CN-122022295-A - Dynamic construction cost and risk prediction system for electric power engineering project based on digital twin

CN122022295ACN 122022295 ACN122022295 ACN 122022295ACN-122022295-A

Abstract

The invention discloses a dynamic construction cost and risk prediction system for an electric power engineering project based on digital twinning, relates to the technical field of electric power engineering project management and digital twinning, and solves the problems that the dynamic and accurate prediction of the construction cost and the construction period is difficult to carry out and the chain influence of a risk event is quantitatively evaluated in the prior art. According to the invention, by constructing the digital twin body with autonomous evolution capability, fusing physical constraint and time-lag causality, combining the structural causal model and the inverse fact reasoning, the probabilistic dynamic prediction of project cost and construction period and the visual map generation of risk situation are realized, and based on interactive decision sand table and multi-objective optimization, multi-scheme comparison and steady decision support are provided. The system also has self-learning optimization capability, and can continuously improve the accuracy of prediction and decision making, thereby obviously improving the fine management and the risk resistance level of the whole life cycle of the power engineering project.

Inventors

XIE WEIBIN
Ning Dongqin
WU JIANGUO
LU JINLIAN
Pu Shengyin
Fan Haoqian

Assignees

肇庆市恒电电力工程有限公司

Dates

Publication Date: 20260512
Application Date: 20260120

Claims (8)

1. The dynamic manufacturing cost and risk prediction system for the electric power engineering project based on digital twinning is characterized by comprising the following modules: the dynamic twin construction module is used for collecting and preprocessing multi-source heterogeneous data of the power engineering project, fusing the preprocessed data, and constructing a digital twin body capable of automatically generating a future state track through coupled physical constraint evolution and time lag causal interaction by adopting a physical information neural network and a causal discovery algorithm; The coupling prediction engine module is used for constructing a structural causal model based on the physical constraint evolution and time-lag causal interaction relation in the digital twin body, executing the simulation deduction of the dynamic cost and the construction period, calculating the chain influence of the risk event on the cost and the construction period through a counter fact reasoning algorithm, and outputting a probabilistic dynamic cost prediction curve and a risk situation map; The interactive decision sand table module visually presents a three-dimensional state, a dynamic cost prediction curve and a risk situation map of the power engineering project generated based on the digital twin body, responds to an intervention strategy input by a user, drives the coupling prediction engine module to perform inverse situation simulation, and adopts an optimization algorithm to process the result of the situation simulation so as to generate and compare a plurality of alternative decision schemes; And the self-learning optimization module performs cooperation and iterative updating on the structural causal model in the coupling prediction engine module and the optimization algorithm adopted by the interactive decision sand table module by circularly executing an online learning process and utilizing continuously acquired full life cycle data of the electric power engineering project.
2. The digital twinning-based power engineering project dynamic cost and risk prediction system according to claim 1, wherein the dynamic twinning construction module comprises the steps of: the method comprises the steps of collecting and preprocessing multi-source heterogeneous data and historical observation data of an electric power engineering project, fusing the preprocessed multi-source heterogeneous data, and inputting the multi-source heterogeneous data to a causal decoupling variable self-encoder, wherein the causal decoupling variable self-encoder is trained in advance, a loss function used for training comprises a data reconstruction item, a priori distribution constraint item, a total correlation penalty item and a causal semantic supervision item; Constructing a physical constraint neural differential equation based on the physical state factors, and describing continuous evolution of the physical state factors, wherein the physical constraint neural differential equation comprises a neural network layer realized by embedding a differentiable physical solver, and the layer applies a physical constraint equation set defined by knowledge in the field of power engineering at each time step As an evolution constraint, wherein At time for physical state factor State vectors of (2); Based on the non-physical factors, a structured time lag causal convolution network is adopted to construct a time lag causal interaction model among the non-physical factors, wherein the structured time lag causal convolution network is used for each pair of non-physical factors Learning an asymmetric causal influence kernel The kernel function is used for characterizing non-physical factors For non-physical factors How the causal effect intensity of (a) varies with delay time Training the structured time lag causal convolution network through sparse regularization; and the system continuously updates and maintains a causal factor state sequence according to the input multi-source heterogeneous data flow, namely a digital twin body.
3. The digital twin-based power engineering project dynamic cost and risk prediction system according to claim 2, wherein the dynamic twin construction module implements an autonomous generation future state trajectory function of the digital twin, and comprises the following steps: based on a generating type dynamic system, solving a random differential equation through numerical calculation to generate a dynamic evolution track of the digital twin body, wherein the track forms an autonomous generation result of the digital twin body on a future state, and the random differential equation has the expression: Wherein, the Is that State vectors of all causal factors at the moment; a physical evolution vector field defined by the physical constraint neural differential equation; A causal interaction vector field defined for the structured time-lag causal convolutional network; Representing intercepted slaves in a causal factor state sequence maintained by a generative dynamic system To the point of A time-of-day historical state vector segment; in order to randomly spread the term(s), Is a standard wiener process.
4. The digital twin-based power engineering project dynamic cost and risk prediction system according to claim 1, wherein the coupling prediction engine module constructs a structural causal model based on physical constraint evolution and time lag causal interaction in the digital twin, and the method comprises the following steps: Acquiring an output causal factor state sequence from a digital twin body, and determining a comprehensive time-lag causal graph represented by a structured time-lag causal convolutional network based on a connection weight obtained after training the network; determining a causal father variable set of each endogenous variable based on the comprehensive time-lag causal graph, wherein the endogenous variable comprises a cost variable, a construction period variable and a risk state variable in the electric power engineering project; constructing a structural equation for each endogenous variable, wherein each structural equation is specifically formed by: Wherein, the Is the first A plurality of endogenous variables; As a variable Is a causal parent variable set of (1); for calculating A deterministic function of the expected value under the influence of its causal parent variable; Is a function used to represent how its noise standard deviation depends on causal parent variable state changes; To obey the standard normal distribution of independent random noise, ; And (3) with Respectively as a function of And (3) with Is to be estimated; based on historical observation data of the power engineering project, parameters are mapped by maximizing a log likelihood function of the structural causal model consisting of structural equations of all endogenous variables And (3) with And carrying out joint estimation so as to complete the construction of the structural causal model.
5. The digital twin-based dynamic construction cost and risk prediction system for electric power engineering projects according to claim 4, wherein in the coupling prediction engine module, simulation deduction of the dynamic construction cost and construction period is executed, chain influence of risk events on the construction cost and construction period is calculated through a counterfactual reasoning algorithm, and a probabilistic dynamic construction cost prediction curve and risk situation map are output, and the method comprises the following steps: based on the structural causal model, a cost variable of the endogenous variables of the structural causal model Construction period variable Mathematically describing the incremental dynamics of (a) as an affine jump-diffusion process driven by the states of all endogenous variables in the structural causal model, the dynamics of the process being defined by a stochastic differential equation: Wherein, the Is shown at the moment State vectors of all endogenous variables in the structural causal model; As a drift term, And (3) with Is a coefficient matrix to be determined; For the diffusion term coefficient matrix, satisfy , And (3) with Is a pending parameter; Representing a two-dimensional standard Brownian motion; Representing a compensated poisson random measure of intensity And jump amplitude distribution Depending on the state vector ; Carrying out numerical solution and path sampling on a random differential equation by a Monte Carlo simulation method to generate a plurality of groups of cost variable and construction period variable tracks in future scenes, wherein the tracks are defined as reference prediction tracks; When a particular risk event is identified, and the event corresponds to an endogenous variable in the structural causal model In the method, based on the historical observation data of the electric power engineering project, the posterior probability distribution of the state variable of the structural causal model at the moment of identifying the risk event is deduced, and in the structural causal model, the endogenous variable is deduced Based on the structure causal model modified by the intervention operation and the initial state sampled from posterior probability distribution, forward Monte Carlo simulation is carried out to generate a group of cost variable and construction period variable tracks under the counter fact scene, which are defined as counter fact prediction tracks; The method comprises the steps of comparing a ground truth prediction track with a reference prediction track, calculating expected values and probability distribution of the change of a manufacturing cost variable and a construction period variable caused by a risk event, constructing a risk situation map based on the calculated expected values and probability distribution, wherein the map comprises causal risk probability that the manufacturing cost variable exceeds a preset threshold value and causal condition risk values when the manufacturing cost variable exceeds the threshold value in a specified future time point or time period, and the risk situation map is subjected to visual mapping in time and space dimensions.
6. The digital twin-based power engineering project dynamic cost and risk prediction system according to claim 1, wherein the interactive decision sand table module visually presents a three-dimensional state, a dynamic cost prediction curve and a risk situation map of the power engineering project generated based on the digital twin, and responds to an intervention strategy input by a user, drives the coupling prediction engine module to perform a counter fact scenario simulation, and processes the results of the scenario simulation by adopting an optimization algorithm to generate a plurality of alternative decision schemes, and the method comprises the following steps: Constructing a multi-participant causal gaming framework in which a set of cost, construction and risk state variables among endogenous variables in a structural causal model are abstracted to virtual participants Each participant Having a policy network Wherein Is an identifier and , As a trainable parameter of the policy network, the policy network is based on the current state User-entered intervention policies Output an intervention operation Current state of Based on the three-dimensional state, dynamic cost prediction curve and risk situation map information which are visually presented; combining the intervention operations of all participants to form a joint intervention operation vector, which is recorded as Based on the joint intervention operation vector The coupling prediction engine module is driven to simulate a counter-facts situation to obtain the corresponding expected variation of the manufacturing cost variable Expected change in construction period variable Risk state variable change vector And thus constitute and the vector Corresponding target vector ; Defining participants The loss function of (2) is: , wherein, Representing a target vector Middle and participants A corresponding component; traversing all other participant identifiers except k for the sum index; A reference target value set according to the reference predicted trajectory; The multi-participant causal game framework and the defined loss function thereof together form a causal game, the causal game is solved by an optimization method to obtain a group of equalization strategies, and for each equalization strategy, the determined joint intervention operation vector is marked as an equalization joint intervention operation vector Its corresponding target vector is denoted as equalization target vector The theory is that% , ) The constructed scheme is taken as an alternative decision scheme; For each obtained alternative decision scheme, if the alternative decision scheme is not governed by any existing scheme in P, adding the alternative decision scheme into P, and simultaneously removing all the schemes governed by the alternative decision scheme in P; Design factor kernel function , wherein, And For two different joint interventional operation vectors, And Respectively representing the application of joint intervention operation vectors to endogenous variables in the structural causal model And The resulting distribution of the interventions is then obtained, Representing a wasperstein distance measure between two interventional distributions, The weight coefficient is preset; constructing a gaussian process proxy model for approximating a scalar objective function using the causal kernel Wherein Representing an arbitrary joint intervention operation vector, defining an intervention entropy acquisition function , wherein, As the reference target value(s), Is the conditional mutual information entropy; searching for new joint intervention operation vectors by maximizing an intervention entropy acquisition function Calculation of Corresponding target vector Forming a new alternative decision scheme And iteratively executing the updating solution set process until the preset optimal iteration times are reached, wherein the finally obtained pareto front edge solution set P is a plurality of generated alternative decision schemes.
7. The digital twinning-based power engineering project dynamic cost and risk prediction system according to claim 6, wherein the interactive decision sand table module compares a plurality of alternative decision schemes and supports interactive decision optimization, and comprises the following steps: Embedding a plurality of alternative decision schemes in the pareto front solution set into a two-dimensional coordinate system for visual comparison by adopting a supervision comparison manifold learning algorithm, wherein each alternative decision scheme in the pareto front solution set corresponds to an equalization strategy, and each alternative decision scheme is marked as Balanced joint intervention operation vectors contained by their corresponding equalization strategies And equalizing the target vector thereof Co-representation, i.e Wherein the subscripts A sequence number representing a scheme in the solution set; for each alternative decision scheme Computing environment robustness scoring , wherein, Representing a set of predefined environmental parameter disturbances, Representing a collection The number of disturbances in (c) is, Representing a collection Is a disturbance of an environmental parameter in the air, The reference environmental parameter is represented by a reference parameter, Representing disturbances in environmental parameters Down-execution balanced joint intervention operation vector The target vector to be obtained later is a vector, Expressed in a reference environment The next corresponding target vector is then used to determine, Scoring based on the environmental robustness With a preset scoring threshold In the two-dimensional coordinate system, satisfy Is marked prominently, wherein A preset scoring threshold value; preference weight vector responsive to user input based on the visual presentation Updating policy network parameters of participants in the multi-participant causal gaming framework by a preference-guided causal policy gradient method Generating an alternative decision scheme more in line with user preferences by adjusting the policy network; based on the updated policy network parameters, the pareto front solution set formed by the updated alternative decision schemes is regenerated and output.
8. The digital twinning-based power engineering project dynamic cost and risk prediction system according to claim 1, wherein the self-learning optimization module performs the following steps in a loop to achieve collaborative and iterative updating of the structural causal model and the optimization algorithm: collecting full life cycle data of an electric power engineering project in a continuous time data stream form, calculating path signature characteristics of the data, and filtering through a self-adaptive forgetting gating unit to obtain time sequence characteristic representation ; Performing an online bayesian graph structure reconstruction for the structure causal model using a continuously relaxed probability adjacency matrix Graph structure representing its integrated time-lapse causal graph and by optimizing evidence lower bound Updating the variational posterior distribution of the matrix , wherein, To synthesize a priori distributions of the graph structure of the time-lapse causal graph, The optimization process simultaneously applies causal invariance regularization constraint; performing a meta-evolution of a multi-participant causal gaming optimization algorithm in the interactive decision sand table module, the meta-evolution passing through a super-network Realizing the decision parameters formed by the super network according to the real-time state, dynamic cost prediction curve, risk situation map and user input intervention strategy of the current electric power engineering project Outputting the internal parameters of the optimization algorithm I.e. By calculating the verification loss Parametrics to super network Is of the meta-gradient of (2) To update the super network, the calculation formula is as follows: ; Maintaining a causal memory store storing a plurality of causal patterns learned from historical power engineering project data, each causal pattern being identified as Wherein the subscripts For the sequence number of the cause and effect pattern in the cause and effect memory, An adjacency matrix representing a historical integrated time-lapse causal graph, Representing the corresponding historical structural equation parameter set, and marking the corresponding causal mode to be learned as a new power engineering project task Calculation of And each causal pattern in the causal memory Causal Wasserstein distance between According to the distance calculation result, the history causal mode which is most matched with the new power engineering project task is obtained Adjacent matrix in (a) Parameters of structural equation And migrating to the learning process of the structural cause and effect model on the new power engineering project task.

Description

Dynamic construction cost and risk prediction system for electric power engineering project based on digital twin Technical Field The invention belongs to the technical field of power engineering project management and digital twinning, and particularly relates to a dynamic manufacturing cost and risk prediction system for a power engineering project based on digital twinning. Background With the deep application of the digital twin technology in the field of engineering management, the value of the digital twin technology in the electric power engineering project is increasingly highlighted. The digital twinning provides dynamic and visual decision support for project whole process management by mapping physical entity states in real time and fusing multi-source heterogeneous data. Especially in terms of cost and construction period management, the traditional static prediction method is difficult to cope with the coupling influence of dynamic change and complex risk in the project execution process, so that the combination of digital twin to realize dynamic cost and risk prediction becomes an important direction for improving the fine management level of the electric power engineering. In this context, intelligent systems integrating causal reasoning, physical information learning and interactive simulation are becoming research hotspots. The system not only can construct a digital twin body reflecting the inherent physical laws and causal relationships of the power engineering, but also can dynamically predict the construction cost and the construction period evolution trend through analog deduction and counterfactual analysis, and evaluate the chain effect of the risk event. The interactive sand table and the self-learning mechanism are further combined, multi-scheme comparison and strategy optimization can be supported, real-time, reliable and interactive decision basis is provided for project managers, and development of power engineering project management towards the intelligent and self-adaptive direction is promoted. The following problems exist in the prior art: The traditional power engineering project cost and risk prediction method is static and isolated, and the dynamic chain effect of a risk event is difficult to simulate; The power engineering project management decision lacks an interactive simulation tool, and the comprehensive effects of different intervention strategies are difficult to evaluate and compare quickly; the prediction and decision model is fixed and cannot be optimized by itself along with the progress of the power engineering project and the accumulation of new data; The model training cost is high and the period is long when the new power engineering project is started, and a means for effectively utilizing the experience of the history project is lacked. Disclosure of Invention With the deep application of the digital twin technology in the field of engineering management, the value of the digital twin technology in the electric power engineering project is increasingly highlighted. The digital twinning provides dynamic and visual decision support for project whole process management by mapping physical entity states in real time and fusing multi-source heterogeneous data. Especially in terms of cost and construction period management, the traditional static prediction method is difficult to cope with the coupling influence of dynamic change and complex risk in the project execution process, so that the combination of digital twin to realize dynamic cost and risk prediction becomes an important direction for improving the fine management level of the electric power engineering. In this context, intelligent systems integrating causal reasoning, physical information learning and interactive simulation are becoming research hotspots. The system not only can construct a digital twin body reflecting the inherent physical laws and causal relationships of the power engineering, but also can dynamically predict the construction cost and the construction period evolution trend through analog deduction and counterfactual analysis, and evaluate the chain effect of the risk event. The interactive sand table and the self-learning mechanism are further combined, multi-scheme comparison and strategy optimization can be supported, real-time, reliable and interactive decision basis is provided for project managers, and development of power engineering project management towards the intelligent and self-adaptive direction is promoted. The following problems exist in the prior art: The traditional power engineering project cost and risk prediction method is static and isolated, and the dynamic chain effect of a risk event is difficult to simulate; The power engineering project management decision lacks an interactive simulation tool, and the comprehensive effects of different intervention strategies are difficult to evaluate and compare quickly; the prediction and decision model is fixed a