Intelligent preparation and dynamic adjustment method for power engineering budget based on reinforcement learning

Abstract

The application discloses a power engineering budget intelligent compilation and dynamic adjustment method based on reinforcement learning, analyzes initial data to construct a continuous relaxation multi-degree-of-freedom network topology model and extracts a pseudo-Jacobi matrix; collects external environment state vectors to input a reinforcement learning decision network and a graph neural network decoding module to generate a virtual damping and virtual stiffness matrix; converts a fund gap and a compliance parameter into an external impact force and an internal physical constraint force vector, combines the pseudo-Jacobi matrix to construct a discrete first-order impedance response model, and reversely solves an output node optimal adjustment vector; when a physical deadlock state is entered, a physical construction scheme dynamic replacement operation is performed to introduce a replacement process node and solve an extended pseudo-Jacobi matrix; compliance verification is performed, instructions are issued, deviation data is collected, a total reward feedback signal is constructed, and parameters are updated. The application realizes smooth dispersion of external fund gaps along a network topology, and avoids global conflicts caused by local adjustment.

Description

Technical Field

[0001] This invention relates to the field of power engineering cost and engineering management technology, specifically to a method for intelligent compilation and dynamic adjustment of power engineering budget estimates based on reinforcement learning. Background Technology

[0002] In the construction of power engineering projects, especially substation projects, the preparation and dynamic adjustment of budget estimates are fundamental to ensuring the implementation of the project and the compliant use of funds. Substation projects involve a large number of equipment and materials, such as transformers, switchgear, cables, and hardware, as well as corresponding engineering quota items. These types of projects have long execution cycles and contain a large number of interconnected underlying process nodes.

[0003] Under the existing project management model, the industry typically relies on traditional cost estimation software for the preparation and management of initial quota budgets. In the traditional cost estimation process, cost estimators need to manually search for the corresponding quota items for equipment and materials and manually calculate the quantity conversion. This method relies on manual experience for item-by-item matching, which is time-consuming and repetitive. Furthermore, differences in understanding and judgment among different personnel can easily lead to inconsistent cost estimates. In addition, the diverse specifications and models of equipment and materials, along with the complex quota items, make manual matching prone to omissions or errors. Moreover, the lack of structured mapping relationships and intelligent quantity conversion logic fails to provide the underlying data support for subsequent automated cost calculations.

[0004] Due to the lack of the aforementioned intelligent mapping relationship, existing cost estimation software suffers from a technical deficiency when dealing with funding gaps caused by external market price fluctuations. It cannot distribute these gaps along the network topology, making it prone to global funding allocation conflicts due to local adjustments. When changes in the external environment lead to funding gaps at local construction nodes, existing technologies generally adopt a static numerical deduction approach for the individual nodes experiencing cost overruns. Because there are physical, spatiotemporal, and temporal dependencies between various process nodes within a substation project, funding and schedule reduction operations at local nodes, detached from the constraints of the project's network topology, will alter the resource flow between preceding and subsequent related nodes, thereby triggering global funding allocation conflicts and causing a disconnect between actual construction progress and the budget allocation plan. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for intelligent compilation and dynamic adjustment of power engineering budget estimates based on reinforcement learning. Existing technologies typically employ reinforcement learning methods based on operations research rules or conventional Markov decision processes when handling dynamic adjustments to power engineering budget estimates. Conventional reinforcement learning methods treat each budget item as an independent continuous variable, directly outputting the budget adjustment amount within a large action space. This approach suffers from the technical drawback of excessively high system state dimensions, leading to difficulty in model convergence. Furthermore, existing reinforcement learning models do not incorporate the physical and spatiotemporal dependencies of the actual power engineering project, easily outputting invalid instructions that violate physical constraints such as construction logic sequence or grid dispatch timing. In addition, when fluctuations in external commodity market prices cause local process costs to exceed the budget, existing methods mainly rely on static numerical reductions; when numerical adjustments reach their limits and trigger budget network deadlock, existing technologies cannot reorganize the underlying physical construction scheme at the system level, leading to project stagnation.

[0006] To achieve the above objectives, the present invention provides a method for intelligent compilation and dynamic adjustment of power engineering budget estimates based on reinforcement learning, comprising the following steps: The initial data of the power engineering project are analyzed, and a continuously relaxed multi-degree-of-freedom network topology model is constructed to establish a forward kinematic mapping model between the continuous system state vector and the target state vector. The partial derivatives of the forward kinematic mapping model are then solved to extract the pseudo-Jacobi matrix. An external environment state vector containing external market fluctuations and internal deviation data is collected, and the external environment state vector is input into a reinforcement learning decision network and a graph neural network decoding module to generate a virtual damping matrix and a virtual stiffness matrix. The funding gap caused by the external environment state vector is converted into an external impact force vector, and the compliance parameters are extracted and converted into an internal physical constraint force vector. A discrete first-order impedance response model is constructed by combining the pseudo Jacobian matrix, the virtual damping matrix and the virtual stiffness matrix. Based on the discrete first-order impedance response model, the external impact force vector and the internal physical constraint force vector are solved in reverse to output the optimal adjustment vector of the node. During the reverse calculation process, the singularity measure of the pseudo-Jacobi matrix is ​​monitored in real time. When it is determined that a physical deadlock is trapped, a dynamic replacement operation of the physical construction scheme is performed to introduce alternative process nodes. The extended pseudo-Jacobi matrix is ​​re-solved based on the total number of newly added nodes including the alternative process nodes, so as to return to the step of performing the reverse calculation. Based on the optimal adjustment vector of the node, the cost data compliance verification is performed and the adjustment instruction is issued; the actual deviation data within the actual project execution cycle is collected to construct the system total reward feedback signal, and the system total reward feedback signal is fed back to the reinforcement learning decision network for model parameter iterative update.

[0007] Preferably, the step of analyzing the initial data of the power engineering project, constructing a continuously relaxed multi-degree-of-freedom network topology model to establish a forward kinematic mapping model between the continuous system state vector and the target state vector, and solving the partial derivatives of the forward kinematic mapping model to extract the pseudo-Jacobi matrix includes: Based on the cost mapping allocation matrix, the amount of funds is projected and integrated onto the corresponding node identifier in the process node set, and the upper limit of the fixed budget funds and the base number of planned construction period days corresponding to each process node in the process node set are extracted; The baseline construction rate constant and the midpoint time parameter of the schedule are calculated using the planned construction period baseline days, and then substituted into the growth curve model to output the continuous progress status parameters. Establish a benchmark consumption baseline based on the product of the continuous progress status parameter and the upper limit of the fixed budget funds. Summate the benchmark consumption baseline with a continuous Gaussian noise variable generated using the standard deviation of historical price fluctuations to generate a continuous funding status parameter. The continuous system state vector is reconstructed by dimensional concatenating the continuous progress state parameter and the continuous funding state parameter. The total cost and total duration of the power engineering project are extracted to construct the target state vector. The forward kinematic mapping model between the target state vector and the continuous system state vector is established. The partial derivatives of the forward kinematic mapping model are calculated to obtain the pseudo-Jacobi matrix that quantifies the sensitivity to local state changes.

[0008] Preferably, the acquisition of an external environment state vector containing external market fluctuations and internal deviation data, and inputting the external environment state vector into a reinforcement learning decision network and a graph neural network decoding module to generate a virtual damping matrix and a virtual stiffness matrix, includes: Long Short-Term Memory (LSTM) networks are used to extract fixed-dimensional hidden state features from historical price fluctuation sequence data. The external environment state vector is constructed by combining the internal deviation data between the fixed-dimensional hidden state features, the meteorological warning frequency parameters and the actual consumption funds and the upper limit of the fixed budget funds through vector dimension splicing. The reinforcement learning decision network is used to perform feedforward neural network calculations based on the external environment state vector, and outputs a low-dimensional global policy feature vector.

[0009] Preferably, after outputting the low-dimensional global policy feature vector, the method further includes: The low-dimensional global strategy feature vector is distributed and broadcast, and then concatenated with the local continuous state features by vector dimension to generate an initial node feature matrix. The feature vector of the initial node feature matrix is ​​input into a single-layer feedforward neural network to calculate the attention score. The attention score is normalized using a normalized exponential function to obtain the attention weight coefficient. Based on the attention weight coefficient, the feature vectors of adjacent process nodes are weighted and summed to output a high-dimensional local node feature vector. A positive definite nonlinear activation function is connected to the output layer at the end of the independent and parallel multilayer perceptron network branch. The positive definite nonlinear activation function is used to monotonically map the unbounded real values ​​of the network output to positive real number interval values. The positive real number interval values ​​are used to construct a diagonal matrix to generate the virtual damping matrix and the virtual stiffness matrix.

[0010] Preferably, the step of converting the funding gap caused by the external environment state vector into an external impact force vector and extracting compliance parameters into an internal physical constraint force vector includes: Using the material node mapping matrix, the material price deviation value is projected and calculated as the budget overrun pressure value of the corresponding process node, and the external impact force vector is generated by summarizing. Calculate the time difference between the planned start time and the start timestamp of the power outage window for each process node, as well as the time difference between the planned completion time and the end timestamp. A maximum repulsive force truncation threshold is set, and a truncated inverse proportional repulsive function is constructed based on the time difference and the artificial potential field algorithm logic. When the output resistance force value exceeds the maximum repulsive force truncation threshold, it is forcibly assigned the maximum repulsive force truncation threshold value. The resistance force values ​​are then summarized to generate the internal physical constraint force vector.

[0011] Preferably, the step of constructing a discrete first-order impedance response model by combining the pseudo-Jacobi matrix, the virtual damping matrix, and the virtual stiffness matrix, and then performing inverse calculations on the external impact force vector and the internal physical constraint force vector based on the discrete first-order impedance response model to output the optimal adjustment vector for the nodes includes: Extract the positive real number interval values ​​on the main diagonal of the virtual stiffness matrix to construct a diagonal weight matrix, and perform a combination operation using the transpose of the pseudo-Jacobi matrix, the pseudo-Jacobi matrix itself, and the inverse of the diagonal weight matrix to obtain a weighted generalized pseudo-inverse matrix; The discrete first-order impedance response model is established by combining the virtual damping matrix, the virtual stiffness matrix, the external impact force vector, and the internal physical constraint force vector, and the desired state adjustment vector is obtained by solving the model. Perform a matrix multiplication operation between the weighted generalized pseudo-inverse matrix and the desired state adjustment vector to output the basic state adjustment vector; Non-critical nodes belonging to non-critical paths and having budget surplus values ​​are identified to construct a compliance optimization target vector. The weighted generalized pseudo-inverse matrix and the pseudo-Jacobi matrix are combined to generate a null space projection operator matrix. The null space state adjustment vector is calculated by performing matrix multiplication between the null space projection operator matrix and the compliance optimization target vector. Perform a vector addition operation between the basic state adjustment vector and the null space state adjustment vector to output the optimal adjustment vector for the node.

[0012] Preferably, during the reverse calculation process, the singularity measure of the pseudo-Jacobi matrix is ​​monitored in real time, and when it is determined that a physical deadlock has occurred, the following steps are taken: Perform singular value decomposition on the pseudo-Jacobi matrix to extract the main diagonal matrix, and perform chain multiplication algebra calculation on all non-zero singular values ​​inside the main diagonal matrix to obtain the operability measure value. Singular value decomposition is performed on the initial pseudo-Jacobi matrix to calculate the initial operability measure value. The standard deviation of historical price fluctuations is extracted, and a nonlinear mapping between the standard deviation of historical price fluctuations and the initial operability measure value is established to obtain the scaling factor. The initial operability measure value is multiplied by the scaling factor to generate the singular safety threshold. The operability measure value is compared with the singularity security threshold. When the operability measure value is less than the singularity security threshold, the pseudo-Jacobi matrix is ​​determined to be trapped in a singular configuration indicating that the predictive network is in the physical deadlock state.

[0013] Preferably, the dynamic replacement operation of the physical construction scheme introduces alternative process nodes, including: Extract the right singular vector matrix output by the singular value decomposition operation, locate the singular value with the smallest value in the main diagonal matrix, and extract the target column vector corresponding to the smallest singular value from the right singular vector matrix. The target element with the largest absolute value in the target column vector is selected by numerical sorting, the position index is extracted and mapped and matched with the node identifier, and the bottleneck process node that caused the physical deadlock is locked. Cut off all preceding and following directed edges connected to the bottleneck process node in the directed acyclic graph, remove the bottleneck process node, and add the alternative process nodes retrieved from the standardized alternative construction scheme database into the process node set. A new directed edge containing the alternative process node is established based on the alternative process flow template.

[0014] Preferably, the step of recalculating the extended pseudo-Jacobi matrix based on the total number of new nodes including the replacement process nodes includes: The actual amount of funds consumed and the number of days consumed before the bottleneck process node was eliminated are defined as sunk costs. The sunk costs are transferred and mapped to the alternative process node according to the weight of the project quantity and initialized to ensure the conservation of the total amount of funds consumed. The state vector of the continuous system is reconstructed based on the expanded total number of newly added nodes, and the partial derivatives of the forward kinematic mapping model are resolved to generate the extended pseudo-Jacobi matrix.

[0015] Preferably, the step of collecting actual deviation data within the actual engineering execution cycle to construct a system total reward feedback signal, and feeding the system total reward feedback signal back to the reinforcement learning decision network for iterative update of model parameters, includes: Compare the actual cost data with the upper limit of the budgeted funds to calculate the cost overrun penalty value, and compare the actual progress data with the planned construction period base days to calculate the construction period delay penalty value; Retrieve the directed acyclic graph, detect whether there is any over-level payment behavior that is reversed from the directed edge in the actual time sequence of fund disbursement to the node, and generate a topology violation penalty value based on the amount of funds disbursed over-level. Assign weight coefficients to each penalty value, and set the weight coefficient of the topology violation penalty value to be greater than the weight coefficient of the cost overrun penalty value and the weight coefficient of the construction period delay penalty value. The penalty values ​​are combined with the configured weight coefficients and weighted summation is performed to output a negative total reward feedback signal of the system, which is then fed back to the reinforcement learning decision network to adjust the connection weights using the policy gradient optimization mechanism.

[0016] This invention provides a method for intelligent compilation and dynamic adjustment of power engineering budget estimates based on reinforcement learning. It has the following beneficial effects: 1. This invention constructs a network topology model by analyzing initial data of power engineering projects, and uses a discrete first-order impedance response model to perform inverse calculations of external impact forces and internal constraints. It establishes an automatic mapping mechanism for equipment quotas and a conversion mechanism for engineering quantities, thereby transforming traditional manual matching into automated processing to significantly improve the efficiency of cost preparation and reduce the error rate. At the same time, it smoothly disperses external funding gaps along the network to avoid triggering global funding allocation conflicts.

[0017] 2. This invention employs a reinforcement learning decision network and a graph neural network decoding module to automatically generalize budget control strategies to the underlying process nodes. It also introduces topology violation penalty feedback signals to perform closed-loop iterative updates of model parameters. This achieves the elimination of subjective human differences through intelligent computational logic, thereby ensuring a high degree of consistency in cost results and laying the foundation for subsequent dynamic adjustments. At the same time, it constrains the model's exploration boundaries to prevent outputs from violating construction logic instructions.

[0018] 3. This invention locks the bottleneck process node when it is determined to be in a physical deadlock state by real-time monitoring and evaluation matrix singularity measure, and automatically calls the alternative nodes in the standardized alternative construction scheme database to replace the scheme and reconstruct the funding dimension. It realizes the dynamic and flexible iteration of system mapping rules and transformation logic. While relying on the underlying scheme reorganization to resolve the computational deadlock and ensure the operation of the system, it also gives the engineering cost system good scalability to adapt to the introduction of new materials and new processes. Attached Figure Description

[0019] Figure 1 The main flowchart of the intelligent compilation and dynamic adjustment method for power engineering budget based on reinforcement learning provided in the embodiments of the present invention is shown. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To achieve intelligent compilation and dynamic adjustment of power engineering budget estimates based on reinforcement learning, this invention provides a multi-level intelligent closed-loop system architecture. The system architecture includes a data perception and constraint layer, a core logic operation layer, and an adjustment instruction execution layer in terms of physical deployment and logical control.

[0022] The data perception and constraint layer is used to collect, transform, and persist physical entity data and business specification data of power engineering projects. Specifically, it extracts initial design drawings, work breakdown structure node sequences, cost breakdown structure funding benchmark data, and physical constraint boundaries for power grid dispatch. These physical constraint boundaries explicitly cover the time limits of power grid outage windows, the three-dimensional mutual exclusion conditions of construction work areas, and the minimum procurement lead time for bulk material supply chains. The data perception and constraint layer transforms the original engineering documents into structured graph node and edge relationship data, providing a data foundation for the core logic operation layer.

[0023] The core logic operation layer performs dynamic calculations of the budget and generates strategies. Internally, it integrates a continuous relaxation mapping module, a reinforcement learning decision network, a graph neural network decoding module, and an inverse kinematics optimization solver. The continuous relaxation mapping module uses a nonlinear smoothing function to convert discrete construction progress instructions and step-type payment milestone data from the data perception and constraint layer into real-valued state vectors that satisfy continuous differentiation conditions. The reinforcement learning decision network includes a strategy output action network and a value assessment network, receiving time-series data from the external market environment and execution feedback data. The graph neural network decoding module includes a multi-layer graph attention mechanism network that performs message passing and feature aggregation along the edges of the engineering topology graph, generalizing the low-dimensional spatial features output by the reinforcement learning decision network to each node, generating the diagonal parameter matrix required for the discrete impedance control model. The inverse kinematics optimization solver utilizes a generalized pseudo-inverse matrix operation mechanism to solve for the optimal node redistribution vector for funding deviations while satisfying physical constraint boundaries.

[0024] The adjustment instruction execution layer includes a variable degree-of-freedom (DOF) reconstruction engine and a closed-loop delivery module. The DDF reconstruction engine communicates bidirectionally with the inverse kinematics optimization solver in the core logic operation layer. When the inverse kinematics optimization solver determines that the current computation matrix is ​​in a singular deadlock state, the DDF reconstruction engine retrieves alternative construction schemes from a pre-defined standardized alternative construction scheme database. For example, alternative construction schemes might replace on-site foundation pouring nodes with new combinations including prefabricated module assembly nodes and prefabricated module transportation nodes. After retrieving the alternative scheme, the DDF reconstruction engine modifies the network topology and data dimensions of the multi-level intelligent closed-loop system architecture. The closed-loop delivery module compiles the final calculated node funding and schedule adjustment vectors into standard-format budget change detail reports and construction log update instructions, writes them to the corresponding project management information terminal, and feeds back the actual execution deviations to the reinforcement learning decision network.

[0025] Please see Figure 1 Based on the above system architecture, the reinforcement learning-based intelligent compilation and dynamic adjustment method for power engineering budgeting provided in this embodiment of the invention includes the following macroscopic working steps: Step S1: Analyze the initial data of the power engineering project and construct a continuously relaxed multi-degree-of-freedom network topology model. The system reads the work breakdown structure and cost breakdown structure data, extracting the underlying construction and procurement sub-item nodes. The progress and funding statuses of the underlying construction and procurement sub-item nodes are concatenated to construct a continuous system state vector. A nonlinear growth function is used to continuously fit the historical data of discrete jump nodes, constructing a forward kinematic mapping model between the continuous system state vector and the target state vector containing the total cost and total construction period. After establishing the forward kinematic mapping model, the system calculates the partial derivatives of the target state with respect to the system state based on the forward kinematic mapping model, solving for the pseudo-Jacobi matrix representing the sensitivity of the system to local state changes.

[0026] Step S2 involves sensing external environmental state data and generating first-order discrete impedance parameters through graph neural network decoding. A reinforcement learning decision network collects the current external environmental state vector and outputs low-dimensional global policy features. The external environmental state vector includes time-series features of bulk engineering material market prices and meteorological warning frequency parameters. The graph neural network decoding module uses a multi-degree-of-freedom network topology model as its graph structure, mapping the low-dimensional global policy features to high-dimensional features at the node level. These features are then transformed into virtual damping and virtual stiffness matrices with a diagonal matrix structure through a multilayer perceptron network.

[0027] Step S3 involves constructing a discrete first-order impedance control response mechanism and inversely calculating the expected state adjustments for funding and schedule at nodes. The system converts the funding gap caused by external environmental conditions into an external impact force vector applied to associated nodes, and extracts power grid safety and audit compliance parameters to convert them into internal physical constraint force vectors. Combining the virtual damping matrix and virtual stiffness matrix generated in step S2, a discrete first-order impedance response model for the expected state adjustments is established. Based on this, the inverse kinematics optimization solver constructs a weighted generalized pseudo-inverse matrix based on the pseudo-Jacobi matrix, and calculates the optimal adjustment vector for nodes that satisfies the internal physical constraint force vector constraints using null space projection optimization technology.

[0028] Step S4: Monitor the singularity measure of the pseudo-Jacobi matrix in real time to determine the physical deadlock state of the budget network. During the reverse calculation in step S3, perform singular value decomposition on the pseudo-Jacobi matrix to extract the non-zero singular values ​​on the main diagonal. Calculate the current operability measure of the system based on the product of all non-zero singular values. Determine if the operability measure is lower than a preset singularity safety threshold. If the operability measure is greater than or equal to the singularity safety threshold, it indicates that the current network topology has redundancy to absorb external shocks, and the system continues to perform conventional calculations and jumps to step S6; if the operability measure is less than the singularity safety threshold, it indicates that the system operation is trapped in a singular configuration, and the conventional numerical adjustment capability of the original network topology has reached its limit, and the system process proceeds to step S5.

[0029] Step S5: Trigger the variable degree-of-freedom reconstruction engine to perform a dynamic replacement operation of the physical construction scheme. The system traces the corresponding restricted node identifiers in the pseudo-Jacobi matrix that lead to singular states and retrieves feasible alternative process nodes from the candidate construction scheme database. These alternative process nodes are added as new mathematical degrees of freedom to the multi-degree-of-freedom network topology model, and the system's state vector dimension is reconstructed based on the new nodes. The corresponding extended pseudo-Jacobi matrix is ​​re-solved based on the new state vector dimension, removing the singular state restrictions of the pseudo-Jacobi matrix. The system then returns to step S3 with the updated extended pseudo-Jacobi matrix for the redistribution of funds.

[0030] Step S6: Perform cost data compliance verification and update the feedback parameters of the reinforcement learning decision network. The system summarizes the optimal adjustment vectors of nodes obtained under non-singular states. The updated node budget amount and schedule node update values ​​after superimposing the adjustment vectors are sent to the built-in power cost quota standard compliance verification model for boundary comparison. After verification, the closed-loop distribution module generates the corresponding budget change details report and structured construction log update instructions, and distributes them to the entity management system to complete the dynamic adjustment of the budget. The closed-loop distribution module collects cost deviation and schedule deviation data within the actual project execution cycle, constructs the reward function value of reinforcement learning, and feeds it back to the reinforcement learning decision network to complete the closed-loop iterative update of the model parameters.

[0031] In the process of parsing initial data from power engineering projects to construct a multi-degree-of-freedom network topology model, the system performs the following steps to complete the data extraction and graph definition of discrete engineering nodes: Step S101: Analyze the initial design data of the power engineering project and extract discrete process nodes and funding benchmark data. The system connects to the project's engineering management information terminal and reads the initial design drawings and associated cost accounting documents. Based on the engineering quantity quota pricing standard, the system breaks down the entire power engineering project into multiple underlying construction operation sub-items and material procurement sub-items. The specific implementation forms of the underlying construction operation sub-items include substation foundation excavation and transformer equipment hoisting. The specific implementation forms of the material procurement sub-items include special flame-retardant cable procurement and core electrical equipment customization. The system assigns a unique node identifier to each decomposed construction operation sub-item and material procurement sub-item, and aggregates all sub-items after assigning node identifiers to form a process node set. To solve the problem of misalignment between the work breakdown structure and the cost breakdown structure, the system reads the cost item codes in the cost accounting documents and establishes a cost feature vector containing labor costs, material costs, and machinery usage costs. The system constructs a multi-dimensional cost mapping and allocation matrix based on the engineering quantity measurements of underlying physical materials. According to this matrix, the funding amount under specific cost item codes is calculated proportionally to the construction workload and projected onto the corresponding node identifiers. By performing the cost mapping and allocation matrix projection operation, the system simultaneously extracts the upper limit of the quota budget funding and the planned base number of days for each work process node in the work process node set. The upper limit of the quota budget funding and the planned base number of days constitute the numerical hard boundaries for subsequent fund reallocation and schedule adjustments.

[0032] Step S102: Define the discrete state-space data structure for each process node. For each independent process node in the process node set, the system instantiates a node state vector in memory containing progress and funding status parameters. The progress status parameter represents the percentage of completion of the process node at the physical construction site. The funding status parameter represents the cumulative amount of actual funds written off since the start of the process node. Progress data entry and funding disbursement operations in physical project management are constrained by preset milestone acceptance window timeframes, causing the progress and funding status parameters to exhibit discrete, step-like values ​​when acceptance conditions are not triggered, forming a discontinuous state-space data foundation that cannot be directly differentiated. The system concatenates and merges the progress and funding status parameters of all process nodes in the process node set according to the lexicographical order of the node identifiers, outputting the system's total discrete state vector representing the current execution state of the entire power engineering project.

[0033] Step S103: Construct a directed acyclic graph (DAG) structure containing spatiotemporal dependencies based on the entity construction logic constraints. The system reads and parses the construction organization design documents in the initial design data of the power engineering project, identifying the pre- and post-process association constraints between different process nodes. The pre- and post-process association constraints reflect the rigid dependency relationships between construction operations in terms of physical spatial occupancy order and temporal sequence. In the entity project, the high-voltage transmission line erection operation must be carried out after the tower foundation excavation operation and tower erection operation have been fully accepted. The system uses the process nodes in the process node set as the vertices of the graph data structure, and the parsed pre- and post-process association constraints as the directed edges of the graph data structure, generating a DAG describing the network topology of the power engineering budget. The system sets the numerical weights of the directed edges to be a composite weight that includes the constant time lag of capital flow between adjacent process nodes and the spatial material transfer cost.

[0034] Step S104 maps the directed acyclic graph (DAG) to a multi-degree-of-freedom kinematic linkage system topology. The system establishes a correspondence between the graph data structure and the kinematic model, mapping process nodes in the DAG to motion joints in the multi-degree-of-freedom kinematic linkage system, and mapping directed edges in the DAG to rigid constraint links connecting these motion joints. The rigid constraint links define the relative numerical change constraint domains of different process nodes in terms of progress and cost consumption. By establishing this mapping correspondence, the system transforms the constraints of the physical construction logic of power engineering on budget execution into a set of graph topological mathematical constraints readable by the reinforcement learning computation layer, completing the topological graph transformation of discrete business data.

[0035] After completing the topological graph definition of the discrete engineering nodes, the system performs the following steps to complete the continuous relaxation process of the system state space: In step S105, the system reads the total discrete state vector of the system output in step S102. To satisfy the operational condition of continuous differentiability of the state matrix for inverse kinematics solution, the system performs continuous relaxation operation. The system uses a Logistic growth curve model to generate continuous progress state parameters for process nodes. Construction operations exhibit an objective physical law of slow preparation in the early stage, accelerated input of factors in the middle stage, and slowed down in the later stage. The Logistic growth curve model fits the aforementioned construction operation law. The system reads the baseline Gantt chart schedule plan from the initial design data of the power engineering project and extracts the planned start time and planned completion time of the process nodes from the baseline Gantt chart schedule plan. The system uses the planned start time and planned completion time to calculate the baseline construction rate constant to measure the construction intensity and locate the midpoint time parameter when the construction progress reaches 50%. The system substitutes the baseline construction rate constant and the midpoint time parameter into the Logistic growth curve model and outputs smooth and continuously differentiable continuous progress state parameters.

[0036] In step S106, the system synchronously processes the funding status parameters of each work process node continuously. The system extracts the upper limit of the budgeted funding for each work process node and establishes a baseline consumption rate based on the product of the continuous progress status parameter and the upper limit of the budgeted funding. Project funding reimbursement is strictly dependent on the physical progress of the project; the baseline consumption rate establishes the underlying mathematical coupling relationship between cost data and progress data. The system collects historical price fluctuation data from the external engineering material market and calculates the standard deviation of the price fluctuations. The system uses the square of the standard deviation as the variance parameter of the Gaussian distribution to generate a continuous Gaussian noise variable. The system sums the baseline consumption rate and the continuous Gaussian noise variable to generate a continuous funding status parameter that includes external environmental disturbances.

[0037] In step S107, the system concatenates the continuous progress state parameter and the continuous funding state parameter after continuous mapping processing to reconstruct the time-space data. continuous system state vector The system extracts the total cost and total construction period of a power engineering project and constructs a target state vector. The system establishes the target state vector. With the state vector of a continuous system Forward kinematic mapping model between Forward kinematic mapping model It is a globally smooth and differentiable nonlinear multivariate function. The system, based on a forward kinematic mapping model, treats the state vector of the continuous system. Solve for the partial derivatives: In the formula, Indicates time The pseudo-Jacobi matrix, The differential operator represents the positive kinematic mapping model. Differential operators representing the state vector of a continuous system. Pseudo-Jacobi matrix. The internal real-valued elements quantify the sensitivity of micro-level process nodes' funding changes or schedule delays to the total project cost and duration. The system uses a pseudo-Jacobi matrix. Stored in memory as underlying derivative data for dynamic reverse adjustment of the budget.

[0038] After constructing the mapping model containing continuous system state vectors, the system performs environmental state awareness and reinforcement learning action decoding, executing the following steps: Step S201: The system constructs an environment state vector for the reinforcement learning decision network. This environment state vector aggregates external market data and internal deviation data of the physical project from a time-series perspective. External market data includes historical price fluctuation data of bulk engineering materials and meteorological warning frequency parameters. Internal deviation data includes the difference between actual expenditure and the upper limit of the budgeted expenditure. The historical price fluctuation data is a multi-dimensional time-series matrix; the system uses a Long Short-Term Memory (LSTM) network to extract fixed-dimensional hidden state features from this data. The system then performs vector dimensionality reduction and alignment, and dimensional concatenation combination on the fixed-dimensional hidden state features, meteorological warning frequency parameters, and internal deviation data, before inputting them into the reinforcement learning decision network.

[0039] In step S202, the reinforcement learning decision network, based on the environment state vector, calculates and outputs a low-dimensional global policy feature vector through a feedforward neural network. The dimension of the low-dimensional global policy feature vector is less than the total number of nodes in the process node set. The numerical elements of the low-dimensional global policy feature vector determine the system's constraint strength on capital reserves and the extent of intervention and allocation on construction progress. By outputting the low-dimensional global policy feature vector, the system avoids the computational resource exhaustion and model non-convergence problems caused by the reinforcement learning decision network directly outputting independent adjustment parameters to the large number of process nodes.

[0040] In step S203, the system establishes a feature mapping relationship between the low-dimensional global policy feature vector and the process nodes. The system extracts the continuous system state vector output from the continuous relaxation processing stage and decomposes it into local continuous state features corresponding to each independent process node according to the node identifier. The system distributes and broadcasts the low-dimensional global policy feature vector to each process node in the process node set, and concatenates it with the local continuous state features of each process node in memory space to generate an initial node feature matrix for all process nodes. The initial node feature matrix generated by concatenation serves as the input data for the decoding stage of the graph attention network.

[0041] In step S204, the system uses a graph attention network to perform feature aggregation and message passing operations on the initial node feature matrix. The graph attention network uses the directed acyclic graph containing pre- and post-association constraints constructed above as its graph structure parameters. In the hidden layer of the graph attention network, the system calculates the attention weight coefficients between adjacent process nodes. The feature vector of the process node and the ... The feature vectors of each process node are concatenated and input into a single-layer feedforward neural network to calculate the attention score. The system uses a normalized exponential function to normalize the attention score and obtain the attention weight coefficient. The attention weight coefficient considers the constant time lag of capital flow and the spatial material transfer cost between adjacent process nodes, quantifying the feature correlation strength of adjacent process nodes in responding to capital shocks. The system performs a weighted summation of the feature vectors of adjacent process nodes based on the attention weight coefficient, outputting a high-dimensional local node feature vector: In the formula, Indicates the first The process node at the _ The high-dimensional local node feature vectors output by the layered graph attention network In a directed acyclic graph, the term is related to the first... Each process node has a set of adjacent nodes connected by directed edges. Represents the calculated first... The process node and the first Attention weight coefficients between process nodes Indicates the first The learnable parameter matrix of the layer, Indicates the first The process node at the _ Feature vectors output by layered graph attention networks This represents a nonlinear activation function. The system transforms and maps the low-dimensional global policy feature vector into a high-dimensional feature matrix that matches the size of the process node set, providing a data foundation for generating the budget control parameters for the process nodes.

[0042] After the graph attention network outputs high-dimensional local node feature vectors, the system performs the following steps to complete the mapping generation of impedance control parameters and iterative model updates: In step S205, the system inputs the high-dimensional local node feature vectors output by the graph attention network into an independent and parallel multilayer perceptron network branch. The multilayer perceptron network branch includes a damping mapping network and a stiffness mapping network. To ensure the stability of the subsequently constructed discrete first-order impedance control system at the mathematical solution level, the system restricts the output virtual damping parameters and virtual stiffness parameters to be strictly greater than zero. A positive definite nonlinear activation function is connected to the output layer of the damping mapping network and the stiffness mapping network, using this function to monotonically map the unbounded real values ​​output by the network to a range of positive real numbers. The system constructs a diagonal matrix using the mapped positive real number range values, generating virtual damping matrices and virtual stiffness matrices that correspond one-to-one with the underlying process nodes. The parameters within the virtual damping matrix quantify the numerical dissipation capability of the process nodes to absorb external financial shocks, while the parameters within the virtual stiffness matrix quantify the numerical resistance to deformation of the process nodes to maintain the upper limit of the fixed budget funds.

[0043] After generating the virtual damping matrix and virtual stiffness matrix, the system performs the following steps to complete the external impact transformation and construct the discrete first-order impedance response equation: Step S301: The system extracts the capital fluctuation values ​​caused by the external market environment and constructs an external impact force vector applied to the multi-degree-of-freedom network topology model. The system reads the real-time material unit price from the external engineering material market, calculates the difference between the real-time material unit price and the benchmark material unit price set in the quota budget, and obtains the material price deviation value. The system establishes a material node mapping matrix between material types and process nodes. Elements within the material node mapping matrix mark the consumption ratio of a specific material type for a specific process node. Using the material node mapping matrix, the system projects the material price deviation value as the budget overrun pressure value for the corresponding process node. The system arranges the budget overrun pressure values ​​of all process nodes in order of node identifier, generating an external impact force vector. The external impact force vector serves as the input variable that triggers changes in the state matrix in the impedance response model.

[0044] In step S302, the system extracts parameters related to power grid safety and audit compliance, and constructs an internal physical constraint force vector. The system obtains the start and end timestamps of the power outage window from the entity project scheduling management system. The system calculates the time difference between the planned start time and the start timestamp, and the time difference between the planned completion time and the end timestamp for each process node. To prevent numerical overflow during mathematical calculations from causing system crashes, the system sets a maximum repulsive force truncation threshold. Based on the time difference and the artificial potential field algorithm logic, the system constructs a truncated inverse proportional repulsive function. When the time difference shrinks and approaches zero, the inverse proportional repulsive function outputs a non-linearly increasing resistance force value pointing towards the safe time interval within the outage window. If the resistance force value exceeds the maximum repulsive force truncation threshold, the system forcibly assigns the resistance force value to the maximum repulsive force truncation threshold. The system summarizes the resistance force values ​​output by all process nodes to generate an internal physical constraint force vector. This internal physical constraint force vector serves as the boundary value matrix that restricts the out-of-bounds offset of the node state vector.

[0045] In step S303, the system combines the virtual damping matrix and the virtual stiffness matrix to establish a discrete first-order impedance response model for the desired state adjustment. The system obtains the reporting and update frequency of actual project cost data and actual progress data. Due to the low-frequency sparsity of the reporting and update frequency, the system discards the second-order inertial mass parameter to avoid numerical oscillations. The system sets the discrete control time step to 1 and establishes a discrete first-order impedance response mathematical model including the virtual damping matrix, the virtual stiffness matrix, the external impact force vector, and the internal physical constraint force vector. In the formula, Indicates time The desired state adjustment vector obtained by solving, Indicates time The virtual damping matrix, Indicates time The virtual stiffness matrix, Indicates time The external impact force vector, Indicates time The internal physical constraint force vector, This represents the matrix inversion operator. The expected state adjustment vector contains the theoretically anticipated values ​​of the adjustments needed to the funding and schedule parameters of the process nodes under the dual influence of external market fluctuations and internal compliance constraints. The system inputs the expected state adjustment vector into the inverse kinematics optimization solver module as the target input variable for performing the funding reallocation solution.

[0046] After establishing the discrete first-order impedance response model and outputting the desired state adjustment vector, the system performs the following steps to complete the compliant reallocation of funds: In step S304, the system reads the virtual stiffness matrix output by the graph neural network decoding module and extracts the positive real number intervals on the main diagonal of the virtual stiffness matrix as the adjustment resistance coefficients of the nodes. The system constructs a diagonal weight matrix based on these adjustment resistance coefficients. The larger the value of the diagonal elements within the diagonal weight matrix, the smaller the compression of the capital or schedule parameters during the redistribution process of the corresponding process node. The system retrieves the pseudo-Jacobi matrix generated by the continuous relaxation mapping module. The system performs matrix multiplication and matrix inversion operations using the transpose of the pseudo-Jacobi matrix, the pseudo-Jacobi matrix itself, and the inverse of the diagonal weight matrix to obtain the weighted generalized pseudo-inverse matrix. The system performs matrix multiplication between the weighted generalized pseudo-inverse matrix and the desired state adjustment vector, outputting a basic state adjustment vector that satisfies the macro-control targets of total cost and total project duration.

[0047] In step S305, the system performs a null-space projection optimization operation to generate an optimal adjustment vector for nodes that meets physical construction topology compliance requirements. The system superimposes the basic state adjustment vector onto the continuous system state vector to generate a redistribution intermediate state vector. The system extracts the updated budget funding values ​​for each process node from the redistribution intermediate state vector. The system compares the updated budget funding values ​​with the actual funds consumed by each process node on the physical site. If the system determines that the updated budget funding value is less than the actual funds consumed, it indicates a violation of the budget inversion rule, where the budget funds cannot cover sunk costs. The system compares the progress status parameters of each process node in the process node set with the critical path identifier of the physical project, identifying non-critical nodes that belong to the non-critical path and have budget surplus values. The system constructs a compliance optimization target vector with the same number of dimensions as the total number of process nodes. The system sets the numerical elements of the corresponding non-critical nodes in the compliance optimization target vector to the difference between the initial quota budget baseline and the current budget funding value, and sets the numerical elements of other nodes in the compliance optimization target vector to zero. The system establishes a difference matrix between the identity matrix and the product of the weighted generalized pseudo-inverse matrix and the pseudo-Jacobi matrix, generating a null space projection operator matrix. The system performs matrix multiplication on the null space projection operator matrix and the compliance optimization objective vector to calculate a null space state adjustment vector that does not change the overall project cost or schedule. The system then performs vector addition on the basic state adjustment vector and the null space state adjustment vector, outputting the node-optimal adjustment vector that integrates internal compliance constraints. The node-optimal adjustment vector indicates the amount of additional or reduced funding and schedule days required for each process node to cope with current external market price fluctuations.

[0048] During the system's calculation of fund compliance redistribution, to prevent external market price fluctuations from causing the fund allocation scheme to violate the underlying physical constraints, the system simultaneously performs the following steps to evaluate matrix singularity and determine physical deadlock: In step S401, the system performs singular value decomposition on the pseudo-Jacobi matrix stored in memory. The system extracts the main diagonal matrix generated after singular value decomposition of the pseudo-Jacobi matrix. The system filters all non-zero singular values ​​within the main diagonal matrix. The system performs algebraic multiplication on all filtered non-zero singular values ​​to obtain the operability measure value. The operability measure value quantitatively reflects the redundancy adjustment space of the current budgeting network topology of the power engineering project, absorbing external funding fluctuations while maintaining physical constraints.

[0049] Step S402: The system obtains the singular safety threshold. To adapt to power engineering projects with different investment scales and numbers of nodes, the system reads the initial pseudo-Jacobi matrix generated by the continuous relaxation mapping module. The system performs singular value decomposition on the initial pseudo-Jacobi matrix and calculates the initial operability measure value. The system extracts the standard deviation of historical price fluctuations in the external engineering material market, establishes a nonlinear mapping between the historical price fluctuation standard deviation and the initial operability measure value, and obtains a scaling factor. The system multiplies the initial operability measure value with the nonlinear mapping scaling factor to generate a singular safety threshold that dynamically adapts to the current project scale. The system compares the operability measure value with the singular safety threshold. If the system finds that the operability measure value is greater than or equal to the singular safety threshold, the system maintains the existing multi-degree-of-freedom network topology and performs a capital redistribution solution task through the inverse kinematics optimization solver. If the system finds that the operability measure value is less than the singular safety threshold, it indicates that the column vectors of the pseudo-Jacobi matrix are linearly correlated, and the pseudo-Jacobi matrix is ​​trapped in a singular configuration. Singular configurations are mapped at the physical engineering level as a physical deadlock in the budget network. A physical deadlock indicates that all process nodes under the current work breakdown structure have reached the quota funding limit or the planned schedule limit, the numerical allocation of state variables cannot offset the external funding gap, and executing the numerical allocation will cause the underlying construction operations to violate mandatory engineering safety specifications.

[0050] In step S403, after determining that the system has entered a physical deadlock state, the system performs a bottleneck process node location operation. The system extracts the right singular vector matrix output by the singular value decomposition operation. The system iterates through the numerical elements in the main diagonal matrix to locate the singular value with the smallest value. The system extracts the target column vector corresponding to the smallest singular value from the right singular vector matrix. The system iterates through all elements within the target column vector and calculates the absolute value of all elements. The system sorts the values ​​to select the target element with the largest absolute value and extracts its position index within the target column vector. The system maps and matches the position index with the node identifier of the multi-degree-of-freedom network topology model to locate the bottleneck process node that caused the physical deadlock state of the project. The bottleneck process node identifies the associated work steps in the current construction plan where funding and schedule cannot be further adjusted. The system outputs the node identifier of the bottleneck process node to the variable-degree-of-freedom reconstruction engine, providing a clear physical operation object for dynamically replacing the physical construction scheme.

[0051] After the system identifies the bottleneck process node that causes the physical deadlock in the engineering project, the variable degree of freedom reconfiguration engine performs the following steps to complete the dynamic replacement of the physical construction scheme and the mathematical reconstruction of the system dimension: Step S501: The variable degree of freedom reconfiguration engine receives the node identifier of the bottleneck process node. The engine connects to a pre-defined standardized alternative construction scheme database. This database stores mapping and replacement rules for entity construction processes and corresponding quota funding limits. The engine extracts the original construction process label corresponding to the bottleneck process node based on the node identifier. It then retrieves a set of alternative process nodes from the standardized alternative construction scheme database that match the original construction process label. Each set contains one or more alternative process nodes (i.e., new construction sub-items). Replacing the on-site foundation pouring node with a new combination including prefabricated module assembly and prefabricated module transportation nodes is a specific implementation of calling the alternative process node set.

[0052] In step S502, the variable degree of freedom (DOF) reconstruction engine performs a graph structure topology update operation on the multi-DOF network topology model. The engine cuts all preceding and following directed edges connecting the bottleneck process node to the directed acyclic graph. It removes the bottleneck process node from the process node set. It then adds all alternative process nodes from the alternative process node set to the process node set. The engine reads the alternative process flow templates recorded in the standardized alternative construction scheme database. Based on the alternative process flow templates, the engine establishes new directed edges between the alternative process node and the original preceding and following nodes. After completing the graph structure topology update operation, the total number of nodes in the multi-DOF network topology model expands from the original number to the total number of newly added nodes including the alternative process nodes.

[0053] Step S503: The variable degree-of-freedom (DOF) reconfiguration engine reconstructs the continuous system state vector based on the expanded total number of nodes. The engine extracts the actual funds and time days consumed by the bottleneck process node before its removal, defining these as sunk costs. The engine reads the workload allocation weights of the alternative process nodes from the standardized alternative construction scheme database. Based on these weights, the engine transfers and maps the sunk costs to the alternative process nodes, completing the initialization of the funding and schedule state parameters for the alternative process nodes. This sunk cost transfer mapping ensures the conservation of total funding consumption for the entire power engineering project before and after process replacement.

[0054] In step S504, based on the topology-updated directed acyclic graph and the reconstructed continuous system state vector, the system recalculates the partial derivatives of the forward kinematic mapping model to generate an extended pseudo-Jacobi matrix. The number of columns in the extended pseudo-Jacobi matrix increases synchronously with the number of newly added replacement process nodes, thus expanding the system's mathematical degrees of freedom. Since the attention mechanism and feature aggregation operations of the graph neural network decoding module are independent of the change in the total number of nodes in the directed acyclic graph, the reinforcement learning decision network and the graph neural network decoding module can directly adapt to the expanded multi-degree-of-freedom network topology model without performing a network structure reset operation. The system uses the extended pseudo-Jacobi matrix to re-execute the singular value decomposition operation and calculates a new operability measure value. After confirming that the new operability measure value is greater than or equal to the singular safety threshold, the system releases the physical deadlock state. The system returns to the discrete first-order impedance response model with the extended pseudo-Jacobi matrix to perform a fund compliance redistribution solution operation until the final optimal node adjustment vector is output.

[0055] After the system resolves the physical deadlock state based on the extended pseudo-Jacobi matrix, or outputs the optimal adjustment vector for the node in a non-singular state, the closed-loop command module executes the following steps to complete command issuance and model parameter updates: Step S601: The closed-loop distribution module receives the optimal adjustment vector for the node. The module then inputs the updated node budget amount and schedule value (after overlaying the adjustment vector) into the built-in compliance verification model for the power cost quota standard. The compliance verification model checks whether the updated values ​​exceed the rigid red line of the quota. If the verification is successful, the closed-loop distribution module converts the optimal adjustment vector for the node into a standard-format budget change detail report and construction log update instructions. The closed-loop distribution module then writes the budget change detail report and construction log update instructions into the physical engineering project management information terminal through the system interface, completing the adjustment distribution at the physical engineering level.

[0056] In step S602, the closed-loop distribution module collects cost and schedule deviation data within the actual project execution cycle and constructs a system total reward feedback signal to update the parameters of the reinforcement learning decision network. The closed-loop distribution module compares the actual cost data with the upper limit of the budgeted funds to calculate the cost overrun penalty value, and compares the actual progress data with the planned schedule base days to calculate the schedule delay penalty value. The closed-loop distribution module retrieves the directed acyclic graph and detects whether there are any over-level payment behaviors reversing the directed edges to generate topology violation penalty values. The closed-loop distribution module assigns the highest weight coefficient to the topology violation penalty values. The closed-loop distribution module weights and sums the above three penalty values, outputting a negative system total reward feedback signal.

[0057] In step S603, the system feeds back the total system reward feedback signal to the reinforcement learning decision network. The reinforcement learning decision network uses a policy gradient optimization mechanism to calculate the gradient update vectors of its internal neuron parameters. Based on these gradient update vectors, the reinforcement learning decision network adjusts the connection weights between the action network and the judge network. The system initiates multiple rounds of environmental interaction actions by collecting external environment state vectors, completing the closed-loop iteration of the reinforcement learning decision network parameters.

[0058] For the hidden layer structure of long short-term memory networks in reinforcement learning, the forward and backward propagation algorithms of multilayer perceptrons, the basic algebraic operations of matrix singular value decomposition, the basic equations of message passing in graph neural networks, and the reading and writing operations of relational databases, those skilled in the art can write code and deploy systems based on conventional computer science and computational mathematics theories. These are well-known technologies in the field and will not be elaborated here.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent compilation and dynamic adjustment of power engineering budget estimates based on reinforcement learning, characterized in that, Includes the following steps: The initial data of the power engineering project are analyzed, and a continuously relaxed multi-degree-of-freedom network topology model is constructed to establish a forward kinematic mapping model between the continuous system state vector and the target state vector. The partial derivatives of the forward kinematic mapping model are then solved to extract the pseudo-Jacobi matrix. An external environment state vector containing external market fluctuations and internal deviation data is collected, and the external environment state vector is input into a reinforcement learning decision network and a graph neural network decoding module to generate a virtual damping matrix and a virtual stiffness matrix. The funding gap caused by the external environment state vector is converted into an external impact force vector, and the compliance parameters are extracted and converted into an internal physical constraint force vector. A discrete first-order impedance response model is constructed by combining the pseudo Jacobian matrix, the virtual damping matrix and the virtual stiffness matrix. Based on the discrete first-order impedance response model, the external impact force vector and the internal physical constraint force vector are solved in reverse to output the optimal adjustment vector of the node. During the reverse calculation process, the singularity measure of the pseudo-Jacobi matrix is ​​monitored in real time. When it is determined that a physical deadlock is trapped, a dynamic replacement operation of the physical construction scheme is performed to introduce alternative process nodes. The extended pseudo-Jacobi matrix is ​​re-solved based on the total number of newly added nodes including the alternative process nodes, so as to return to the step of performing the reverse calculation. Based on the optimal adjustment vector of the node, perform cost data compliance verification and issue adjustment instructions; Collect actual deviation data during the actual engineering execution cycle to construct the system total reward feedback signal, and feed the system total reward feedback signal back to the reinforcement learning decision network for iterative update of model parameters.

2. The method according to claim 1, characterized in that, The process involves analyzing initial power engineering data, constructing a continuously relaxed multi-degree-of-freedom network topology model, establishing a forward kinematic mapping model between the continuous system state vector and the target state vector, and solving for the partial derivatives of the forward kinematic mapping model to extract the pseudo-Jacobi matrix. Based on the cost mapping allocation matrix, the amount of funds is projected and integrated onto the corresponding node identifier in the process node set, and the upper limit of the fixed budget funds and the base number of planned construction period days corresponding to each process node in the process node set are extracted; The baseline construction rate constant and the midpoint time parameter of the schedule are calculated using the planned construction period baseline days, and then substituted into the growth curve model to output the continuous progress status parameters. Establish a benchmark consumption baseline based on the product of the continuous progress status parameter and the upper limit of the fixed budget funds. Summate the benchmark consumption baseline with a continuous Gaussian noise variable generated using the standard deviation of historical price fluctuations to generate a continuous funding status parameter. The continuous system state vector is reconstructed by dimensional concatenating the continuous progress state parameter and the continuous funding state parameter. The total cost and total duration of the power engineering project are extracted to construct the target state vector. The forward kinematic mapping model between the target state vector and the continuous system state vector is established. The partial derivatives of the forward kinematic mapping model are calculated to obtain the pseudo-Jacobi matrix that quantifies the sensitivity to local state changes.

3. The method according to claim 1, characterized in that, The process involves collecting an external environment state vector containing data on external market fluctuations and internal deviations. This external environment state vector is then input into a reinforcement learning decision network and a graph neural network decoding module to generate a virtual damping matrix and a virtual stiffness matrix, including: Long Short-Term Memory (LSTM) networks are used to extract fixed-dimensional hidden state features from historical price fluctuation sequence data. The external environment state vector is constructed by combining the internal deviation data between the fixed-dimensional hidden state features, the meteorological warning frequency parameters and the actual consumption funds and the upper limit of the fixed budget funds through vector dimension splicing. The reinforcement learning decision network is used to perform feedforward neural network calculations based on the external environment state vector, and outputs a low-dimensional global policy feature vector.

4. The method according to claim 3, characterized in that, After outputting the low-dimensional global policy feature vector, the following is also included: The low-dimensional global strategy feature vector is distributed and broadcast, and then concatenated with the local continuous state features by vector dimension to generate an initial node feature matrix. The feature vector of the initial node feature matrix is ​​input into a single-layer feedforward neural network to calculate the attention score. The attention score is normalized using a normalized exponential function to obtain the attention weight coefficient. Based on the attention weight coefficient, the feature vectors of adjacent process nodes are weighted and summed to output a high-dimensional local node feature vector. A positive definite nonlinear activation function is connected to the output layer at the end of the independent and parallel multilayer perceptron network branch. The positive definite nonlinear activation function is used to monotonically map the unbounded real values ​​of the network output to positive real number interval values. The positive real number interval values ​​are used to construct a diagonal matrix to generate the virtual damping matrix and the virtual stiffness matrix.

5. The method according to claim 1, characterized in that, The process of converting the funding gap caused by the external environment state vector into an external impact force vector and extracting compliance parameters into an internal physical constraint force vector includes: Using the material node mapping matrix, the material price deviation value is projected and calculated as the budget overrun pressure value of the corresponding process node, and the external impact force vector is generated by summarizing. Calculate the time difference between the planned start time and the start timestamp of the power outage window for each process node, as well as the time difference between the planned completion time and the end timestamp. A maximum repulsive force truncation threshold is set, and a truncated inverse proportional repulsive function is constructed based on the time difference and the artificial potential field algorithm logic. When the output resistance force value exceeds the maximum repulsive force truncation threshold, it is forcibly assigned the maximum repulsive force truncation threshold value. The resistance force values ​​are then summarized to generate the internal physical constraint force vector.

6. The method according to claim 1, characterized in that, The method involves constructing a discrete first-order impedance response model by combining the pseudo-Jacobi matrix, the virtual damping matrix, and the virtual stiffness matrix. Based on this model, the method performs inverse calculations on the external impact force vector and the internal physical constraint force vector, outputting the optimal adjustment vector for each node. Extract the positive real number interval values ​​on the main diagonal of the virtual stiffness matrix to construct a diagonal weight matrix, and perform a combination operation using the transpose of the pseudo-Jacobi matrix, the pseudo-Jacobi matrix itself, and the inverse of the diagonal weight matrix to obtain a weighted generalized pseudo-inverse matrix; The discrete first-order impedance response model is established by combining the virtual damping matrix, the virtual stiffness matrix, the external impact force vector, and the internal physical constraint force vector, and the desired state adjustment vector is obtained by solving the model. Perform a matrix multiplication operation between the weighted generalized pseudo-inverse matrix and the desired state adjustment vector to output the basic state adjustment vector; Non-critical nodes belonging to non-critical paths and having budget surplus values ​​are identified to construct a compliance optimization target vector. The weighted generalized pseudo-inverse matrix and the pseudo-Jacobi matrix are combined to generate a null space projection operator matrix. The null space state adjustment vector is calculated by performing matrix multiplication between the null space projection operator matrix and the compliance optimization target vector. Perform a vector addition operation between the basic state adjustment vector and the null space state adjustment vector to output the optimal adjustment vector for the node.

7. The method according to claim 1, characterized in that, During the reverse calculation process, the singularity measure of the pseudo-Jacobi matrix is ​​monitored in real time. When a physical deadlock is detected, the following steps are taken: Perform singular value decomposition on the pseudo-Jacobi matrix to extract the main diagonal matrix, and perform chain multiplication algebra calculation on all non-zero singular values ​​inside the main diagonal matrix to obtain the operability measure value. Singular value decomposition is performed on the initial pseudo-Jacobi matrix to calculate the initial operability measure value. The standard deviation of historical price fluctuations is extracted, and a nonlinear mapping between the standard deviation of historical price fluctuations and the initial operability measure value is established to obtain the scaling factor. The initial operability measure value is multiplied by the scaling factor to generate the singular safety threshold. The operability measure value is compared with the singularity security threshold. When the operability measure value is less than the singularity security threshold, the pseudo-Jacobi matrix is ​​determined to be trapped in a singular configuration indicating that the predictive network is in the physical deadlock state.

8. The method according to claim 1, characterized in that, The dynamic replacement operation of the physical construction scheme introduces alternative process nodes, including: Extract the right singular vector matrix output by the singular value decomposition operation, locate the singular value with the smallest value in the main diagonal matrix, and extract the target column vector corresponding to the smallest singular value from the right singular vector matrix. The target element with the largest absolute value in the target column vector is selected by numerical sorting, the position index is extracted and mapped and matched with the node identifier, and the bottleneck process node that caused the physical deadlock is locked. Cut off all preceding and following directed edges connected to the bottleneck process node in the directed acyclic graph, remove the bottleneck process node, and add the alternative process nodes retrieved from the standardized alternative construction scheme database into the process node set. A new directed edge containing the alternative process node is established based on the alternative process flow template.

9. The method according to claim 8, characterized in that, The process of recalculating the extended pseudo-Jacobi matrix based on the total number of new nodes including the replacement process nodes includes: The actual amount of funds consumed and the number of days consumed before the bottleneck process node was eliminated are defined as sunk costs. The sunk costs are transferred and mapped to the alternative process node according to the weight of the project quantity and initialized to ensure the conservation of the total amount of funds consumed. The state vector of the continuous system is reconstructed based on the expanded total number of newly added nodes, and the partial derivatives of the forward kinematic mapping model are resolved to generate the extended pseudo-Jacobi matrix.

10. The method according to claim 1, characterized in that, The process of collecting actual deviation data within the actual engineering execution cycle to construct the system's total reward feedback signal, and then feeding the system's total reward feedback signal back to the reinforcement learning decision network for iterative model parameter updates, includes: Compare the actual cost data with the upper limit of the budgeted funds to calculate the cost overrun penalty value, and compare the actual progress data with the planned construction period base days to calculate the construction period delay penalty value; Retrieve the directed acyclic graph, detect whether there is any over-level payment behavior that reverses the actual disbursement time sequence of node funds, and generate a topology violation penalty value based on the amount of funds over-leveled. Assign weight coefficients to each penalty value, and set the weight coefficient of the topology violation penalty value to be greater than the weight coefficient of the cost overrun penalty value and the weight coefficient of the construction period delay penalty value. The penalty values ​​are combined with the configured weight coefficients and weighted summation is performed to output a negative total reward feedback signal of the system, which is then fed back to the reinforcement learning decision network to adjust the connection weights using the policy gradient optimization mechanism.

Images