Data-driven enhanced optimization solving method for security constraint unit commitment problem

Abstract

The invention discloses a data-driven enhanced optimization solving method for a security constraint unit commitment problem. The method comprises the following steps: constructing a security constraint unit commitment problem model of a power system; historical operation load data of a power system is collected, a confidence data set is constructed by using a Turing statistical method, and the confidence data set is divided into a training set and a verification set according to a proportion; constructing a neural network model, and inputting the training set and the verification set for network training; new load data of a power system is used as input, a multi-layer neural network is used as a classifier, and a logic strategy of unit operation is predicted online; by eliminating integer variables, an SCUC problem is converted into a continuous convex optimization problem, and by sorting a plurality of candidate integer solutions, feasibility of the model is simplified and optimized, and rapid online solution of the method is realized. According to the method, a high-quality approximate optimal solution can be provided, the solving speed is high, and the expandability is high. Not only can efficient power system scheduling optimization be realized, but also the peak regulation capability and the emergency response capability of the power grid can be enhanced.

Description

Technical Field

[0001] The invention relates to a fast solution method for a safety constrained unit commitment (SCUC) problem, in particular to a data-driven enhanced optimization solution method for the safety constrained unit commitment problem. Background Art

[0002] With the booming development of new energy, the power system is undergoing a profound transformation. The core features of this transformation are reflected in the high penetration rate of new energy, the widespread application of power electronic equipment, and the deep coupling of AC and DC power grids. These factors work together to make the operating characteristics of the power system more complex. In order to cope with this complexity, the dispatching decision of the power system needs to be faster and more intelligent. At present, the operation and dispatch of the power system mainly rely on traditional optimization methods, but these methods are stretched when dealing with the growing system complexity. Therefore, the power system has begun to actively embrace digital, networked, and intelligent technologies, especially artificial intelligence technologies, such as machine learning, which play an increasingly important role in many aspects of the power system, including power generation planning, energy dispatching, load forecasting, and risk management. Applying artificial intelligence technology to the optimization of the power system is expected to greatly improve the efficiency and effectiveness of decision-making, showing great application prospects.

[0003] Unit Commitment (UC) is a fundamental and critical task in the daily dispatch of power systems. It meets power demand at the lowest cost while ensuring system safety and efficiency by carefully planning the startup and shutdown of equipment and power generation plans. The UC problem requires flexible response to load fluctuations in the power market under the background of continuous access to renewable energy. However, with the continuous growth of power demand, the transmission and distribution network may approach its carrying limit, and traditional UC models often ignore network security constraints such as transmission capacity and voltage limitations, which may lead to grid congestion and increased operating costs. Therefore, it becomes crucial to construct a unit commitment model that considers network security constraints (security-constrained unit commitment, SCUC).

[0004] From the perspective of mathematical programming, the SCUC problem belongs to the Mixed-Integer Convex Programming (MICP) problem, which is usually solved by optimizing solvers and algorithms such as branch and bound. However, due to the huge scale of generators and transmission networks involved in the power system, the difficulty of solving the MICP problem increases accordingly, and the calculation time increases sharply as the scale of the problem increases, constituting an NP-hard problem. Considering the uncertainty and volatility of renewable energy output, how to obtain a high-quality solution to the MICP problem within a limited time has become a major challenge in power system scheduling.

[0005] At present, some studies have attempted to directly apply machine learning models to end-to-end constrained optimization from unit output to load data. However, these prediction results often fail to meet the actual operating constraints, resulting in the need for additional post-processing steps to adjust infeasible scheduling schemes to feasible areas, which undoubtedly increases the computational burden. In addition, some works use machine learning models to predict the start and stop status of units, but their output dimensions are high and are easily affected by uneven sample distribution, resulting in overfitting and insufficient generalization ability. At the same time, in the training process of machine learning models, existing studies have not fully considered the required sample size, resulting in insufficient coverage of training data. Summary of the invention

[0006] The purpose of the present invention is to provide a new fast optimization solution method for the safety constrained unit commitment (SCUC) problem, which is more intelligent and has a higher accuracy.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] The data-driven enhanced optimization solution method for the safety-constrained unit commitment problem is characterized by comprising the following steps:

[0009] Step 1: Construct a security constrained unit commitment (SCUC) problem model for the power system;

[0010] Step 2: Collect historical operating load data of the power system, construct a credible data set using Turing statistics, and divide the credible data set into a training set and a validation set in proportion;

[0011] Step 3: Build a neural network model and input the training set and validation set for network training;

[0012] Step 4: Use the historical operating load data in the validation set as input, use a multi-layer neural network as a classifier, and predict the logic strategy of the unit operation online;

[0013] Step 5: By eliminating integer variables, the SCUC problem is transformed into a continuous convex optimization problem. By sorting multiple candidate integer solutions, the feasibility of the simplified optimization model is optimized and a fast online solution is achieved.

[0014] In the above technical solution, the present invention has the following beneficial effects:

[0015] This paper proposes a new method that combines machine learning models with traditional optimization techniques, and adds an energy storage system to the original model to further improve the flexibility and reliability of power grid dispatching. The historical data accumulated when solving the safety constrained unit commitment (SCUC) problem is used to extract effective experience in new situations to achieve real-time online solution of power grid dispatch optimization problems. Specifically, we first model the SCUC problem and construct a "prediction + decision-making" optimization framework to convert the operating state changes of the unit into low-dimensional logical strategies and convert the decision-making problem into a classification problem. Then, the Turing statistical method is used to estimate the required sample size to ensure that the probability of encountering a scenario that does not appear during data sampling is controlled at a low level. In response to the problem of model unsolvability that may be caused by the classification result, we sort multiple candidate integer solutions to ensure that the simplified optimization model remains feasible. The results show that this method can quickly find the approximate optimal solution to the SCUC problem, which has significant advantages over other data-driven methods, and performs outstandingly in both computing time and objective function optimization; compared with conventional solvers, the solution speed is increased by more than an order of magnitude.

[0016] The advantage of the present invention is that the fast intelligent dispatch fully utilizes the application of data-driven enhancement technology in the safety constrained unit commitment (SCUC) problem, thereby improving the safety and economy of the power system. The proposed algorithm framework "prediction + decision" can effectively handle mixed integer programming problems and achieve fast online solutions, significantly improving the solution speed and success rate. The online prediction of logical strategies based on a multi-layer neural network classifier avoids complex integer variable processing, ensures high-quality approximate optimal solutions for the optimization results, has strong scalability, and is suitable for dispatching scenarios of various power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0018] Figure 1 A flowchart for implementing the data-driven enhanced optimization solution method provided by the present invention;

[0019] Figure 2 A typical daily load curve of a city in Hubei Province is input into the data set provided by the present invention;

[0020] Figure 3 A data-driven enhancement algorithm framework diagram provided by the present invention;

[0021] Figure 4The optimal scheduling scheme for the generator set and the energy storage battery provided by the present invention; wherein (a) is the output power distribution of the generator set, and (b) is the charge and discharge power and SoC of the energy storage battery;

[0022] Figure 5 A comparison chart of solution success rate and solution time between the data-driven enhanced optimization solution method provided by the present invention and other methods; wherein (a) is a comparison chart of solution success rate, and (b) is a comparison chart of solution time. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0024] like Figure 1-3 As shown, the data-driven enhanced optimization solution method for the safety constraint unit combination problem disclosed in the present invention is as follows: Figure 1 The implementation steps are as follows:

[0025] Step 1: First, construct a model of the safety constrained unit commitment (SCUC) problem of the power system. It is necessary to focus on analyzing the access method of the machine learning algorithm in the mixed integer programming. The present invention puts the machine learning model in front of the optimization solver. In this way, the machine learning model can provide heuristic information for the optimization process, optimize the initial conditions, and improve the efficiency and accuracy of the solution. In order to more accurately simulate the role of the energy storage system in power dispatching, a series of constraints for energy storage batteries are introduced into the model. These constraints include the charging and discharging power constraints and capacity limitations of the energy storage battery to ensure that the model can fully reflect the actual operation of the energy storage system.

[0026] The safety constrained unit commitment (SCUC) problem model is:

[0027] Objective function:

[0028]

[0029] In the formula, F G is the total operating cost of power generation; T represents the total dispatching time; N G Represents the total number of generators; P Gi,t is the output power of the i-th generator in the t-th time interval; U Gi,t represents the operating status of the i-th generator in the t-th time interval; a i , b i 、c i d i 、e i Respectively represent the operating cost coefficient of the i-th generator. Gi,tis the generator startup indicator variable, which is 1 when the i-th generator is started in the t-time interval; ID Gi,t is the shutdown indicator variable of the generator, which is 1 when the i-th generator stops at time t; SU Gi ,SD Gi They represent the start-up and shutdown costs of the generator, P sci,t and P sdi,t Represent the charging and discharging power of the ith battery at time t respectively.

[0030] The constraints include formulas (2)-(13):

[0031]

[0032] Where: N D is the total number of loads, D j,t is the load power of the jth load in time interval t.

[0033]

[0034] Where: is the minimum output of the ith unit, is the maximum output of the ith unit.

[0035] P Gi,t -P Gi,t-1 ≤RU i (4)

[0036] P Gi,t-1 -P Gi,t ≤RD i (5)

[0037] Where: RU i is the output increase ramp rate limit of the i-th unit, RD i is the output decline slope limit of the i-th unit.

[0038]

[0039] IU Gi,τ +ID Gi,τ ≤1 (8)

[0040] Where: .TU i . is the minimum startup time of the i-th unit, TD i is the minimum shutdown time of the ith unit.

[0041]

[0042] Where: represents the power flow change limit of line l, N lIndicates the total number of branches, H Gi represents the output power transfer distribution factor of the i-th unit, H Dj represents the j-th load power transfer distribution factor.

[0043]

[0044]

[0045] Where: and Indicates the charging and discharging status of the energy storage battery. is the battery energy storage state at time, η c and η d are the efficiency of charging and discharging of the energy storage battery, Δt is the time interval, and are the minimum and maximum limits of battery energy storage.

[0046] Step 2: Collect historical load data of the power system, such as 24-hour load data. Figure 2 As shown, the Turing statistical method is used to construct a credible data set to ensure that the training data covers various possible operation scenarios, and the data set is divided into a training set and a validation set in proportion; the Turing statistical method is calculated as follows:

[0047] Assume that 500 independent samples L are extracted from the historical operating load data of the power system. N ={L1,...,L N}, from which 264 different logical strategies S(L N )={S1,...,S M}, then in the next sample L N+1 The probability of obtaining a new strategy is:

[0048]

[0049] Among them, G is the Good-Turing estimator, which is used to estimate the probability of an unseen event under finite samples, expressed as:

[0050]

[0051] N1 represents the number of logical strategies that appear only once, N represents the total number of samples, and the constant Using this estimation method to calculate the unknown strategy S(L N+1 ) has a confidence level of 0.5.

[0052] At a confidence level of 50%, the probability of encountering a new logic strategy does not exceed 50%. A data set is generated and divided into two parts, 90% of which is taken as a training set to train the deep learning framework, and 10% is taken as a validation set to verify the network's classification effect on the unit's operating status.

[0053] Step 3: Build a neural network model and input the training set and validation set for network training.

[0054] The neural network construction method is: construct a five-layer network, including input layer, fully connected layer, softmax layer, classification output layer, where the fully connected layer has 96 hidden units and its activation function is 'RELU'. The specific hyperparameter values ​​of the neural network model are shown in Table 1.

[0055] Table 1. Hyperparameter values ​​of neural network models

[0056]

[0057] The neural network architecture consists of a standard feedforward network. The input layer corresponds to the hourly load of the receiving system, and the number of neurons is T×N D The hidden layer uses the ReLU activation function, expressed as

[0058] x k+1 =ReLU(w k x k +b k ),k=1,...,z-1 (16)

[0059] Among them, x k is the input of the kth layer, w k and b k is the weight and bias of the kth layer. The ReLU activation function introduces nonlinearity, that is, ReLU(x)=max(0,x). The output layer dimension is the strategy dictionary set S D The rank of , and the Softmax activation function is used to normalize the original output of the network, mapping it to the (0,1) interval with a sum of 1, and the probability of each strategy in the dictionary is obtained as

[0060]

[0061] In the formula, x z,i Represents the output of the i-th neuron in the last hidden layer. For sample i, its true class label is y i ∈0,1,...n,; the category probability predicted by the model is in Indicates the probability that the model predicts that sample i belongs to the jth class. The class with the highest probability is selected as the predicted label value output by the network.

[0062] The training of the neural network model uses the cross entropy loss function to measure the difference between the predicted output and the actual class label y i The error between

[0063]

[0064] Among them, m is the number of samples in the batch, y ij =1 means that the output classification of sample i is the jth class, otherwise y ij =0. The Adam algorithm is used in the training process, and the first-order moment estimation and second-order moment estimation of the gradient are used to dynamically adjust the learning rate, which can effectively reduce the oscillation phenomenon caused by the learning rate during the gradient descent process. The adjustment direction of the network parameters is calculated according to the loss function, and the error is transmitted from the output layer back to each hidden layer of the network by back propagation. The network parameters are updated according to formulas (19)-(21):

[0065]

[0066] Among them, φ k-1 and φ k is the neural network parameter value before and after the update, η is the learning rate, e is the smoothing parameter, m k and m k is the weighted average of the gradient before and after error correction, v k and is the weighted bias variance of the gradient before and after error correction, and β1 and β2 are attenuation factors. Repeat the above steps to gradually adjust the parameters of the neural network until the set maximum number of iterations is reached.

[0067] Step 4: Take the historical operating load data in the validation set as input, use a multi-layer neural network as a classifier, and predict the logical strategy of the unit operation online.

[0068] The softmax layer takes the output of the fully connected layer in the network as the input vector and outputs the probability distribution of the unit operation status logic strategy, as follows:

[0069] ξ1=Softmax(θ d x1+b d ) (twenty two)

[0070] Where x1 represents the output of the fully connected layer, that is, the input of the softmax layer, θ d and b d is the weight matrix and bias matrix of the softmax layer, ξ1 represents the final output of the softmax layer, and softmax is the activation function.

[0071] Step 5: By eliminating integer variables, the SCUC problem is transformed into a continuous convex optimization problem. By sorting multiple candidate integer solutions, the feasibility of the simplified optimization model is ensured, and fast online solution is achieved. The compact form of the SCUC optimization model and the model of the solution algorithm are as follows:

[0072] Objective function:

[0073] minf0(P G ,P sc ,P sd ,U,δ;L)(23)Constraints:

[0074] f i (P G ,P sc ,P sd ,U,δ;L)≤0,i=1,...,K1 (24)

[0075] h j (U,δ;L)≤0,j=1,...,K2 (25)

[0076]

[0077] in is a continuous decision variable, i.e., the active output of the generator and the battery charging and discharging power during the dispatch period. U and δ are binary decision variables, representing the start / stop state and charging / discharging state of the generator. The inequality constraint f i Contains both continuous and discrete variables, h j Include only discrete variables.

[0078] The binary decision variables U and δ are complex factors in solving the optimization model. If U and δ take fixed values, the original problem can be simplified to a convex optimization problem containing only continuous variables P, that is:

[0079] Objective function:

[0080] minf0(P G ,P sc ,P sd ,U * ,δ * ; L) (27) Constraints:

[0081] f i (P G ,P sc ,P sd ,U * ,δ * ; L)≤0,i=1,...,K1 (28)

[0082] By identifying the optimal unit start-stop plan U for the SCUC problem* , δ * , the optimization problem formulas (23)–(26) can be transformed into the quadratic programming problem of formulas (27)–(28). Quadratic programming is a special nonlinear programming problem and a standard convex optimization problem that can be solved quickly.

[0083] Finally, the optimal binary decision U * , δ * Take them as candidate logic strategies in turn to solve the QP problem. If the problem is feasible, the algorithm terminates and outputs the optimal solution. Otherwise, select the next candidate logic strategy and repeat the above steps. Finally, the optimal scheduling plan for the generator set and the energy storage battery is used as the solution to the SCUC problem. Figure 4 .

[0084] Depend on Figure 4 It can be seen that when the load is low, the lower-cost unit 1 and the highest-maximum-output unit 3 are given priority to bear the load. The switching time limit of unit 2 is more flexible, which allows it to be quickly adjusted according to load fluctuations, so that unit 2 can be started to provide power support during peak load periods.

[0085] Figure 5 The box plots of the four solvers (Gurobi, the proposed method (LO), Regressor and KNN) are compared in terms of solution time distribution and solution success rate. The solution time of the conventional solver Gurobi is 10 -0.5 s, Gurobi's solution time is relatively stable but longer. LO's median solution time is about 10 -2.5 s, the solution time fluctuates greatly, which indicates that LO may take longer to complete the solution task in some cases. Reg has the shortest solution time, with a median of about 10 -2.8 s, however, the success rate of Reg is only 44%, which is much lower than other solvers. Although Reg is stable and fast in solving time, its low success rate significantly affects the overall performance and greatly reduces its reliability in practical applications. The solving time of KNN is close to that of LO, with a median of about 10 -2.5 s, but the success rate of solving is only 66%.

[0086] It can be seen that the method of the present invention performs well in terms of solution time and success rate and is a reliable solver. Compared with the traditional method, the method proposed by the present invention has a two-order-of-magnitude improvement in solution time, achieving rapid problem solving.

[0087] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A data-driven enhanced optimization solution method for the safety-constrained unit commitment problem, characterized in that: The following steps are involved: Step 1: Construct a security constrained unit commitment (SCUC) problem model for the power system; Step 2: Collect historical operating load data of the power system, construct a credible data set using Turing statistics, and divide the credible data set into a training set and a validation set in proportion; Step 3: Build a neural network model and input the training set and validation set for network training; Step 4: Use the historical operating load data in the validation set as input, use a multi-layer neural network as a classifier, and predict the logic strategy of the unit operation online; Step 5: By eliminating integer variables, the SCUC problem is transformed into a continuous convex optimization problem. By sorting multiple candidate integer solutions, the feasibility of the simplified optimization model is optimized and a fast online solution is achieved.

2. The data-driven enhanced optimization solution method for the safety-constrained unit commitment problem according to claim 1 is characterized in that: In the step 1, in the safety constrained unit commitment (SCUC) problem model, the constraints include at least the charging and discharging power constraints and capacity limitations of the energy storage battery.

3. The data-driven enhanced optimization solution method for the safety-constrained unit commitment problem according to claim 2 is characterized in that: The objective function of the safety constrained unit commitment (SCUC) problem model in step 1 is expressed as follows: In the formula, F G is the total operating cost of power generation; T represents the total dispatching time; N G Represents the total number of generators; P Gi,t is the output power of the i-th generator in the t-th time interval; U Gi,t represents the operating status of the i-th generator in the t-th time interval; a i 、b i 、c i d i 、e i Represent the operating cost coefficient of the i-th generator. Gi,t is the generator startup indicator variable, which is 1 when the i-th generator is started in the t time interval; ID Gi,t is the shutdown indicator variable of the generator, which is 1 when the i-th generator stops in the time interval t; SU Gi ,SD Gi Respectively represent the start-up and shutdown costs of the generator; P sci,t and P sdi,t Respectively represent the charging and discharging power of the ith battery in the time interval t; The constraints of the safety constrained unit commitment (SCUC) problem model are expressed as follows: Where N D is the total number of loads, D j,t is the load power of the jth load in time interval t; In the formula, is the minimum output of the ith unit, is the maximum output of the i-th unit; P Git -P Git-1 ≤RU i (4) P Gi,t-1 -P Gi,t ≤RD i (5) In the formula, RU i is the output increase ramp rate limit of the i-th unit, RD i is the output decline rate limit of the i-th unit; yl Gi,τ +ID Gi,τ ≤1 (8) In the formula, TU i is the minimum startup time of the i-th unit, TD i is the minimum shutdown time of the i-th unit; In the formula, represents the power flow change limit of line l, N l Indicates the total number of branches, H Gi represents the output power transfer distribution factor of the i-th unit, H Dj represents the jth load power transfer distribution factor; In the formula, and Indicates the charging and discharging status of the energy storage battery. is the battery energy storage state at time, η c and η d is the efficiency of charging and discharging of the energy storage battery, Δt is the time interval, and are the minimum and maximum limits of battery energy storage.

4. The data-driven enhanced optimization solution method for the safety-constrained unit commitment problem according to claim 1 is characterized in that: The Turing statistics method in step 2 includes: Set up N independent samples L obtained from the historical operating load data of the power system N ={L1,...,L N }, from which we get M different logical strategies S(L N )={S1,...,S M }, then in the next sample L N+1 The probability of obtaining a new strategy is: In the formula, G is the Good-Turing estimator, which is used to estimate the probability of an unseen event under finite samples, and is expressed as follows: N1 represents the number of logical strategies that appear only once, N represents the total number of samples, and the constant Using Turing statistics method to calculate the unknown strategy S(L N+1 ) has a confidence level of 1-β.

5. The data-driven enhanced optimization solution method for the safety-constrained unit commitment problem according to claim 1 is characterized in that: The step 2 divides the credible data set into a training set and a validation set, wherein η% is taken as the training set to train the deep learning framework, and (1-η%) is taken as the validation set to verify the classification effect of the network on the unit operation status.

6. The data-driven enhanced optimization solution method for the safety-constrained unit commitment problem according to claim 1 is characterized in that: The method for constructing a neural network model in step 3 includes: constructing a five-layer network, including an input layer, a fully connected layer, a softmax layer, and a classification output layer, wherein the fully connected layer has 96 hidden units and its activation function is 'RELU'.

7. The data-driven enhanced optimization solution method for the safety-constrained unit commitment problem according to claim 6 is characterized in that: In step 4, the softmax layer uses the output of the fully connected layer as the input vector, and the probability distribution of the output unit operation state logic strategy is expressed as follows: ξ1=Softmax(θ d ·x1+b d ) (16) Where x1 represents the output of the fully connected layer and is also the input of the softmax layer, θ d and b d are the weight matrix and bias matrix of the softmax layer respectively, ξ1 is the final output of the softmax layer, and softmax is the activation function.

8. The data-driven enhanced optimization solution method for the safety-constrained unit commitment problem according to claim 1 is characterized in that: In step 5, the SCUC problem can be transformed into a continuous convex optimization problem and solved; The compact form of the SCUC optimization model is as follows: Among them, the objective function is: minf0(P G ,P sc ,P sd ,U,δ;L)(17) where the constraints are expressed as follows (18)-(20): f i (P G ,P sc ,P sd ,U,δ;L)≤0,i=1,...,K1 (18) h j (U,δ;L)≤0,j=1,...,K2 (19) In the formula is a continuous decision variable, i.e., the active output of the generator and the battery charging and discharging power during the dispatch period. U and δ are binary decision variables, representing the start / stop state and charging / discharging state of the generator. The inequality constraint f i Contains both continuous and discrete variables, h j Contains only discrete variables; The binary decision variables U and δ are set to fixed values. The compact form of the SCUC optimization model expressed by equations (17)-(20) can be simplified to a convex optimization problem containing only continuous variables P. The objective function is expressed as follows: minf0(P G ,P sc ,P sd ,U * ,δ * ; L)(21) The constraint condition is expressed as follows: f i (P G ,P sc ,P sd ,AT * ,δ * ;L)≤0,i=1,...,K1 (22) The optimal binary decision U * , δ * The candidate logical strategies are selected in turn to solve the quadratic programming problem. If the problem is feasible, the algorithm terminates and outputs the optimal solution; otherwise, the next candidate logical strategy is repeatedly selected to solve the quadratic programming problem until the problem is feasible.

Images