complex functions, including active participation in demand response in energy and ancillary service markets, and improving adaptability to intermittent and volatile renewable energy through various energy storage options. In order to meet these needs, the modeling and solution of SCUC will become more and more complex. Not surprisingly, traditional acceleration methods will increasingly increase their algorithmic complexity. Therefore, data-driven and machine learning have been cleverly applied in artificial intelligence technology. The purpose of this invention is to use artificial intelligence algorithms to improve the computing power and efficiency of large-scale unit commitment decision-making. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a method, device and storage medium for optimizing and accelerating unit commitment considering safety constraints.

[0004] The object of the present invention can be achieved by the following technical solution: a unit commitment optimization acceleration method considering safety constraints, the method comprising the following steps:

[0005] Given a power grid system with known topology and line parameters, the system uses historical user load information in the region of the power grid system with known topology and line parameters. By predicting and optimizing the start and stop status of units within a day or week and the day-ahead real-time power generation scheduling, a unit combination SCUC model that considers safety constraints is established, and a SCUC scenario is established based on the SCUC model.

[0006] According to the data characteristics of the established SCUC scenario, the load fluctuation data integrated with the known grid system dispatch information is used as the collected training data features to construct the input parameters of the acceleration strategy;

[0007] Based on the input parameters, a graph neural network multivariate classifier is used to solve the MILP mixed integer linear programming model of the SCUC problem under different parameters using a universal solver. This model learns and identifies two data labels (0-1 variables) and tight constraints, and establishes a mapping between data features and data labels to form an acceleration strategy.

[0008] Training an acceleration strategy identifies tight constraints based on input parameters and uses it to output the value of binary variables. A graph convolutional neural network (GCN) is used to extract and classify potential data features in SCUC. The acceleration strategy class is applied to simplify the SCUC problem into a linear programming problem with only tight constraints. The acceleration strategy is optimized through exploration and pruning to train the multivariate GCN classifier. Training results for binary variables and tight constraints in the SCUC model are obtained. The KKT condition is then used to solve the value of continuous variables for large-scale, rapid SCUC decision-making.

[0009] Evaluate the training results, calculate the optimal solution based on the Karush-Kuhn-Tucker KKT condition of nonlinear programming optimization theory, and determine whether the optimal solution is feasible, so as to correct the solution;

[0010] The constructed acceleration strategy is validated and the results are analyzed, focusing on the computational efficiency and accuracy of the acceleration strategy.

[0011] Preferably, the process of establishing the SCUC scenario includes the following steps:

[0012] Establish the cost-optimal objective function:

[0013]

[0014] Among them, F i is the power generation cost function of generator set i, P it is the output of generator i at time t, I it is the on / off state of generator i at time t, α i is the comprehensive cost of startup and equipment maintenance, SU it is the startup cost of generator i at time t, SD it is the downtime cost of generator i at time t;

[0015] Set the constraints of the SCUC model:

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] Among them D it is the load of node i at time t, SR it is the spinning reserve requirement of generator i at time t, OR it is the operating reserve requirement of generator i at time t, UR it is the operating ramp coefficient of generator i at time t, DR it is the ramp coefficient of generator i at time t, UP it is the operating unit i at time t, DP it is the shutdown unit i at time t, X on / off,it is the minimum start-up and shutdown time of generator i at time t, T on / off,i is the minimum on-off time of generator i, PL t is the line power flow matrix at time t, θ t is the node voltage phase angle matrix at time t, V t is the node voltage amplitude matrix at time t, V min / max Indicates the maximum and minimum values ​​of the node voltage, PL max represents the maximum power flow capacity limit of the line, and f is the objective function expressed in (1).

[0028] Preferably, the constraints of the SCUC include constraints on basic unit combinations and transmission network security constraints; the variables appearing in the constraints of the SCUC model include system states and control variables; line flow, bus voltage amplitude and bus voltage angle belong to the set of system state variables, and the target value is optimized by adjusting the control variables, that is, the actual power output of each generator set. Equation (2) represents the system power balance, including general UC constraints and system rotation (3) and operation (4) reserve requirements, actual power generation capacity limitations (5), minimum opening (6) and closing (7) time limitations, and generator set ramp-up (8) and ramp-down (9) limitations; considering the transmission network constraints, equation (10) represents the transmission flow, and inequalities (11) and (12) enforce transmission flow limitations and bus voltage limitations.

[0029] Preferably, according to the data features of the established SCUC scenario, extracting and sending signals and collecting training data through simulation calculation results of the SCUC model include the following steps:

[0030] Selecting Data Features for SCUC

[0031] In the SCUC problem, the load variation trend is repetitive and small. Analysis shows that the system network topology and line parameters are predetermined and difficult to change. The system network topology and line parameters are considered as known fixed values, while the user load demand is regarded as an independent variable and a variable input parameter. Therefore, the user load demand is selected as the data feature label of the SCUC problem, and a parameter optimization formula is introduced to establish a supervised learning task.

[0032] Abstract and parameterize the data features of the SCUC model:

[0033] The parameterization formula of SCUC is as follows:

[0034]

[0035] in

[0036] f i (b,x)=w i x(j|x(j)=I i,t )p(j|p(j)=Load i,t ) (14)

[0037] The constraint expression is:

[0038] g k (x,p)≤0,k=1,...,m (15)

[0039] Where x is the vector of all integer and continuous decision variables in SCUC, p is the concatenated vector of load demands p = {Systemload}, gk (x,p) represents the kth constraint of SCUC, and m is the number of constraints.

[0040] Preferably, according to the parameterized formula of SCUC, the SCUC problem is determined and solved by giving an input parameter p.

[0041] Preferably, the process of learning and identifying two data labels includes:

[0042] Based on the input parameters, the general solver Gurobi is used to calculate and collect training data to form a network learning memory and collect n input parameter samples P. n ={p1,...,p n}, through the GCN acceleration strategy, output strategy s(p i ) corresponds to the value with parameter p i ∈P N The parameterized MILP problem is the optimal decision value and tight constraint index of the binary variables,

[0043]

[0044] Optimizing 0-1 variables

[0045] One solution to the SCUC problem is to find the optimal start and stop states for the units and then solve the day-ahead and real-time generation scheduling problems. Therefore, we only need to obtain the 0-1 state of the I variable for each generator. Therefore, let the optimal value of the binary variable be expressed as

[0046] x * (p i )=(I * ) (17)

[0047] Finding tight constraints

[0048] Define optimality tight constraints as another acceleration metric, which means that some inequality constraints are equal to each other to achieve optimality. Such inequality constraints are identified as tight constraints and retained in the acceleration strategy.

[0049]

[0050] where x*(p i ) is the parameter p i The optimal solution under , the number of tight constraints is less than the total number of constraints, which means that there are a lot of redundant constraints in the original SCUC problem. Then, through the learning strategy Ignoring a large number of redundant constraints, the acceleration strategy is implemented. The acceleration model for the SCUC optimization problem for GCN learning is as follows:

[0051]

[0052] where x 0-1 Represents a binary variable.

[0053] Preferably, the process of training the acceleration strategy includes the following steps:

[0054] Exploring and pruning accelerated branches

[0055] Construct a multi-class classification problem from sample parameters to acceleration strategies. The sample set contains the optimal acceleration strategies corresponding to many load scenarios. In order to select the optimal acceleration strategy under a certain sample parameter, the acceleration strategy identification of the SCUC problem is regarded as a multi-class classification task in supervised learning. The purpose of this multi-class classification task is to select the optimal acceleration strategy under a given sample parameter p. i Predict the corresponding acceleration strategy under the circumstances;

[0056] Suppose that given n independent samples P n ={p1,...,p n} to find m unique strategies S(P n )={s1,...,s m}, then new policies will emerge when continuously sampling data, so a good Turing estimator is used to evaluate the probability of discovering an unseen policy, encountering parameters p corresponding to the unseen policy n+1 The probability satisfies the following expression with a confidence level of 1-β:

[0057]

[0058] Where G = n1 / n, the number of different strategies corresponding to n1 only appears once, and the coefficient is

[0059] Learning acceleration strategies using multi-class classifiers

[0060] Based on the analysis of the parameter SCUC problem, the input and output parameters used in the classification task are as follows:

[0061] p={Systemload} (21)

[0062] s(p)={value of binary variables,index of tight constraints} (22)

[0063] During the classification process, s i is the hot code used to input parameter p i Rewrite the labels of the policy classes; graph neural networks and decision tree ensembles are combined into a multi-class classifier;

[0064] Graph Convolutional Neural Network Learning Strategies

[0065] The algorithmic functionality of the GCN model consists of L layers, defined as follows:

[0066]

[0067] where s i is the prediction strategy class, h l It is the lth hidden layer in the neural network. Since the data transmitted in the graph neural network must be represented by the structure of the graph, we must first start from the parameter p i To node feature X i The mapping is represented by h0:

[0068] X i =h0(p i ) (twenty four)

[0069] The characteristics of each node X i Contains two dimensions, representing the node load and fault indicator respectively. If the node adjacent line fails, the fault indicator is 1, otherwise it is 0. The components of each hidden layer are

[0070] y l =h l (y l-1 ,A)=σ l (Ay l-1 W l +b l ),l=1,...,L (25)

[0071] where y l is the data obtained through layer operations, A is the adjacency matrix of the power network, W l and b l is the network weight parameter, the input layer is defined as l = 1, the output layer is defined as l = L, so y0 = X i ,y L =s i , each layer uses parameter W l and b l Perform affine transformations;

[0072] In addition, the activation function σ l It is integrated into the nonlinear model and the activation function ReLU, that is, the linear rectification function, is introduced as follows:

[0073] σ l (x)=max(x,0),l=1,...,L-1 (26)

[0074] Among them, the max operator is executed bitwise, and the ReLU operator shows higher performance and can avoid the vanishing gradient problem that occurs in the usual sigmoid function. The output layer is set to the Softmax function to provide a standardized sorting between acceleration strategies and evaluate the possibility of them becoming the correct strategy. The Softmax function is very common in multi-classification problems due to its smoothness and probabilistic interpretation. The selection layer of the acceleration strategy, namely σ L Expressed as

[0075]

[0076] During the training of the GCN network, the cross entropy loss function is applied.

[0077]

[0078] The training process uses the stochastic gradient descent method to obtain the loss function derivative and trains based on the back-propagation rule.

[0079] Preferably, the process of evaluating the obtained training results includes:

[0080] The acceleration strategy of the GCN classifier is used to transform the SCUC problem of the MILP into the SCUC problem of the linear programming LP. Then, the KKT condition is used to calculate the optimal solution of the transformed SCUC model. If the solution under the acceleration strategy is feasible, it is output as the solution to the SCUC problem. If the acceleration strategy encounters an infeasible problem, the predicted data is set for pre-solution, and the traditional solver Gurobi is used to calculate and output the optimization result of the large-scale SCUC problem.

[0081] Beneficial effects of the present invention:

[0082] This invention utilizes a simpler sampling method to implement and further innovate the practical application of the GCN acceleration framework, meeting the requirements of engineering applications. While some new proposals can accelerate SCUC, prediction accuracy, feature recognition complexity, and acceleration multipliers can be developed and improved. This invention introduces a graph neural network multivariate classifier for learning policies from data. When performing day-ahead SCUC scheduling and start / stop prediction, the trained multivariate classifier identifies and collects binary variable values ​​and tight constraints to ensure safe unit commitment decisions based on extracted characteristic load parameters. The ReLU activation function further enhances the expressive power of the graph neural network. Thus, this invention transforms the MILP SCUC problem into a computationally less computationally intensive LP problem. In this process, more network constraints are eliminated compared to existing methods, tight constraints are identified, and a higher computational cost is saved. GCN offers a good balance between complexity and prediction accuracy. Furthermore, the size of the training set can be limited by existing theories. Therefore, learning results can be obtained faster and more accurately without wasting large amounts of data, which can be applied to large-scale, rapid SCUC decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0084] Figure 1 Flowchart of the present invention;

[0085] Figure 2 This is a single-line diagram of an IEEE 30-bus system according to an embodiment of the present invention;

[0086] Figure 3 Schematic diagram of superimposed optimal generator output given by GCN in a 30-node system according to an embodiment of the present invention;

[0087] Figure 4 Schematic diagram of superimposed optimal generator output given by Gurobi in a 30-node system according to an embodiment of the present invention;

[0088] Figure 5 This is a convergence curve diagram of GCN in a 30-node system according to an embodiment of the present invention;

[0089] Figure 6 This is a single-line diagram of the IEEE 118-node system according to an embodiment of the present invention;

[0090] Figure 7 Schematic diagram of superimposed optimal generator output provided by GCN in the 118-node system of Example 118 of the present invention;

[0091] Figure 8 This is a schematic diagram of the superimposed optimal generator output given by Gurobi in the 118-node system of Example 1 of the present invention. DETAILED DESCRIPTION

[0092] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0093] like Figure 1-8 As shown, the unit commitment optimization acceleration method considering safety constraints includes the following steps:

[0094] Step 1: Based on the existing theories of unit commitment, establish the cost-optimal objective function under the SCUC problem. On this basis, set physical and operational constraints to construct a unit commitment model scenario considering safety issues. The specific method is as follows:

[0095] Establish the objective function of minimizing the cost of the generator set:

[0096]

[0097] Among them, F i is the power generation cost function of generator set i, P it is the output of generator i at time t, I it is the on / off state of generator i at time t, α i is the comprehensive cost of startup and equipment maintenance, SU it is the startup cost of generator i at time t, SD it is the downtime cost of generator i at time t

[0098] Set the constraints of the SCUC model:

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110] Among them D it is the load of node i at time t, SR it is the spinning reserve requirement of generator i at time t, OR it is the operating reserve requirement of generator i at time t, UR it is the operating ramp coefficient of generator i at time t, DR it is the ramp coefficient of generator i at time t, UP it is the operating unit i at time t, DP it is the shutdown unit i at time t, X on / off,it is the minimum start-up and shutdown time of generator i at time t, T on / off,i is the minimum on-off time of generator i, PL t is the line power flow matrix at time t, θ t is the node voltage phase angle matrix at time t, V t is the node voltage amplitude matrix at time t, V min / max Indicates the maximum and minimum values ​​of the node voltage, PL max represents the maximum power flow capacity limit of the line. f is the objective function expressed in (29).

[0111] The SCUC constraints include constraints on the basic unit combination and transmission network security constraints. The variables that appear in the transmission network security constraints in the SCUC formula include system states and control variables. Line power flow, bus voltage amplitude and bus voltage angle belong to the set of system state variables. By adjusting the control variables, that is, the actual power output of each generator unit, the target value can be optimized. Equation (30) represents the system power balance, which includes general UC constraints as well as system rotation (31) and operation (32) reserve requirements, actual generation capacity limitations (33), minimum on (34) and off (35) time limitations, and generator unit ramp-up (36) and ramp-down (37) limitations. Considering the transmission network security constraints, equation (38) describes the transmission flow, while inequalities (39) and (40) enforce transmission flow limitations and bus voltage limitations.

[0112] Step 2: Based on the data characteristics of the established SCUC scenario, parameterized model features are selected. The process of extracting, sending signals, and collecting training data includes the following steps:

[0113] Selecting Data Features for SCUC

[0114] In the SCUC problem, load changes are repetitive and small. Analysis shows that the system network topology and line parameters are predetermined and difficult to change. Considering these system network topology and line parameters as known fixed values, and user load demands as independent variables and variable input parameters, we select user load demands as the data feature labels for the SCUC problem and introduce the following parameter optimization formula to establish the supervised learning task.

[0115] Abstracting and parameterizing data features of the SCUC model

[0116] The parameterization formula of SCUC is as follows:

[0117]

[0118] in

[0119] f i (b,x)=wix(j|x(j)=I i,t )p(j|p(j)=Load i,t ) (42)

[0120] The constraint expression is:

[0121] g k (x,p)≤0,k=1,...,m (43)

[0122] Where x is the vector of all integer and continuous decision variables in SCUC, p is the concatenated vector of load demands p = {Systemload}, g k (x,p) represents the kth constraint of SCUC, and m is the number of constraints.

[0123] According to the parameterized formula model of SCUC, the SCUC problem is determined and solved by giving the input parameter p.

[0124] Step 3: Based on the parameterized data features of SCUC, a graph neural network multivariate classifier is used to learn and identify two data labels, and then a mapping is established between the features and labels. The process of forming an acceleration strategy includes:

[0125] Based on the input parameters, the general solver Gurobi is used to calculate and collect training data to form a network learning memory to solve the SCUC problem in different situations. The present invention proposes that if n input parameter samples P are collected under historical data training, n ={p1,...,p n}, through the GCN acceleration strategy, output strategy s(p i ) corresponds to the value with parameter p i ∈P N The parameterized MILP problem is the optimal decision value and tight constraint index of the binary variables,

[0126]

[0127] Optimizing 0-1 variables

[0128] One solution to the SCUC problem is to find the optimal start and stop states for the units and then solve the day-ahead and real-time generation scheduling problems. Therefore, we only need to obtain the 0-1 state of the I variable for each generator. Therefore, let the optimal value of the binary variable be expressed as

[0129] x * (p i )=(I * ) (45)

[0130] Finding tight constraints

[0131] Define optimality tight constraints as another acceleration metric, which means that some inequality constraints can be equal to each other to achieve optimality. Such inequality constraints are identified as tight constraints and retained in the acceleration strategy.

[0132]

[0133] where x*(p i ) is the parameter p i The optimal solution under , the number of tight constraints is less than the total number of constraints, which means that there are a lot of redundant constraints in the original SCUC problem. Then, through the learning strategy Ignoring a large number of redundant constraints, the acceleration strategy is implemented. The acceleration model for the SCUC optimization problem for GCN learning is as follows:

[0134]

[0135] where x 0-1 Represents a binary variable.

[0136] Step 4: Training the acceleration strategy, that is, identifying tight constraints based on input parameters and outputting the value of the binary variable through the acceleration strategy. GCN's accelerated classification simplifies the SCUC problem into a linear programming problem with only tight constraints. The acceleration strategy is optimized through exploration and pruning. The process of training the multivariate classifier GCN to make large-scale rapid decisions on SCUC based on the extracted feature load parameters includes:

[0137] Exploring and pruning accelerated branches

[0138] A class is a custom reference data type, also known as a class type. Each class contains a data description and a set of functions that operate on data or pass messages. Instances of a class are called objects. The recognition of acceleration strategies for the SCUC problem can be viewed as a multi-class classification task in supervised learning. The purpose of the multi-class classification task is to find the best solution for a given sample parameter p. i Predict the corresponding acceleration strategy class under the circumstances;

[0139] Suppose that given n independent samples P n ={p1,...,p n} to find m unique strategies S(P n )={s1,...,s m}, then new policies will emerge when continuously sampling data, so a good Turing estimator is used to evaluate the probability of discovering an unseen policy, encountering parameters p corresponding to the unseen policy n+1 The probability satisfies the following expression with a confidence level of 1-β:

[0140]

[0141] Where G = n1 / n, the number of different strategies corresponding to n1 only appears once, and the coefficient is

[0142] Learning acceleration strategies using multi-class classifiers

[0143] Based on the analysis of the parameter SCUC problem, the input and output parameters used in the classification task are as follows:

[0144] p={Systemload} (49)

[0145] s(p)={value of binary variables,index of tight constraints} (50)

[0146] During the classification process, s i Hot code is used to input parameter p iRewrite the labels of the policy classes; graph neural networks and decision tree ensembles are combined into a multi-class classifier;

[0147] Graph Convolutional Neural Network Learning Strategies

[0148] The algorithmic functionality of the GCN model consists of L layers, defined as follows:

[0149]

[0150] where s i is the prediction strategy class, h l It is the lth hidden layer in the neural network. Since the data transmitted in the graph neural network must be represented by the structure of the graph, we must first start from the parameter p i To node feature X i The mapping is represented by h0:

[0151] X i =h0(p i ) (52)

[0152] The characteristics of each node X i It contains two dimensions, representing the node load and the fault indicator. If the adjacent line of the node fails, the fault indicator is 1, otherwise it is 0. The components of each hidden layer are

[0153] y l =h l (y l-1 ,A)=σ l (Ay l-1 W l +b l ),l=1,...,L (53)

[0154] where y l is the data obtained through layer operations, A is the adjacency matrix of the power network, W l and b l is the network weight parameter, the input layer is defined as l = 1, the output layer is defined as l = L, so y0 = X i ,y L =s i , each layer uses parameter W l and b l Perform affine transformations;

[0155] In addition, the activation function σ l It is integrated into the nonlinear model and the activation function ReLU, that is, the linear rectification function, is introduced as follows:

[0156] σ l (x)=max(x,0),l=1,...,L-1 (54)

[0157] Among them, the max operator is executed bitwise, and the ReLU operator shows higher performance and can avoid the vanishing gradient problem that occurs in the usual sigmoid function. The output layer is set to the Softmax function to provide a standardized sorting between acceleration strategies and evaluate the possibility of them becoming the correct strategy. The Softmax function is very common in multi-classification problems due to its smoothness and probabilistic interpretation. The selection layer of the acceleration strategy, namely σ L Expressed as

[0158]

[0159] During the training of the GCN network, the cross entropy loss function is applied.

[0160]

[0161] The training process uses the stochastic gradient descent method to obtain the loss function derivative and trains based on the back-propagation rule.

[0162] Step 5: Conduct targeted testing on the accelerated training results, calculate the optimal solution based on the KKT conditions, and determine whether the optimal solution is feasible. The process of testing the training results includes:

[0163] The acceleration strategy of the GCN classifier is used to transform the SCUC problem of the MILP into the SCUC problem of the linear programming LP. Then, the KKT condition is used to calculate the optimal solution of the transformed SCUC model. If the solution under the acceleration strategy is feasible, it is output as the solution to the SCUC problem. If the acceleration strategy encounters an infeasible problem, the predicted data is set for pre-solution, and the traditional solver Gurobi is used to calculate and output the optimization result of the large-scale SCUC problem.

[0164] Step 6: Verify the effectiveness of the constructed acceleration algorithm and analyze the results. Simulation experiments were conducted on a laptop equipped with a 2.8GHz Intel Core i7-11700H CPU and 16GB of RAM. Examples of IEEE 30-node and 118-node systems were simulated. The process of analyzing their computational efficiency and accuracy included the following:

[0165] The solutions given by Gurobi and GCN are compared in terms of correctness and computational efficiency to demonstrate the effectiveness of the proposed model. Simulation experiments are conducted on IEEE 30-bus and 118-bus systems. The variable parameter system load data here is obtained from the historical 11-year cumulative hourly load of the default service (including residential, commercial, and industrial customer groups) in three applicable load areas (SEMA, NEMA, and WCMA) in Massachusetts from January 1, 2011 to December 31, 2021. Data scaling is performed on the daily load data in Massachusetts:

[0166]

[0167] in is the normalized k-th hour load in the test system, is the kth hourly load in Massachusetts, is the kth hour load in the original test system. To simulate the actual half-hour behavior of the unit, the half-hour load data is padded with the average of two adjacent full-hour data. GCN is built using the deep learning framework PyTorch, and the network hyperparameters are automatically selected by Optuna.

[0168] Simulation experiment of IEEE 30-bus system

[0169] The topology of the test system is as follows: Figure 1 As shown in Figure 2, GCN is trained using a total of 3653 load data samples from the first 10 years and tested using load data from the past year. Figure 4 The convergence process of the network loss recorded with the tension plate is depicted, showing smooth convergence. The parameter settings of the generator are shown in Table 1, corresponding to the system parameters in the MILP model. Table 2 summarizes the performance of GCN compared with Gurobi (MIPGap=0.01). The results show that the average computation speed is improved by a factor of 13 while maintaining high feasibility. For infeasible cases, the predicted solution is selected as a reasonable initial guess for starting the MIP in Gurobi, thereby reducing the Gurobi solution time.

[0170] Table 1. Generator parameters for IEEE 30-bus system

[0171]

[0172] Table 2 Comparison of GCN and Gurobi solutions in a 30-node system

[0173]

[0174] To better illustrate the comparison between the solution results of GCN and Gurobi, Figure 2 and Figure 3The superimposed optimal generator output for a typical day is shown. In this case, both solutions are feasible. The target value of the GCN solution is 55964796, while the target value of the Gurobi solution is 55926228 (MIPGap=0.001). The relative gap of the predicted solution is 0.07%, which shows that it is a close to optimal solution. As can be seen from the figure, the main difference between the two solutions is the "turn on" operation of the first unit. The Gurobi solution turns on at 8:30, while the GCN solution turns on at 7:30. This shows that the operating cost of the first unit from 7:30 to 8:30 is slightly higher than the cost increment brought by the power allocated to other generators.

[0175] (2) Simulation experiment of IEEE 118-node system

[0176] The network topology of the 118-node system is as follows Figure 5 As shown in Table III, the number of variables and constraints of the MILP model in the 118-node system increases exponentially compared to the 30-node system. In addition, in such a large system, the computational efficiency of Gurobi is very low and it cannot reach the predetermined MIPGap. In contrast, the proposed GCN accelerates by about 40 times and reaches a feasible solution. Figure 6 and Figure 7 In the typical case shown, the superimposed optimal generator output of the two methods is also given. The target value of the GCN solution is 12915143, while the target value of the Gurobi solution is 12894330 (MIPGap=0.001). The relative gap between the predicted solutions is 0.16%, and there is almost no difference between the changes in the unit operating status.

[0177] Table 3 Comparison of GCN and Gurobi solutions in a 30-node system

[0178]

[0179] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0180] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention, and such changes and modifications are intended to fall within the scope of the present invention.

Claims

1. An accelerated method for unit commitment optimization considering safety constraints, characterized in that: The method comprises the following steps: Given a power grid system with known topology and line parameters, the system uses historical user load information in the region of the power grid system with known topology and line parameters. By predicting and optimizing the start and stop status of units within a day or week and the day-ahead real-time power generation scheduling, a unit combination SCUC model that considers safety constraints is established, and a SCUC scenario is established based on the SCUC model. According to the data characteristics of the established SCUC scenario, the load fluctuation data integrated with the known grid system dispatch information is used as the collected training data features to construct the input parameters of the acceleration strategy; Based on the input parameters, a graph neural network multivariate classifier is used to solve the MILP mixed integer linear programming model of the SCUC problem under different parameters using a universal solver. This model learns and identifies two data labels (0-1 variables) and tight constraints, and establishes a mapping between data features and data labels to form an acceleration strategy. Training an acceleration strategy identifies tight constraints based on input parameters and uses it to output the value of binary variables. A graph convolutional neural network (GCN) is used to extract and classify potential data features in SCUC. The acceleration strategy class is applied to simplify the SCUC problem into a linear programming problem with only tight constraints. The acceleration strategy is optimized through exploration and pruning to train the multivariate GCN classifier. Training results for binary variables and tight constraints in the SCUC model are obtained. The KKT condition is then used to solve the value of continuous variables for large-scale, rapid SCUC decision-making. Evaluate the training results, calculate the optimal solution based on the Karush-Kuhn-Tucker KKT condition of nonlinear programming optimization theory, and determine whether the optimal solution is feasible, so as to correct the solution; The constructed acceleration strategy is validated and the results are analyzed, focusing on the computational efficiency and accuracy of the acceleration strategy.

2. The unit commitment optimization acceleration method considering safety constraints according to claim 1 is characterized in that: The process of establishing the SCUC scenario includes the following steps: Establish the cost-optimal objective function: Among them, F i is the power generation cost function of generator set i, P it is the output of generator i at time t, I it is the on / off state of generator i at time t, α i is the comprehensive cost of startup and equipment maintenance, SU it is the startup cost of generator i at time t, SD it is the downtime cost of generator i at time t; Set the constraints of the SCUC model: Among them D it is the load of node i at time t, SR it is the spinning reserve requirement of generator i at time t, OR it is the operating reserve requirement of generator i at time t, UR it is the operating ramp coefficient of generator i at time t, DR it is the ramp coefficient of generator i at time t, UP it is the operating unit i at time t, DP it is the shutdown unit i at time t, X on / off,it is the minimum start-up and shutdown time of generator i at time t, T on / off,i is the minimum on-off time of generator i, PL t is the line power flow matrix at time t, θ t is the node voltage phase angle matrix at time t, V t is the node voltage amplitude matrix at time t, V min / max Indicates the maximum and minimum values ​​of the node voltage, PL max represents the maximum power flow capacity limit of the line, and f is the objective function expressed in (1).

3. The unit commitment optimization acceleration method considering safety constraints according to claim 2 is characterized in that: The constraints of the SCUC include the constraints of the basic unit combination and the transmission network security constraints; the variables that appear in the constraints of the SCUC model include system state and control variables; line flow, bus voltage amplitude and bus voltage angle belong to the system state variable set, and the target value is optimized by adjusting the control variables, that is, the actual power output of each generator set. Equation (2) represents the system power balance, including general UC constraints and system rotation (3) and operation (4) reserve requirements, actual generation capacity restrictions (5), minimum opening (6) and closing (7) time restrictions, and generator set ramp-up (8) and ramp-down (9) restrictions; considering the transmission network constraints, equation (10) represents the transmission flow, while inequalities (11) and (12) enforce transmission flow restrictions and bus voltage restrictions.

4. The unit commitment optimization acceleration method considering safety constraints according to claim 1 is characterized in that: According to the data characteristics of the established SCUC scenario, extracting, sending signals and collecting training data through the simulation calculation results of the SCUC model include the following steps: Selecting Data Features for SCUC In the SCUC problem, the load variation trend is repetitive and small. Analysis shows that the system network topology and line parameters are predetermined and difficult to change. The system network topology and line parameters are considered as known fixed values, while the user load demand is regarded as an independent variable and a variable input parameter. Therefore, the user load demand is selected as the data feature label of the SCUC problem, and a parameter optimization formula is introduced to establish a supervised learning task. Abstract and parameterize the data features of the SCUC model: The parameterization formula of SCUC is as follows: in f i (b,x)=w i x(j|x(j)=I i,t )p(j|p(j)=Load i,t ) (14) The constraint expression is: g k (x,p)≤0,k=1,...,m (15) Where x is the vector of all integer and continuous decision variables in SCUC, p is the concatenated vector of load demands p = {Systemload}, g k (x,p) represents the kth constraint of SCUC, and m is the number of constraints.

5. The unit commitment optimization acceleration method considering safety constraints according to claim 4 is characterized in that: According to the parameterized formula of SCUC, the SCUC problem is determined and solved by giving the input parameter p.

6. The unit commitment optimization acceleration method considering safety constraints according to claim 1 is characterized in that: The process of learning and identifying two data labels includes: Based on the input parameters, the general solver Gurobi is used to calculate and collect training data to form a network learning memory and collect n input parameter samples P. n ={p1,...,p n }, through the GCN acceleration strategy, output strategy s(p i ) corresponds to the value with parameter p i ∈P N The parameterized MILP problem is the optimal decision value and tight constraint index of the binary variables, Optimizing 0-1 variables One solution to the SCUC problem is to find the optimal start and stop states for the units and then solve the day-ahead and real-time generation scheduling problems. Therefore, we only need to obtain the 0-1 state of the I variable for each generator. Therefore, let the optimal value of the binary variable be expressed as x * (p i )=(I * ) (17) Finding tight constraints Define optimality tight constraints as another acceleration metric, which means that some inequality constraints are equal to each other to achieve optimality. Such inequality constraints are identified as tight constraints and retained in the acceleration strategy. where x*(p i ) is the parameter p i The optimal solution under , the number of tight constraints is less than the total number of constraints, which means that there are a lot of redundant constraints in the original SCUC problem. Then, through the learning strategy Ignoring a large number of redundant constraints, the acceleration strategy is implemented. The acceleration model for the SCUC optimization problem for GCN learning is as follows: where x 0-1 Represents a binary variable.

7. The unit commitment optimization acceleration method considering safety constraints according to claim 1 is characterized in that: The process of training the acceleration strategy includes the following steps: Exploring and pruning accelerated branches Construct a multi-class classification problem from sample parameters to acceleration strategies. The sample set contains the optimal acceleration strategies corresponding to many load scenarios. In order to select the optimal acceleration strategy under a certain sample parameter, the acceleration strategy identification of the SCUC problem is regarded as a multi-class classification task in supervised learning. The purpose of this multi-class classification task is to select the optimal acceleration strategy under a given sample parameter p. i Predict the corresponding acceleration strategy under the circumstances; Suppose that given n independent samples P n ={p1,...,p n } to find m unique strategies S(P n )={s1,...,s m }, then new policies will emerge when continuously sampling data, so a good Turing estimator is used to evaluate the probability of discovering an unseen policy, encountering parameters p corresponding to the unseen policy n+1 The probability satisfies the following expression with a confidence level of 1-β: Where G = n1 / n, the number of different strategies corresponding to n1 only appears once, and the coefficient is Learning acceleration strategies using multi-class classifiers Based on the analysis of the parameter SCUC problem, the input and output parameters used in the classification task are as follows: p={Systemload} (21) s(p)={value of binary variables,index of tight constraints} (22) During the classification process, s i is the hot code used to input parameter p i Rewrite the labels of the policy classes; graph neural networks and decision tree ensembles are combined into a multi-class classifier; Graph Convolutional Neural Network Learning Strategies The algorithmic functionality of the GCN model consists of L layers, defined as follows: where s i is the prediction strategy class, h l It is the lth hidden layer in the neural network. Since the data transmitted in the graph neural network must be represented by the structure of the graph, we must first start from the parameter p i To node feature X i The mapping is represented by h0: X i =h0(p i ) (24) The characteristics of each node X i Contains two dimensions, representing the node load and fault indicator respectively. If the node adjacent line fails, the fault indicator is 1, otherwise it is 0. The components of each hidden layer are y l =h l (y) l-1 ,A)=σ l (Yes l-1 W l +b l ),l=1,...,L (25) where y l is the data obtained through layer operations, A is the adjacency matrix of the power network, W l and b l is the network weight parameter, the input layer is defined as l = 1, the output layer is defined as l = L, so y0 = X i ,y L =s i , each layer uses parameter W l and b l Perform affine transformations; In addition, the activation function σ l It is integrated into the nonlinear model and the activation function ReLU, that is, the linear rectification function, is introduced as follows: σ l (x)=max(x,0),l=1,...,L-1 (26) Among them, the max operator is executed bitwise, and the ReLU operator shows higher performance and can avoid the vanishing gradient problem that occurs in the usual sigmoid function. The output layer is set to the Softmax function to provide a standardized sorting between acceleration strategies and evaluate the possibility of them becoming the correct strategy. The Softmax function is very common in multi-classification problems due to its smoothness and probabilistic interpretation. The selection layer of the acceleration strategy, namely σ L Expressed as During the training of the GCN network, the cross entropy loss function is applied. The training process uses the stochastic gradient descent method to obtain the loss function derivative and trains based on the back-propagation rule.

8. The unit commitment optimization acceleration method considering safety constraints according to claim 1 is characterized in that: The process of evaluating the training results includes: The acceleration strategy of the GCN classifier is used to transform the SCUC problem of the MILP into the SCUC problem of the linear programming LP. Then, the KKT condition is used to calculate the optimal solution of the transformed SCUC model. If the solution under the acceleration strategy is feasible, it is output as the solution to the SCUC problem. If the acceleration strategy encounters an infeasible problem, the predicted data is set for pre-solution, and the traditional solver Gurobi is used to calculate and output the optimization result of the large-scale SCUC problem.

9. A device, characterized in that include: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the safety-constrained unit combination optimization acceleration method as described in any one of claims 1 to 8.

10. A storage medium containing computer-executable instructions, characterized in that: The computer executable instructions are used to execute the safety constraint-considered unit commitment optimization acceleration method as described in any one of claims 1 to 8 when executed by a computer processor.

Images