Search

CN-122022028-A - Power grid procedure quantization-oriented manufacturing cost parameter self-adaptive optimization method and system

CN122022028ACN 122022028 ACN122022028 ACN 122022028ACN-122022028-A

Abstract

The invention relates to the technical field of parameter self-adaptive control, in particular to a cost parameter self-adaptive optimization method and system for quantifying power grid procedures, comprising the following steps of firstly, acquiring a procedure unit set and a material price index sequence of a target power grid project, constructing a physical information neural network, and forming an initial set of cost parameters; step two, constructing a mixed integer programming model based on the initial set of the cost parameters, setting constraint conditions of the mixed integer programming model, solving the mixed integer programming model by adopting a branch cutting method to obtain an optimal solution, and obtaining a cost parameter reference set, and step three, constructing a layered reinforcement learning intelligent body, defining a state vector and a reward signal of the layered reinforcement learning intelligent body, and outputting the cost parameters after self-adaption optimization. The invention realizes the deep fusion of data driving and knowledge in the manufacturing cost field, and solves the problem of collaborative optimization of continuous parameter adjustment and discrete quota selection.

Inventors

  • JIANG YONG
  • Zhan Lihui
  • TU LEI

Assignees

  • 湘能卓信项目管理有限公司

Dates

Publication Date
20260512
Application Date
20260121

Claims (10)

  1. 1. The cost parameter self-adaptive optimization method facing the quantification of the power grid process is characterized by comprising the following steps of: Step one, acquiring a process unit set and a material price index sequence of a target power grid project, constructing a physical information neural network, taking a hierarchical summarization rule of power grid cost as a training process of embedding the physical information neural network into physical constraints, training the physical information neural network, inputting the process unit set into the trained physical information neural network, and acquiring a unit price predicted value and a consumption coefficient predicted value of each process unit to form an initial set of cost parameters; Step two, constructing a mixed integer programming model based on the initial set of the cost parameters, setting constraint conditions of the mixed integer programming model, solving the mixed integer programming model by adopting a branch cutting method to obtain an optimal solution, and adjusting the initial set of the cost parameters according to the optimal solution to obtain a reference set of the cost parameters; Step three, constructing a layered reinforcement learning intelligent agent, defining a state vector and a reward signal of the layered reinforcement learning intelligent agent, training the layered reinforcement learning intelligent agent, inputting real-time cost data of a construction project into the trained layered reinforcement learning intelligent agent, generating a cost parameter adjustment range, correcting a cost parameter reference set, and outputting self-adaptive optimized cost parameters.
  2. 2. The method of claim 1, wherein each process unit in the process unit set comprises a process code, a sub-engineering code, a measurement unit and an engineering quantity, wherein the input of the physical information neural network is an input vector formed by splicing a 64-dimensional vector obtained by converting the process code through an embedding layer and a material price index, the physical information neural network comprises 3 hidden layers, each hidden layer comprises 128 neurons and adopts a hyperbolic tangent activation function, and the output layer outputs a unit price predicted value and a consumption coefficient predicted value of the corresponding process unit.
  3. 3. The method according to claim 1, wherein the hierarchical aggregation rule is that for any partial engineering, the product of the engineering quantity of all process units to which any partial engineering belongs and the corresponding unit price predicted value is summed to obtain a cost calculated value of any partial engineering, the cost calculated value is subtracted from a cost marked value of any partial engineering, an absolute value is taken as a conservation residual of any partial engineering, the training goal of the physical information neural network is to simultaneously minimize the sum of the mean square error between the output predicted value and the marked value and the conservation residual of all partial engineering, the iterative training is performed by adopting an Adam optimizer, the learning rate is set to be 0.001, and the training is stopped when the change quantity of the loss value of 50 continuous iterations is smaller than 0.0001.
  4. 4. The method according to claim 1, wherein the decision variables of the mixed integer programming model include a unit price adjustment quantity variable and a consumption coefficient adjustment quantity variable set for each process unit, the unit price adjustment quantity variable and the consumption coefficient adjustment quantity variable are continuous variables, the rated gear variable set for each process unit has values of 1,2,3,4 and 5, respectively corresponding to 5 standard ratings selected by the process unit, and the mixed integer programming model aims to minimize the sum of absolute values of the unit price adjustment quantity variables and the sum of absolute values of the consumption coefficient adjustment quantity variables of all the process units.
  5. 5. The method according to claim 4, wherein the constraint conditions include a class 1 constraint that the absolute value of the unit price adjustment amount variable of each process unit is required to be not more than 15% of the unit price prediction value of the corresponding process unit, the absolute value of the consumption coefficient adjustment amount variable of each process unit is required to be not more than 10% of the consumption coefficient prediction value of the corresponding process unit, a class 2 constraint that the budget compliance constraint requires that the total construction cost be within a range of 95% to 105% of the budget control price, a class 3 constraint that the quota matching constraint requires that the adjusted unit price be within a unit price floating interval corresponding to a standard quota specified by the quota gear variable, and a class 4 constraint that the conservation transfer constraint requires that the conservation residual of each subsection engineering be not more than 2% of the corresponding subsection construction cost labeling value.
  6. 6. The method of claim 4, wherein solving the mixed integer programming model by a branch cutting method to obtain an optimal solution comprises relaxing all rated gear variables into continuous variables between 1 and 5, solving a relaxation problem by an interior point method to obtain a relaxed optimal solution, selecting rated gear variables with non-integer values, wherein the rated gear variables with small parts closest to 0.5 as branch variables, respectively establishing two sub-problems by taking down integer values and up integer values of the branch variable values, repeating the relaxation solution and the branching process for each sub-problem, generating a cutting plane at each node according to constraint conditions, cutting out the sub-problem when the optimal target value of the sub-problem is larger than the target value of the current known optimal integer solution, and outputting the integer feasible solution with the minimum target value as the optimal solution after traversing.
  7. 7. The method of claim 1 wherein the hierarchical reinforcement learning agent comprises an upper controller and a lower actuator, wherein the status vector is formed by stitching 3 parts, namely a 1 st part is a cost deviation vector, an i-th component of the cost deviation vector is a difference between a current cost parameter of an i-th process unit and a corresponding value in a cost parameter reference set, a2 nd part is a price fluctuation vector, a j-th component of the price fluctuation vector is a ratio of a current price index of the j-th material to a base price index, a 3 rd part is a progress status vector, and a k-th component of the progress status vector is a current completion proportion of the k-th partial project.
  8. 8. The method of claim 7, wherein the upper controller maintains 1Q-value table, rows of the Q-value table correspond to discretized state space, columns of the Q-value table correspond to 3 macro instructions, the 3 macro instructions are global synchronization adjustment instructions, sub-engineering grouping adjustment instructions and only deviation overrun process instructions, the upper controller performs 1 decision every 10 time steps, queries the Q-value table according to the current state vector, selects a macro instruction with the largest Q-value and outputs the macro instruction to the lower actuator, the lower actuator maintains 1 strategy network, the strategy network is a fully connected neural network comprising 2 hidden layers, each hidden layer comprises 64 neurons and adopts a ReLU activation function, the output of the strategy network is an average value and a standard deviation of the adjustment amplitude of the manufacturing cost parameters of each process unit, the lower actuator constructs normal distribution according to the average value and the standard deviation and samples to obtain the adjustment amplitude of each process unit, and the adjustment amplitude is truncated to be within a range of-3% to +3%.
  9. 9. The method of claim 7 wherein the reward signal includes an instant reward and a periodic reward, the instant reward being a negative of the root mean square deviation between the adjusted overall process unit cost parameter and the set of cost parameter references, the periodic reward being a number of time steps in the settlement period for which the root mean square deviation is below a predetermined threshold, the upper level controller being trained with a Q learning update rule, the learning rate being 0.1, the discount factor being 0.95, the lower level executor storing the current state vector, the adjustment of the execution, the obtained instant reward and the next state vector in an experience buffer, randomly extracting samples of batch size 64 from the experience buffer when the number of samples in the experience buffer reaches 1000, updating parameters of the policy network with a near-end policy optimization algorithm, the clipping ratio being set to 0.2.
  10. 10. The utility model provides a cost parameter self-adaptive optimization system facing power grid process quantification, which is characterized by comprising a conservation constraint solving module, a dynamic self-adaptive correcting module, a multi-constraint collaborative optimizing module, a cost parameter self-adaptive optimizing module and a cost parameter self-adaptive optimizing module, wherein the conservation constraint solving module is used for acquiring a process unit set and a material price index sequence of a target power grid project, training a hierarchical summarizing rule of power grid cost as physical constraint through a physical information neural network, outputting unit price predicted values and consumption coefficient predicted values of each process unit to form a cost parameter initial set, the multi-constraint collaborative optimizing module is used for constructing a mixed integer programming model based on the cost parameter initial set, setting constraint conditions and solving by adopting a branch cutting method to obtain an optimal solution, adjusting the cost parameter initial set according to the optimal solution to obtain a cost parameter reference set, and the dynamic self-adaptive correcting module is used for generating a cost parameter adjustment range according to real-time cost data of a built project through a layered reinforcement learning intelligent body, correcting the cost parameter reference set and outputting the cost parameter after self-adaptive optimization.

Description

Power grid procedure quantization-oriented manufacturing cost parameter self-adaptive optimization method and system Technical Field The invention belongs to the technical field of parameter self-adaptive control, and particularly relates to a manufacturing cost parameter self-adaptive optimization method and system for power grid process quantization. Background Currently, the determination of the power grid engineering cost parameters mainly depends on a combination mode of a rating system and manual experience adjustment. The quota system is a standardized parameter set formed by statistics and induction on the basis of a large number of engineering practices, and provides a basic basis for construction cost. However, the update period of the quota system is usually 3 to 5 years, and the influence of dynamic factors such as material price fluctuation, construction process improvement, labor cost change and the like is difficult to reflect in time. On the basis of a rating system, a manufacturing engineer needs to manually adjust according to specific engineering conditions, including selecting applicable rating gears, determining difference adjustment coefficients, checking consumption and the like. The manual adjustment mode is highly dependent on professional experience and subjective judgment of a cost engineer, and different personnel can have great difference on parameter adjustment results of the same engineering, so that the standardization and consistency of cost management are affected. In recent years, some researchers have attempted to introduce machine learning methods into the field of cost parameter prediction. A training sample set is established by collecting cost data of the completed engineering, and a mapping model between engineering characteristics and cost parameters is established by using algorithms such as regression analysis, a support vector machine, random forests and the like. The data driving method can automatically extract the characteristic rule from a large number of samples, and overcomes the limitation of manual experience to a certain extent. However, the pure data driving method has obvious disadvantages that firstly, the model output result cannot ensure to meet the basic rule of cost calculation, for example, the hierarchical summarization relation between the construction cost of a subsection engineering and the construction cost of a process unit to which the model belongs, so that the prediction result lacks rationality in engineering application, secondly, the model is difficult to effectively process complex constraint conditions of coupling discrete decisions such as rated gear selection, budget control and the like with continuous parameter adjustment, thirdly, the trained model parameters are fixed, dynamic adjustment cannot be carried out according to real-time feedback in the engineering construction process, and the parameter correction requirements caused by material price fluctuation and engineering progress change are difficult to adapt. Disclosure of Invention The invention mainly aims to provide a cost parameter self-adaptive optimization method and system for quantifying the power grid process, which realize the deep fusion of data driving and knowledge in the cost field, solve the problem of collaborative optimization of continuous parameter adjustment and discrete quota selection, realize the dynamic response of the cost parameter to material price fluctuation and engineering progress change through a layered reinforcement learning mechanism, and improve the precision and adaptability of the power grid engineering cost parameter. In order to solve the problems, the technical scheme of the invention is realized as follows: The cost parameter self-adaptive optimization method facing the quantification of the power grid process comprises the following steps: Step one, acquiring a process unit set and a material price index sequence of a target power grid project, constructing a physical information neural network, taking a hierarchical summarization rule of power grid cost as a training process of embedding the physical information neural network into physical constraints, training the physical information neural network, inputting the process unit set into the trained physical information neural network, and acquiring a unit price predicted value and a consumption coefficient predicted value of each process unit to form an initial set of cost parameters; Step two, constructing a mixed integer programming model based on the initial set of the cost parameters, setting constraint conditions of the mixed integer programming model, solving the mixed integer programming model by adopting a branch cutting method to obtain an optimal solution, and adjusting the initial set of the cost parameters according to the optimal solution to obtain a reference set of the cost parameters; Step three, constructing a layered reinforcement learning intelligent agent, de