A large-scale multi-objective unit maintenance optimization method

Abstract

The application provides a large-scale multi-target unit maintenance optimization method, comprising the following steps: establishing a multi-target unit maintenance optimization model; based on an intelligent random sampling method, N feasible solutions satisfying the multi-target unit maintenance optimization model are randomly extracted from a solution space; a rough evaluation model based on an MLP-Bi-LSTM stacked network is established, a traversal set is traversed, and the N feasible solutions are rough evaluated one by one; based on the rough evaluation result, the N feasible solutions are sorted and layered, an OPC curve is drawn according to the comparison result, and an improved DTW method is used to determine the type of the established model; based on the type of the determined multi-target unit maintenance optimization problem and the corresponding rules, a selected set S requiring accurate simulation is determined; the S set is sorted and layered again, a Pareto solution set is obtained, and a suitable solution is selected as an optimal unit maintenance scheme. The method provided by the application can guarantee that a good solution is solved with a high enough probability, and is very suitable for solving unit maintenance optimization problems in engineering practice.

Description

Technical Field

[0001] The present invention relates to power unit maintenance, and in particular to a large-scale multi-objective unit maintenance optimization method. Background Art

[0002] Unit maintenance is a traditional planning issue in power systems. A well-designed unit maintenance plan can not only extend unit lifespan but also preemptively eliminate potential safety hazards such as system downtime caused by unit failures, thus supporting system scheduling and planning. In reality, my country's southwest region is rich in hydropower resources, and a relatively complete and mature cascade hydropower station project has been built, with construction of other cascades also underway. Theoretically, the unit maintenance optimization problem is essentially a mixed-integer nonlinear optimization problem with complex constraints, belonging to the field of optimization. Exploring efficient methods for solving optimization problems is of theoretical significance.

[0003] Regarding the unit maintenance optimization problem, the traditional maintenance optimization model that only considers single objectives such as maintenance costs or operating costs can no longer meet the needs of power grid operation. Taking into account multiple optimization objectives has become the research focus of the maintenance optimization model.

[0004] In recent years, the optimization objectives considered in multi-objective maintenance optimization problems primarily fall into two categories: system economic objectives, such as maintenance costs, generator revenue, and system power purchase costs; and system stability objectives, such as system reserve ratio and power shortage expectations. Methods for solving multi-objective maintenance optimization problems can be broadly categorized into traditional optimization methods and improved intelligent algorithms. Traditional optimization methods simplify the multi-objective maintenance optimization problem into a single-objective problem. These methods primarily include constraint methods, dictionary sorting methods, min-max methods, NNC methods, and linear weighting methods. These methods rely on practical scenarios and expert experience to select appropriate parameters for multiple objectives. Consequently, weight determination is often difficult, and multi-objective trade-offs are often inadequate. Improved intelligent algorithms directly solve the multi-objective maintenance optimization problem, iteratively updating the Pareto frontier solution set. These algorithms primarily include differential evolution algorithms, ant colony algorithms, particle swarm algorithms, and genetic algorithms. However, due to their inherent limitations, improved intelligent algorithms often fall into local optima when solving multi-objective maintenance optimization problems. In addition, in actual engineering applications, the scale of multi-objective maintenance optimization problems is often large. For example, a regional power grid contains 86 units, 202 load nodes, and 1,485 lines. In the maintenance optimization model established based on the actual data of the regional power grid, there are 30,785 continuous variables and 26,055 0-1 integer variables. The entire optimization model is large in scale, has a large number of variables, and has high computational complexity. At this time, the above-mentioned solution methods all have problems such as computational complexity and difficulty in convergence.

[0005] Patent document CN113837494A discloses a multi-objective unit maintenance scheduling optimization method and device. This method takes the minimum standard deviation of system reliability indicators, the minimum regional carbon emission cost, and the minimum total system cost as multiple objectives, and takes the safe operation of the unit as a constraint condition, to establish a multi-objective unit maintenance optimization scheduling model; then a multi-objective particle swarm algorithm is used to solve the unit maintenance optimization scheduling model and optimize the Pareto optimal solution set; finally, a method of sorting close to the ideal solution is used to select the final scheduling solution. This method introduces a carbon trading mechanism into unit maintenance scheduling, which can not only reduce the carbon emissions of the power system to a certain extent, but also ensure the safe and economical operation of the power system. However, this method still uses the ant colony algorithm, and still has problems such as complex calculations and difficulty in convergence. Summary of the Invention

[0006] In order to solve at least one technical problem existing in the above background technology, the present invention provides a large-scale multi-objective unit maintenance optimization method.

[0007] To achieve the above object, the technical solution of the present invention is:

[0008] A large-scale multi-objective unit maintenance optimization method, the method comprising:

[0009] A multi-objective unit maintenance optimization model is established based on unit type, maintenance resources, maintenance optimization objectives, and system operation control constraints;

[0010] Based on the intelligent random sampling method, N feasible solutions that meet the multi-objective unit maintenance optimization model are randomly extracted from the solution space to form a representation set Θ N ;

[0011] Establish a rough evaluation model based on the MLP-Bi-LSTM stacked network, traverse the representation set, and perform rough evaluation on N feasible solutions one by one;

[0012] Based on the rough evaluation results, N feasible solutions are ranked and layered, and the OPC curve is drawn according to the comparison results. The improved DTW method is used to evaluate the type of multi-objective unit maintenance optimization problem.

[0013] Based on the determined type of the multi-objective unit maintenance optimization problem and the corresponding rules, determine the selected set S that needs to be accurately simulated;

[0014] The S set is sorted and stratified again to obtain the Pareto solution set, and the appropriate solution is selected as the optimal unit maintenance plan.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1) The intelligent random sampling technique proposed in this invention boasts high efficiency, rapid solution speed, and strong robustness, enabling it to rapidly extract a specified number of feasible solutions from a vast solution space. Its computational efficiency improves significantly as the number of samples increases. This intelligent random sampling technique is not only a key step in the proposed method but can also be applied to other sampling problems.

[0017] 2) The MLP-Bi-LSTM stacked network proposed for the first time in this paper has strong nonlinear fitting capabilities and can accurately capture the hidden laws of "time series sample + non-time series sample" data.

[0018] 3) The present invention proposes a rough evaluation model suitable for multi-objective unit maintenance optimization problems. The rough evaluation model is based on the MLP-Bi-LSTM stacked network and can quickly and accurately evaluate the representation set Θ N All feasible solutions of the proposed method provide a more accurate data source for the evaluation of sorted hierarchical OPC curve types.

[0019] 4) This paper proposes an improved DTW method for assessing the type of multi-objective unit maintenance optimization problem. Compared to the traditional DTW method, this improved DTW method demonstrates superiority in curve similarity comparison, enabling a more accurate assessment of the type of the multi-objective unit maintenance optimization model established in this paper, providing support for the next step of the proposed method. This improved DTW method can also be applied to other time series matching and curve identification problems.

[0020] 5) This paper proposes an improved vector-order optimization method for finding satisfactory solutions to large-scale multi-objective optimization problems. Compared with traditional vector-order optimization methods, this method is computationally efficient. Compared with other multi-objective methods such as the NNC method and the ε-constraint method, this method demonstrates ease of solution and high quality, ensuring a relatively fast and satisfactory solution within the engineering tolerance. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Implement a flow chart for the intelligent random sampling method;

[0022] Figure 2 This is the combined data structure diagram of "time series samples + non-time series samples";

[0023] Figure 3 This is the MLP-Bi-LSTM stacked network structure diagram;

[0024] Figure 4 This is the application diagram of the MLP-Bi-LSTM stacked network;

[0025] Figure 5 There are three standard OPC curves;

[0026] Figure 6 Figure for DTW-based OPC curve evaluation method;

[0027] Figure 7 Figure for path coordinate updating;

[0028] Figure 8 Figure for overall implementation flow of large-scale multi-objective unit maintenance optimization method provided by the embodiment of the application. DETAILED DESCRIPTION

[0029] Embodiment:

[0030] The technical solutions of the application will be further described below with reference to the drawings and embodiments.

[0031] The large-scale multi-objective unit maintenance optimization method provided by the embodiment mainly includes the following steps:

[0032] 101. Establishing a multi-objective unit maintenance optimization model according to the unit type, maintenance resource, maintenance optimization target and system operation control constraint;

[0033] 102. Randomly extracting N feasible solutions satisfying the multi-objective unit maintenance optimization model from the solution space based on an intelligent random sampling method to form a representation set Θ N ;

[0034] 103. Establishing a rough evaluation model based on a MLP-Bi-LSTM stacked network, traversing the representation set and performing rough evaluation on the N feasible solutions one by one;

[0035] 104. Based on the rough evaluation result, sorting and layering the N feasible solutions, drawing an OPC curve according to the comparison result, and evaluating the multi-objective unit maintenance optimization problem type by using an improved DTW method;

[0036] 105. Based on the type of the determined multi-objective unit maintenance optimization problem and the corresponding rule, determining a selected set S that needs to be accurately simulated;

[0037] 106. Sorting and layering the S set again to obtain a Pareto solution set, and selecting a suitable solution as an optimal unit maintenance scheme.

[0038] Specifically, in step 101, the multi-objective unit maintenance optimization model comprehensively considers the cascade hydropower coupling constraint, maintenance constraint, system operation control constraint and the like, takes the system safety and economy as the optimization target, effectively solves the problem of compiling a hydropower unit joint maintenance plan, and the model is relatively comprehensive and is convenient for engineering application.

[0039] In the actual compilation of annual unit maintenance plans, dispatch agencies usually need to consider multiple optimization objectives, and these multiple optimization objectives are mutually exclusive. The unit maintenance optimization model established in this paper takes into account both the economic and safety of system operation. Three objective functions are selected as the optimization objectives: annual maintenance cost, annual operating cost, and annual maintenance standby equivalent utility. The specific objectives are as follows:

[0040] 1) Annual maintenance costs

[0041] Maintenance costs are the most direct economic indicator of the quality of a unit maintenance plan and are a crucial factor that cannot be ignored when compiling a maintenance plan. In particular, maintenance costs are a key objective of primary concern for power generators in market environments. Due to the limitations of human and material resources, unit maintenance costs per unit of capacity fluctuate throughout the year. The corresponding objective function for annual maintenance costs is expressed as follows:

[0042]

[0043] Where F1 is the annual maintenance cost optimization target of the maintenance model; T is the maintenance period (the present invention takes one year as the maintenance cycle, T = 366); i is the unit number; Z is the total number of all units, Z = N + M, N and M are the number of thermal power units and hydropower units respectively; P max,i is the maximum output of unit i; b i (t) is the maintenance 0-1 variable of unit i; b i (t) = 1 means that unit i is under maintenance at time t, b i (t)=0 means that unit i is in a non-maintenance state at time t.

[0044] 2) Annual operating costs

[0045] In an environment where dispatching agencies coordinate power generation plans, annual system operating costs are a key indicator of system economic efficiency. Typically, unit fuel costs are used to represent system operating costs. For thermal power units, this is represented by a quadratic function relationship between unit coal consumption and generated power, derived using the least squares method. This is described below.

[0046]

[0047] Where, F2 is the annual operating cost optimization target of the maintenance model; P i (t) is the output of unit i at time t; c0, c1, and c2 are the operating cost coefficients of thermal power units.

[0048] 3) Annual maintenance standby equivalent utility

[0049] Optimization objectives F1 and F2 are both indicators that reflect the economic efficiency of system operation. However, system safety is the primary factor that must be considered when formulating maintenance plans. A system with good safety not only ensures stable operation of the power grid, but also supports scheduling and planning, improving the economic efficiency of system operation. Based on the concept of system spare capacity, this paper constructs a maintenance reserve equivalent utility objective that reflects system safety. The specific expression is as follows:

[0050]

[0051] Where F3 is the maintenance standby equivalent utility optimization objective for the maintenance model; Dt is the system's maximum daily load (MW); and the coefficient β is used to convert the value of this objective function to the equivalent dimensions of the previous two optimization objective functions. This value can be determined based on historical operating experience. As can be seen from the above formula, the greater the maintenance standby equivalent utility, the higher the system security.

[0052] The upstream and downstream power stations of a large-scale cascade hydropower station group have a significant hydroelectric coupling relationship. Specifically, the operating status of the upstream power station and the upstream water inflow will directly affect the operation of the downstream power station. If the structure of the hydropower station group and the complex coupling relationship are carefully considered, the number of dimensions and constraints of the cascade hydropower coupling model will be greatly increased. Considering the requirements of practical engineering for model universality, this embodiment, based on existing research, proposes cascade hydropower coupling model constraints applicable to practical engineering. The specific process of establishing the model constraints is as follows:

[0053] 1) Reservoir capacity balance equation

[0054] The reservoir capacity balance equation is the most basic equation that reflects the coupling relationship between upstream and downstream cascade power stations. When the influence of water stagnation is simplified, the reservoir capacity balance equation can be expressed as follows:

[0055] The storage capacity change rate of reservoir k at the current time t is equal to the sum of the natural water inflow at the current time, the change in the power generation flow of the current reservoir k and the upstream reservoir k-1, and the change in the abandoned water, that is:

[0056] Δv k,Δt / Δt=Δh Δk,t +Δs Δk,t +n k,t (4)

[0057] Δv k,Δt =v k,t -v k,t-1 (5)

[0058] Δh Δk,t =h k-1,t -h k,t (6)

[0059] Δs Δk,t=s k-1,t -s k,t (7)

[0060] Where: Δv k,Δt is the storage capacity change of reservoir k; v k,t is the storage capacity of reservoir k at time t; v k,t-1 is the storage capacity of reservoir k at time t-1; Δh Δk,t h is the change in power generation flow between the current reservoir k and the upstream reservoir k-1; k,t is the power generation flow of reservoir k at time t; h k-1,t is the power generation flow of reservoir k at time t-1; n k,t is the natural water flow of reservoir k at time t; Δs Δk,t is the change in water discharge between the current reservoir k and the upstream reservoir k-1; s k,t is the water discharge of reservoir k at time t; s k-1,t is the water discharged from reservoir k-1 at time t.

[0061] For the leading power station of the cascade reservoir, due to the lack of upstream power stations, the reservoir capacity balance equation is revised to:

[0062] Δv 1,t / Δt=n 1,t -h 1,t -s 1,t (8)

[0063] 2) Hydropower unit output characteristic equation

[0064] The power generation of cascade hydropower units is inseparable from the power generation flow. According to the actual project, the output characteristic equation of the hydropower unit is modeled as a linear function of the hydropower unit output and the power generation flow, that is:

[0065] P i (t) = γ i h i,t (9)

[0066] Where h i,t is the power generation flow of hydropower unit i at time t, i∈Ω M ; γ i is the power conversion coefficient of hydropower unit i, which is generally obtained by fitting the historical power generation flow and historical power generation of cascade hydropower stations.

[0067] 3) Reservoir power flow equation

[0068] The sum of the power generation flow of each unit in a reservoir is the total power generation flow of the reservoir, that is:

[0069]

[0070] Where K is the total number of hydropower units in reservoir k.

[0071] 4) Upper and lower limits of hydropower unit output

[0072] For each hydropower unit, its output at any time must meet the upper and lower output constraints, namely:

[0073] (1-b i (t))P i,min ≤P i (t)≤(1-b i (t))P i,max (11)

[0074] Where, P i,min 、P i,max are the lower and upper limits of the output of hydropower unit i, i∈Ω M ; b i (t) is the 0-1 variable for maintenance of hydropower units, b i (t) = 1 means that hydropower unit i is under maintenance, b i (t)=0 means that hydropower unit i is in a non-maintenance state.

[0075] 5) Upper and lower limits of reservoir capacity

[0076] Affected by hydrological factors, the storage capacity of cascade hydropower plants must be limited within a certain range, which can be described as follows:

[0077] v k,min ≤v k,t ≤v k,max (12)

[0078] Where, v k,min 、v k,max are the lower and upper limits of the storage capacity of reservoir k respectively.

[0079] 6) Upper and lower limit constraints on reservoir power generation flow

[0080] The upper and lower output limits of each unit in the reservoir indirectly affect the change of the reservoir power flow, so that the reservoir power flow is kept within a certain range, that is:

[0081]

[0082] Where h k,min 、h k,max are the lower and upper limits of the power generation flow of reservoir k respectively.

[0083] 7) Calculation equation for discarded water

[0084] When the water volume in the reservoir exceeds the preset range at a certain moment, it is necessary to discard water appropriately to prevent the overflow from affecting the safe operation of the reservoir.

[0085]

[0086] 8) Ecological flow constraints

[0087] In order to ensure the ecological service functions of rivers and maintain or restore the basic structure and functions of river, lake and reservoir ecosystems, rivers must have a minimum flow, namely:

[0088] h k,t +s k,t ≥ecoflow k (15)

[0089] Where, ecoflow k The present invention adopts the Tennant method to determine the ecological flow demand of hydropower station k. The formula shows that hydropower station k needs to discharge at least enough water to meet the ecological flow demand during period t.

[0090] For unit maintenance issues, the constraints that need to be considered mainly include maintenance constraints and system operation constraints.

[0091] 1) Maintenance period constraints

[0092] For any unit i, the total maintenance time should be equal to the scheduled maintenance period, that is:

[0093]

[0094] Where, T cx,i The scheduled maintenance period for unit i.

[0095] 2) Maintenance continuity constraints

[0096] For any unit i, once the maintenance task starts, it cannot be stopped, that is, the unit will always remain in maintenance status within the scheduled maintenance period, that is:

[0097] b i (t)-b i (t-1)≤b i (t+T cx,i -1) (17)

[0098] 3) Upper and lower limits of unit output

[0099] For each unit, its output at any time must meet the upper and lower output constraints, that is,

[0100] (1-b i (t))P i,min ≤P i (t)≤(1-b i (t))P i,max(18a)

[0101] 0≤P Ww (t)≤P′ Ww (t) (18b)

[0102] where P i,min is the maximum power of thermal unit i; P Ww (t) is the scheduled output of wind-solar unit w at time period t; P' Ww (t) is the predicted output value of wind-solar unit w at time period t.

[0103] 4) System power balance constraint

[0104] The system power balance constraint is that at any time of system operation, the sum of the output of all units in the region and the power exchange between the region and the external grid equals the system load in the region, which is specifically expressed as:

[0105]

[0106] where P q (t) is the power exchange between the region and the external grid at time t, and the sign before the cumulative term is - when the external region injects into the region, and vice versa; Q is the total number of tie lines; D t is the daily maximum load in the region, and D i,max is the total system load value.

[0107] 5) System reserve capacity constraint

[0108] To ensure the safe operation of the system, a certain reserve margin should be left at any time, which is expressed as:

[0109]

[0110] where P l,max is the maximum power of unit i; and a is the optimal system reserve coefficient.

[0111] 6) Line power flow constraint

[0112] For any transmission line at any time, the transmission power should be controlled within a certain range:

[0113]

[0114] -P l ≤P l,max (22)

[0115] where G l,i , G l,q , G l,w , G l,dare the active power transfer factors between the transmission line l and conventional unit i (hydro and thermal power units), wind and solar units w, external grid tie lines q, and load d; P l,max is the maximum transmission power of transmission line l.

[0116] The multi-objective unit maintenance optimization problem established above can be abstractly defined as follows:

[0117] minF={F1(x), F2(x), ...F k (x)} (23)

[0118] stg(x)={g1(x), g2(x),…g n (x)}≤0 (24)

[0119] x={x1,x2,…x n}∈X (25)

[0120] The multi-objective unit maintenance optimization problem has k optimization objectives, n constraints, and n decision variables.

[0121] Step 102, i.e. randomly extracting N feasible solutions to construct a representation set is the second step of the method proposed in the present invention. The general random sampling technique is to extract feasible solutions that satisfy the constraint g(x) in a huge solution space. The specific method is to decouple discrete variables and continuous variables, and transform the constraints related to continuous variables so that they are indirectly satisfied. However, with the increase of continuous variables, the decoupling variable operation also becomes complicated and cumbersome, and this method cannot completely and strictly satisfy all constraints. Based on this defect, the present invention proposes an intelligent random sampling method, which forms a classifier Q for screening feasible solutions by establishing a mapping model between solution x and solution feasibility state A(0 / 1). The mapping model of solution x and solution feasibility state A is as follows:

[0122] A=f(x1,x2,…x n ) (26)

[0123] The implementation steps of intelligent random sampling method are as follows: Figure 1 shown.

[0124] Combine Figure 1 It can be seen that the specific implementation steps of the intelligent random sampling method are as follows: First, based on historical data, with the help of machine learning ideas, by adjusting parameters and training the mapping model A=f(x1,x2,…x n ) to obtain the classifier Q. Then, the randomly selected M solutions x are input into the trained classifier Q, so that the feasibility status of these M solutions can be obtained, achieving the effect of quickly extracting feasible solutions from a large number of solutions.

[0125] Step 103, i.e., establishing a rough assessment model is the third step of the method proposed in the present invention. The present invention establishes a rough assessment model based on an MLP-Bi-LSTM stacked network for the multi-objective unit maintenance optimization problem.

[0126] The present invention proposes a rough model based on MLP-Bi-LSTM stacked network to obtain the optimization target value F=[F1,…,F3] from the known N historical maintenance parameter information. T Specifically, the parameter X=[X1,…,X s ] T (daily maximum load, unit maintenance cost, unit operating cost coefficient, natural water supply, unit maintenance initial time, etc.) as input, with daily maintenance cost, daily operating cost, daily maintenance standby equivalent utility F′=[F1′,…,F3′] T As the network output, the mapping relationship F between the training unit maintenance parameters and the optimization target value is i ′=f(X), that is:

[0127]

[0128] During a maintenance planning period, the optimization targets F1 and F2 are obtained by adding up F1′ and F2′ on a daily basis; the optimization target F3 is obtained by adding up F3′ on a daily basis and taking the average.

[0129] Mapping relationship F i The N historical maintenance parameter data (based on days, with a total data volume of N*T) with ′=f(X) have the following characteristics: (1) They can be regarded as maintenance parameter data under N different scenarios of the multi-objective maintenance optimization problem of the present invention (i.e., N parallel samples), and there is no time series relationship between the samples; (2) A single sample is the maintenance parameter data within one year. Due to the limitations of maintenance duration constraints, annual maintenance resource planning, etc., the maintenance parameter data within one year have a time series relationship, that is, the maintenance status of the unit on day t {a1(t), a2(t), ...a Z (t)} is affected by the maintenance status of the unit in the adjacent period {{a1(t-d0),…a Z (t-d0)},...,{a1(t+d0),...a Z Based on the above characteristics, historical maintenance parameter data can be regarded as a combination of "time series samples + non-time series samples". The specific data structure is as follows: Figure 2 shown.

[0130] Mathematically, this data structure can be represented as:

[0131]

[0132] Where d2a2 is the non-time series sample data, d1a1 is the time series sample data, and Da is the total data set; 1,i is the index number of N time series samples; d 2,i is the index number of N non-time series samples.

[0133] For the combination data of "time series samples + non-time series samples", this paper proposes an MLP-Bi-LSTM stacked network to achieve the purpose of accurately fitting the nonlinear mapping model between the combination data of "time series samples + non-time series samples". Its structure diagram is as follows Figure 3 shown.

[0134] The MLP-Bi-LSTM stacked network is constructed by stacking multiple MLP-Bi-LSTM networks, each of which is formed by a cross-combination of MLP and Bi-LSTM. Its core concept is "split data and train in batches." Specifically, the process of training the mapping model using the MLP-Bi-LSTM stacked network is as follows: for each of the N samples (T-2d), the internal data is trained using a bidirectional long short-term memory (Bi-LSTM) network; for the connection data (2d) between the N samples, a multilayer perceptron (MLP) is trained. One training run is performed on each of the N sample data (N*T). The abstract mathematical expression is as follows:

[0135]

[0136] net1(x)=Bi-LSTM(x) (30)

[0137] net2(x)=MLP(x) (31)

[0138] The specific application ideas of the MLP-Bi-LSTM stacking network are as follows Figure 4 shown.

[0139] The objective function value obtained by the MLP-Bi-LSTM stacked network is a rough estimate. Due to the inherent prediction uncertainty of the neural network, this rough estimate may deviate from the precise calculation result obtained by the maintenance optimization model. According to the Kolmogorov complexity principle, this deviation can be considered as normally distributed noise with mean and standard deviation . Therefore, noise is superimposed on the target value obtained by the MLP-Bi-LSTM stacked network to form observations, which allows for more accurate sorting and stratification. That is:

[0140]

[0141]

[0142] Step 104, i.e., evaluating the type of multi-objective unit maintenance problem, is the fourth step of the method proposed in the present invention. After making a rough evaluation of all feasible solutions of the representation set and sorting them into layers with the help of the rough evaluation model, the sorted and layered OPC curve of the multi-objective unit maintenance problem can be drawn with the number of layers x as the horizontal axis and the cumulative number of feasible solutions in the first x layers as the vertical axis. The method proposed in the present invention divides the OPC curve into three types: flat type, steep type, and neutral type, as shown in Figure 2. Figure 5 shown.

[0143] In the present invention, the multi-objective unit maintenance problem is consistent with the OPC curve type. Generally, the improved DTW method is used to evaluate the type of multi-objective unit maintenance optimization problem, that is: the type to which the sorted layered OPC curve belongs is evaluated according to its overall trend and characteristics, and then the type of the multi-objective optimization problem is determined. This evaluation method is highly subjective, and as the number of layers decreases, it becomes more difficult to extract the important features of the sorted layered OPC curve, which can easily lead to problems such as inaccurate evaluation of the sorted layered OPC curve and difficulty in evaluating the curve type. In order to more accurately determine the type of the multi-objective optimization problem, the present invention proposes an OPC curve evaluation method based on the improved DTW method, such as Figure 6 Specifically, the improved DTW method is used to calculate the similarity between the sorted layered OPC curves and the standard OPC curves one by one.

[0144] The improved DTW method proposed in the present invention is derived from DTW, and its basic idea is as follows.

[0145] Improved DTW method basic concepts:

[0146] Assume that the sorted hierarchical OPC curve can be regarded as a sequence of n data Q = {q1, q2, ...q n}, the standard OPC curve can be regarded as a curve formed by m sequence data C={c1,c2,…c m}The curve formed (assuming n>m).

[0147] Define the original distance sequence od between two curves i as follows:

[0148]

[0149] The improved DTW distance value M(i,j) between two curves is defined as follows:

[0150] M(i,j)=|q i -c j |,i=1…n,j=1…m (35)

[0151] Define the cumulative improved DTW distance λ(i,j) of the matrix M(i,j) as:

[0152]

[0153] The above formula shows that the initial cumulative improved DTW distance is defined as The λ(i, j) of the current coordinate is related to the cumulative improved DTW distance of the previous coordinate and the improved DTW distance value M(i, j) of the current coordinate. There are only three update paths from the previous coordinate to the current coordinate (i, j), namely:

[0154]

[0155] The specific implementation steps of the improved DTW method for evaluating the multi-objective unit maintenance optimization problem type are as follows:

[0156] 1) Obtain the improved DTW distance matrix M of the two curves and define the initial coordinates of the path (1, 1);

[0157] 2) Start path finding by calculating the cumulative improved DTW distance λ(i, j). When the coordinate (2, j) is found, that is, when i→i+1, the coordinates and the cumulative improved DTW distance are updated using the following formula:

[0158]

[0159]

[0160] The above formula indicates that when n>m, the horizontal coordinate i is used as the reference. When the horizontal coordinate update step size is 1, that is, when the coordinate is updated from (1,1) to (2,j), the current coordinate and the cumulative improved DTW distance value λ(2,j) are determined based on the relationship between the original distance value od2 of the two coordinates and the cumulative improved DTW distance change value λ(2,j)-λ(1,1). When od2 is larger, the improved DTW update method is considered superior, λ(2,j) is selected as the current cumulative improved DTW distance value, and (2,j) is selected as the current position. When λ(2,j)-λ(1,1) is larger, the original update method is considered superior, od2+M(1,1) is selected as the current cumulative improved DTW distance value, and (2,2) is selected as the current position.

[0161] 3) Continue updating the path coordinates as described in step 2, as follows:

[0162]

[0163]

[0164] In the formula, j0 is the corresponding vertical coordinate j when the horizontal coordinate of the current coordinate first reaches i-1 during path optimization. The above formula means that when the horizontal coordinate update step is 1, the current coordinate (i, j) and the cumulative improved DTW distance value λ(i, j) need to be calculated based on the original distance value od between the two coordinates (i, j) and (i-1, j0). i , the relationship between the cumulative improvement DTW distance change value λ(i,j)-λ(i-1,j0) is determined. i When it is larger, the improved DTW update method is considered to be better, and the calculated λ(i, j) is selected as the current cumulative improved DTW distance value, and (i, j) is selected as the current position, that is, according to Figure 7 The path shown in ① updates the coordinates; when λ(i,j)-λ(i-1,j0) is large, the original update method is considered better and od is selected i +M(i-1,j0) is used as the current cumulative improved DTW distance value, and (i, i) is selected as the current position, that is, according to Figure 7 Update the coordinates of the path ② shown.

[0165] 4) By continuously updating the coordinates and accumulating the improved DTW distance, a path from the initial coordinate (1, 1) to the end coordinate (i, j) is eventually formed. The final calculated λ(n, m) is the shortest path.

[0166] 5) Following steps 1)-4), calculate the shortest distances between the sorted hierarchical OPC curve and the standard flat OPC curve, the standard steep OPC curve, and the standard neutral OPC curve, perform sequence matching on each curve, and calculate the similarity between each pair of curves. Based on the calculated similarity results, determine the type of sorted hierarchical OPC curve that is more appropriate.

[0167] Through the above steps, the type of OPC curve can be obtained, and then the type of multi-objective unit maintenance optimization problem can be determined.

[0168] Step 105, i.e., the establishment of an accurate model is the fifth step of the method proposed in the present invention. After determining the type of the sorting layer OPC curve, the first s layers of the sorting layer are determined according to certain rules, and all feasible solutions in the first s layers constitute the selected set S. Next, it is necessary to accurately calculate all feasible solutions in S, that is, to take all feasible solutions as known conditions and solve F under g(x) that meets the constraints under this feasible solution. Specifically, the k objectives F to be optimized are converted into the same dimension (k=3 in the present invention), and the comprehensive evaluation method is used to determine the weight of each objective. The multi-objective unit maintenance optimization problem is converted into a single-objective optimization problem by weighted summation, that is:

[0169]

[0170] Step 106, i.e. selecting a sufficiently good solution set, is the sixth step of the method proposed in the present invention. The accurately calculated solutions S are again hierarchically sorted, and the Pareto optimal solution set s1 is selected. All solutions in the s1 layer are accurately simulated and calculated, and multiple objective functions R = {F1(x), F2(x), ... F k (x)}. Membership degree λ i Characterize the comprehensive deviation degree of each target value in the i-th solution from the optimal solution of each target in Pareto, and select the membership degree λ i The solution with the largest value is regarded as the optimal solution.

[0171] The membership function λ corresponding to the jth objective function in the i-th non-inferior solution i,j Defined as:

[0172]

[0173] Where, F i,j is the jth objective function value in the i-th non-inferior solution; F j min is the minimum value of the jth objective function in the Pareto optimal solution set; F j max is the maximum value of the jth objective function in the Pareto optimal solution set.

[0174] For any solution i in Pareto, its comprehensive membership λ i As shown in the formula:

[0175]

[0176] Where n obj is the number of targets, n obj =k;n design is the number of solutions in the Pareto equation.

[0177] like Figure 8 As shown in FIG, it is a flowchart of the overall implementation of the large-scale multi-objective unit maintenance optimization method provided by an embodiment of the present invention.

[0178] In summary, compared with the prior art, the present invention has the following technical advantages:

[0179] 1) A joint maintenance optimization model for hydropower and thermal power units is established, taking into account the coupling characteristics of cascade hydropower. This model comprehensively considers the cascade hydropower and power coupling constraints, maintenance constraints, and system operation constraints. With system safety and economy as the optimization goals, it effectively solves the problem of compiling joint maintenance plans for hydropower and thermal power units. The model is relatively comprehensive and easy to apply in engineering projects.

[0180] 2) A new approach to solving large-scale multi-objective maintenance optimization problems is proposed. That is, in the optimization process, instead of focusing on finding the global optimal solution, efforts are made to find a "good enough" solution that meets the actual engineering needs, thus satisfying the pursuit of a satisfactory solution in actual engineering.

[0181] 3) A new method for searching for satisfactory solutions is proposed, which greatly improves the solution efficiency by comparing the advantages and disadvantages of the solutions instead of comparing the exact target values ​​of the solutions.

[0182] 4) An intelligent random sampling method with fast solution speed and strong robustness is proposed, which can ensure the rapid extraction of feasible solutions that meet the constraints of the multi-objective unit maintenance optimization problem in the huge solution space. This intelligent random sampling technology is also applicable to any sampling problem.

[0183] 5) A MLP-Bi-LSTM stacked network is proposed, which has strong nonlinear fitting ability when training "time series samples + non-time series samples" combination data and can accurately capture the hidden patterns between data.

[0184] 6) A rough evaluation model based on MLP-Bi-LSTM stacked network is proposed. The rough evaluation model can quickly and accurately evaluate the representation set Θ N All feasible solutions of are obtained, and the annual maintenance cost, annual operating cost, and annual maintenance standby equivalent utility corresponding to N feasible solutions are obtained.

[0185] 7) We propose an improved DTW method for sequence matching. This method, an improvement on DTW, is suitable for determining the similarity between two sequences of different lengths. Compared to DTW, the improved DTW method updates the coordinates and path lengths in real time when finding the shortest path between two sequences, better capturing the similarity between the sequences. This method is not only suitable for evaluating the type of sorted hierarchical OPC curves, but is also applicable to other time series matching and curve identification problems.

[0186] 8) An evaluation method for the type of multi-objective unit maintenance optimization problem based on improved DTW is proposed, which can accurately evaluate the type of the multi-objective unit maintenance optimization model established by the present invention.

[0187] 9) A method for obtaining satisfactory solutions to large-scale multi-objective unit maintenance optimization problems is proposed. When existing multi-objective optimization methods cannot provide feasible solutions, this method can quickly obtain satisfactory solutions within the allowable error range of the project.

[0188] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made based on the essence of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. A large-scale multi-objective unit maintenance optimization method, characterized in that: The method comprises: Establish a multi-objective unit maintenance optimization model based on unit type, maintenance resources, maintenance optimization objectives, and system operation control constraints; Based on the intelligent random sampling method, N feasible solutions that meet the multi-objective unit maintenance optimization model are randomly extracted from the solution space to form a representation set. Establish a rough evaluation model based on the MLP-Bi-LSTM stacked network, traverse the representation set, and perform rough evaluation on N feasible solutions one by one; Based on the rough evaluation results, N feasible solutions are ranked and layered, and the OPC curve is drawn according to the comparison results. The improved DTW method is used to evaluate the type of multi-objective unit maintenance optimization problem. Based on the determined type of the multi-objective unit maintenance optimization problem and the corresponding rules, determine the selected set S that needs to be accurately simulated; Sort and stratify the S set again to obtain the Pareto solution set, and select the appropriate solution as the optimal unit maintenance plan; The multi-objective unit maintenance optimization model takes system safety and economy as optimization objectives, and considers cascade hydropower coupling constraints of hydrological and runoff characteristics, maintenance constraints, and system operation control constraints; The system safety and economy are optimized as follows: 1) Annual maintenance costs The corresponding objective function of annual maintenance cost is expressed as follows: Where F1 is the annual maintenance cost optimization target of the maintenance model; T is the maintenance period; i is the unit number; Z is the total number of all units, Z = N + M, N and M are the number of thermal power units and hydropower units respectively; P max,i is the maximum output of unit i; b i (t) is the maintenance 0-1 variable of unit i; b i (t) = 1 means that unit i is under maintenance at time t, b i (t) = 0 means that unit i is in a non-maintenance state at time t; 2) Annual operating costs The unit fuel cost is used to represent the system operating cost. For thermal power units, the fuel cost is represented by the quadratic function relationship between the unit coal consumption and the generated power obtained by the least squares method. The specific description is as follows: Where, F2 is the annual operating cost optimization target of the maintenance model; P i (t) is the output of unit i at time t; c0, c1, c2 are the operating cost coefficients of thermal power units; 3) Annual maintenance standby equivalent utility The specific expression is as follows: Where F3 is the maintenance standby equivalent utility optimization objective of the maintenance model; Dt is the system daily maximum load (MW); the coefficient β is used to convert the value of this objective function to the equivalent dimension of the first two optimization objective functions and can be determined based on historical operating experience.

2. The large-scale multi-objective unit maintenance optimization method according to claim 1, characterized in that: The cascade hydropower coupling constraints on hydrological and runoff characteristics include: 1) Reservoir capacity balance equation The reservoir capacity balance equation is expressed as follows: The storage capacity change rate of reservoir k at the current time t is equal to the sum of the natural water inflow at the current time, the change in the power generation flow of the current reservoir k and the upstream reservoir k-1, and the change in the abandoned water, that is: Δv k,Δt / Δt=Δh Δk,t +Δs Δk,t +n k,t (4) Δv k,Δt =v k,t -v k,t-1 (5) Δh Δk,t =h k-1,t -h k,t (6) Δs Δk,t =s k-1,t -s k,t (7) Where: Δv k,Δt is the storage capacity change of reservoir k; v k,t is the storage capacity of reservoir k at time t; v k,t-1 is the storage capacity of reservoir k at time t-1; Δh Δk,t h is the change in power generation flow between the current reservoir k and the upstream reservoir k-1; k,t is the power generation flow of reservoir k at time t; h k-1,t is the power generation flow of reservoir k at time t-1; n k,t is the natural water flow of reservoir k at time t; Δs Δk,t is the change in water discharge between the current reservoir k and the upstream reservoir k-1; s k,t is the water discharge of reservoir k at time t; s k-1,t is the water discharged from reservoir k-1 at time t; For the leading power station of the cascade reservoir, due to the lack of upstream power stations, the reservoir capacity balance equation is revised to: Δv 1,t / Δt=n 1,t -h 1,t -S 1,t (8) 2) Hydropower unit output characteristic equation The power generation of cascade hydropower units is inseparable from the power generation flow. The output characteristic equation of the hydropower unit is modeled as a linear function of the hydropower unit output and the power generation flow, that is: P i (t)=γ i h i,t (9) Where h i,t is the power generation flow of hydropower unit i at time t, i∈Ω M ; γ i is the power conversion coefficient of hydropower unit i, which is generally obtained by fitting the historical power generation flow and historical power generation of cascade hydropower stations; 3) Reservoir power flow equation The sum of the power generation flow of each unit in a reservoir is the total power generation flow of the reservoir, that is: Where K is the total number of hydropower units in reservoir k; 4) Upper and lower limits of hydropower unit output For each hydropower unit, its output at any time must meet the upper and lower output constraints, namely: (1-b i (t))P i,min ≤P i (t)≤(1-b i (t))P i,max (11) Where, P i,min 、P i,max are the lower and upper limits of the output of hydropower unit i, i∈Ω M ; b i (t) is the 0-1 variable for maintenance of hydropower units, b i (t) = 1 means that hydropower unit i is under maintenance, b i (t) = 0 means that hydropower unit i is in a non-maintenance state; 5) Upper and lower limits of reservoir capacity The storage capacity of cascade hydropower plants must be limited within certain limits, which can be described as follows: in k,min ≤in k,t ≤in k,max (12) Where, v k,min 、v k,max are the lower and upper limits of the storage capacity of reservoir k respectively; 6) Upper and lower limit constraints on reservoir power generation flow The upper and lower output limits of each unit in the reservoir indirectly affect the change of the reservoir power flow, so that the reservoir power flow is kept within a certain range, that is: Where h k,min 、h k,max are the lower and upper limits of the power generation flow of reservoir k, respectively; 7) Calculation equation for discarded water When the water volume in a reservoir exceeds the preset range at a certain moment, water needs to be appropriately discarded to prevent water overflow from affecting the safe operation of the reservoir, namely: 8) Ecological flow constraints In order to ensure the ecological service functions of rivers and maintain or restore the basic structure and functions of river, lake and reservoir ecosystems, rivers must have a minimum flow, namely: h k,t +s k,t ≥ecoflow k (15) Where, ecoflow k represents the ecological flow demand of hydropower station k, and the value is determined by the Tennant method. This formula shows that hydropower station k needs to discharge at least the amount of water that meets the ecological flow in period t.

3. The large-scale multi-objective unit maintenance optimization method according to claim 1, characterized in that: The intelligent random sampling method includes: firstly, based on historical data, with the help of machine learning ideas, by adjusting parameters and training the mapping model A=f(x1, x2, ... x n ) to obtain the classifier Q; then the randomly selected M solutions x are input into the trained classifier Q to obtain the feasibility status of these M solutions.

4. The large-scale multi-objective unit maintenance optimization method according to claim 1, characterized in that: The establishment of a rough evaluation model based on an MLP-Bi-LSTM stacked network, traversing the representation set, and performing rough evaluation on N feasible solutions one by one includes: The N historical maintenance parameters X=[X1,…,X s ] T As input, take daily maintenance cost, daily operation cost, daily maintenance standby equivalent utility F′=[F1′,…,F3′] T As the network output, the mapping relationship F between the training unit maintenance parameters and the optimization target value is i ′=f(X), that is: During a maintenance planning period, the optimization targets F1 and F2 are obtained by adding up F1′ and F2′ on a daily basis; the optimization target F3 is obtained by adding up F3′ on a daily basis and taking the average. Mapping relationship F i The N historical maintenance parameter data of ′=f(X) are regarded as a combination of "time series samples + non-time series samples". This combination data is trained using an MLP-Bi-LSTM stacked network pair: for the internal data of each of the N samples, a bidirectional long short-term memory network Bi-LSTM is used for training; for the connection data between the N samples, a multi-layer perceptron MLP is used for training. Traversing the N sample data is considered a training session. The objective function value obtained by the MLP-Bi-LSTM stacked network is a rough evaluation result. Noise is superimposed on the objective function value to form an observation value.

5. The large-scale multi-objective unit maintenance optimization method according to claim 4, characterized in that: The MLP-Bi-LSTM stacked network is formed by stacking multiple MLP-Bi-LSTM networks, and each MLP-Bi-LSTM network is formed by a cross combination of MLP and Bi-LSTM.

6. The large-scale multi-objective unit maintenance optimization method according to claim 1, characterized in that: The sorting and stratifying of N feasible solutions and drawing the OPC curve according to the comparison results include: After making a rough assessment of all feasible solutions of the representation set and ranking them in layers with the help of a rough assessment model, the ranked hierarchical OPC curve of the multi-objective unit maintenance problem is drawn with the number of layers x as the horizontal axis and the cumulative number of feasible solutions in the first x layers as the vertical axis; OPC curves are divided into three types: flat, steep, and neutral.

7. The large-scale multi-objective unit maintenance optimization method according to claim 6, characterized in that: The basic concepts of the improved DTW method are as follows: Assume that the sorted hierarchical OPC curve is considered to be composed of n sequence data Q = {q1, q2, ...q n }, the standard OPC curve is considered to be a curve formed by m sequence data C = {c1, c2, ... c m }The curve formed, assuming n>m; Define the original distance sequence od between two curves i as follows: The improved DTW distance value M(i,j) between two curves is defined as follows: M(i,j)=|q i -c j |,i=1…n,j=1…m (35) Define the cumulative improved DTW distance λ(i,j) of the matrix M(i,j) as: The above formula shows that the initial cumulative improved DTW distance is defined as The λ(i, j) of the current coordinate is related to the cumulative improved DTW distance of the previous coordinate and the improved DTW distance value M(i, j) of the current coordinate. There are only three update paths from the previous coordinate to the current coordinate (i, j), namely: The steps of using the improved DTW method to evaluate the multi-objective unit maintenance optimization problem type are as follows: 1) Obtain the improved DTW distance matrix M of the two curves and define the initial coordinates of the path (1, 1); 2) Start path finding by calculating the cumulative improved DTW distance λ(i, j). When the coordinate (2, j) is found, that is, when i→i+1, the coordinates and the cumulative improved DTW distance are updated using the following formula: The above formula indicates that when n>m, the horizontal coordinate i is used as the reference; when the horizontal coordinate update step is 1, that is, when the coordinate is updated from (1,1) to (2,j), the current coordinate and the cumulative improved DTW distance value λ(2,j) need to be determined based on the size relationship between the original distance value od2 of the two coordinates and the cumulative improved DTW distance change value λ(2,j)-λ(1,1); when od2 is larger, the improved DTW update method is considered to be better, and λ(2,j) is selected as the current cumulative improved DTW distance value, and (2,j) is selected as the current position; when λ(2,j)-λ(1,1) is larger, the original update method is considered to be better, and od2+M(1,1) is selected as the current cumulative improved DTW distance value, and (2,2) is selected as the current position; 3) Continue updating the path coordinates as described in step 2, as follows: Where j0 is the corresponding vertical coordinate j when the horizontal coordinate of the current coordinate reaches i-1 for the first time during path optimization; The above formula indicates that when the horizontal coordinate update step is 1, the current coordinate (i, j) and the cumulative improved DTW distance value λ(i, j) need to be calculated based on the original distance value od between the two coordinates (i, j) and (i-1, j0). i , the relationship between the cumulative improvement DTW distance change value λ(i,j)-λ(i-1,j0) is determined; when od i When it is larger, the improved DTW update method is considered to be better, and the calculated λ(i, j) is selected as the current cumulative improved DTW distance value, and (i, j) is selected as the current position; 4) By continuously updating the coordinates and accumulating the improved DTW distance, a path from the initial coordinate (1, 1) to the end coordinate (i, j) is eventually formed. The final calculated λ(n, m) is the shortest path; 5) According to steps 1)-4), the shortest distances between the sorted hierarchical OPC curve and the standard flat OPC curve, the standard steep OPC curve, and the standard neutral OPC curve are calculated respectively, and sequence matching is performed one by one to calculate the similarity between each two curves; based on the calculated similarity results, the type of sorted hierarchical OPC curve that is more appropriate is determined; Through the above steps, the type of OPC curve is obtained, and then the type of multi-objective unit maintenance optimization problem is determined.

Images