A power grid engineering cost risk prediction method and system based on time series data modeling

Abstract

The application discloses a power grid engineering cost risk prediction method and system based on time series data modeling, relates to the technical field of power grid engineering risk prediction, and obtains continuous material price data and discrete node event sequences of a power grid engineering, reconstructs a node time sequence deformation feature vector through time-frequency transformation and anchor point alignment, maps the node time sequence deformation feature vector into a boundary penalty coefficient of a dynamic time warping matrix, generates a cost exposure offset probability tensor containing a lag variable through time axis distortion and reorganization, intercepts a tensor segment to calculate a lead-lag cross entropy, extracts an asymmetric risk conduction feature to generate a dynamically evolving directed cost risk correlation graph sequence, synchronously inputs the graph sequence and the probability tensor into a space-time convolution channel, generates a space-time joint vector through multi-order residual aggregation of a gate unit, and outputs a dynamic cost overrun probability sequence at a preset time step through inverse convolution decoding of the space-time joint vector. The system is used for implementing the above method.

Description

Technical Field

[0001] This invention relates to the field of power grid engineering risk prediction technology, specifically to a method and system for predicting power grid engineering cost risks based on time series data modeling. Background Technology

[0002] Power grid construction projects have long construction cycles and involve the procurement of large quantities of materials. Their cost estimation is affected by both macroeconomic market fluctuations in material prices and changes in on-site construction progress. During the long project duration, the prices of core materials exhibit continuous, high-frequency fluctuations, while on-site delays often occur due to objective factors such as weather changes or work transitions. Accurately predicting cost changes throughout the project cycle is fundamental to ensuring the safe flow of funds in digital power grid construction. With the increasing informatization of engineering projects, cost control terminals have accumulated massive amounts of continuous time-series material price data and discrete project status logs. How to process this high-dimensional, heterogeneous time-series data to accurately analyze the actual probability of cost overruns has become a key aspect of cost risk control.

[0003] Existing cost analysis methods often employ static time windows to perform linear regression calculations on continuous price characteristics, or physically isolate material prices and construction progress for separate calculations. In actual engineering data interaction, discrete construction delays force nonlinear shifts in actual procurement nodes, causing misalignment and overlap between the planned procurement period and the high-frequency fluctuation peaks of actual material prices. Existing static regression networks cannot perform underlying alignment and time-axis distortion resampling operations on this spatiotemporally asynchronous misaligned data. Furthermore, when dealing with the cost characteristics of multiple material categories, conventional analysis networks exhibit symmetrical and static tendencies in handling the cross-influence between various factors, failing to analyze the asymmetric lead-lag transmission relationships between different materials. This makes the model highly susceptible to local warning distortion and feature transmission interruptions when facing short-term non-stationary price noise and long-term cascading reactions. Therefore, it is still necessary to provide a power grid engineering cost risk prediction method based on time-series data modeling to improve the fusion and alignment capabilities of multi-source asynchronous time-series data processing and the output accuracy of long-term warning sequences. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for predicting the cost risk of power grid engineering based on time-series data modeling to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the cost risk of power grid engineering based on time-series data modeling, comprising the following steps: S1, acquiring continuous material price data and discrete construction node event sequences of power grid engineering; performing time-frequency transformation on the continuous material price data to separate high-frequency fluctuation component sequences, and anchoring the discrete construction node event sequences with a preset construction baseline as a reference to reconstruct the node time-series deformation feature vector; S2, mapping the node time-series deformation feature vector to the boundary penalty coefficient of a dynamic time warping matrix to construct deformation time constraints; using the deformation time constraints to perform time axis distortion and reorganization on the high-frequency fluctuation component sequences to generate a cost exposure offset probability tensor containing lag variables; S3, on the time axis... A multi-scale sliding window is set up to extract cost exposure offset probability tensor fragments corresponding to multiple materials. The lead-lag cross-entropy between any two cost exposure offset probability tensor fragments is calculated. The asymmetric risk transmission features in the lead-lag cross-entropy are extracted as network edge weights to generate a dynamically evolving directed cost risk correlation graph sequence. S4. The directed cost risk correlation graph sequence is input into a spatial graph convolution kernel to generate a topology aggregation path. The cost exposure offset probability tensor with lag variables is input into a temporal convolution kernel to generate a time-aware sequence. The topology aggregation path and the time-aware sequence are synchronously input into an adaptive gated fusion unit for multi-order residual aggregation to generate a spatiotemporal joint vector. The spatiotemporal joint vector is deconvolved and decoded to output a dynamic cost overrun probability sequence at a preset time step.

[0006] In a preferred embodiment, the specific process of obtaining continuous material price data and discrete construction node event sequences for a power grid project is as follows: Obtain material price flow data from a bulk commodity trading terminal; extract the daily settlement price sequence and bulk transaction volume sequence from the material price flow data; concatenate the daily settlement price sequence and bulk transaction volume sequence to generate continuous material price data; obtain project log text data and external weather warning text data from a project construction management terminal; perform natural language entity recognition on the project log text data and external weather warning text data to extract a work stoppage status label sequence; perform discretized one-hot encoding mapping on the work stoppage status label sequence according to the timestamp sequence of the project log text data and external weather warning text data to generate a discrete construction node event sequence.

[0007] In a preferred embodiment, the specific process of performing time-frequency transformation on continuous material price data to separate high-frequency fluctuation component sequences, and anchoring discrete construction node event sequences with a preset construction baseline as a reference to reconstruct the node temporal deformation feature vector is as follows: Empirical mode decomposition is performed on continuous material price data to generate multiple intrinsic mode function component sequences; instantaneous frequency measurement is performed on multiple intrinsic mode function component sequences to extract target high-frequency mode function component sequences, and multiple target high-frequency mode function component sequences are fused to generate high-frequency fluctuation component sequences; actual state transition time nodes of discrete construction node event sequences are extracted, and planned milestone time nodes of the preset construction baseline are extracted; phase deviation measurement is performed on the actual state transition time nodes and planned milestone time nodes to generate a multidimensional phase deviation sequence, and the multidimensional phase deviation sequence is mapped to a node temporal deformation feature vector.

[0008] In a preferred embodiment, the specific process of mapping the node temporal deformation feature vector to the boundary penalty coefficient of the dynamic time warping matrix and constructing the deformation time constraint is as follows: Initialize the two-dimensional cost distance grid based on the temporal length of the high-frequency fluctuation component sequence and the temporal length of the preset construction baseline; parse the nonlinear offset coefficient in the node temporal deformation feature vector, input the nonlinear offset coefficient into the asymmetric activation function for nonlinear projection to generate a dynamic penalty weight sequence; map the dynamic penalty weight sequence along the diagonal search path of the two-dimensional cost distance grid to construct the boundary penalty coefficient of the dynamic time warping matrix; set the global search boundary width limit on the dynamic time warping matrix with the boundary penalty coefficient to generate the deformation time constraint.

[0009] In a preferred embodiment, the specific process of generating a cost exposure offset probability tensor containing lag variables by time-axis distortion and reorganization of the high-frequency fluctuation component sequence under deformation time constraints is as follows: Dynamic programming is performed along the global search boundary width limit defined by the deformation time constraints in a dynamic time warping matrix with boundary penalty coefficients to extract the minimum cumulative distance path; the minimum cumulative distance path is used as a time remapping function to perform nonlinear interpolation resampling on the high-frequency fluctuation component sequence to generate a deformed high-frequency fluctuation component sequence; a Gaussian perturbation operator is introduced at each time step of the deformed high-frequency fluctuation component sequence to perform Monte Carlo probability sampling to generate a multidimensional probability distribution matrix; multiple multidimensional probability distribution matrices are stacked in three dimensions according to the power grid engineering material category dimension to generate a cost exposure offset probability tensor containing lag variables.

[0010] In a preferred embodiment, the specific process of setting a multi-scale sliding window on the time axis, extracting cost exposure offset probability tensor segments corresponding to multiple materials, and calculating the lead-lag cross-entropy between any two cost exposure offset probability tensor segments is as follows: A multi-scale sliding window is generated by configuring a multi-band observation step size. The cost exposure offset probability tensor containing lag variables is dynamically sliced ​​on the time axis using the multi-scale sliding window to extract cost exposure offset probability tensor segments corresponding to multiple materials. A multi-level time delay operator is introduced between any two cost exposure offset probability tensor segments for misalignment alignment, and the marginal probability density distribution characteristics under misalignment alignment are analyzed. A joint distribution integral calculus is performed on the marginal probability density distribution characteristics to extract divergence measure variables, and these divergence measure variables are mapped to the lead-lag cross-entropy between any two cost exposure offset probability tensor segments.

[0011] In a preferred embodiment, the specific process of extracting the asymmetric risk transmission features from the lead-lag cross-entropy as network edge weights to generate a dynamically evolving directed cost-risk association graph sequence is as follows: Directional gradient analysis is performed on the lead-lag cross-entropy to separate the forward and backward transmission probability components; the forward and backward transmission probability components are input into an asymmetric activation kernel for significance testing, and the topological response variables that pass the significance test are extracted as asymmetric risk transmission features; the entity codes corresponding to multiple materials are extracted as a network node matrix, and the asymmetric risk transmission features are mapped to network edge weights and filled into the network node matrix; Markov state iteration updates are performed on the network node matrix filled with network edge weights along the time axis to generate a dynamically evolving directed cost-risk association graph sequence.

[0012] In a preferred embodiment, the specific process of inputting the directed cost-risk association graph sequence into a spatial graph convolution kernel to generate a topological aggregation path, and inputting the cost exposure offset probability tensor containing lagged variables into a temporal convolution kernel to generate a time-aware sequence is as follows: Extracting the Laplace smoothing operator within the spatial graph convolution kernel, and using the Laplace smoothing operator to perform spectral domain mapping on the directed cost-risk association graph sequence to analyze its spatial spectral features; performing multi-hop node neighborhood message passing aggregation on the spatial spectral features along the network edge weights of the directed cost-risk association graph sequence to generate a topological aggregation path; activating the dilated causal convolution channel within the temporal convolution kernel, and inputting the cost exposure offset probability tensor containing lagged variables into the dilated causal convolution channel for long-range receptive field sequence penetration analysis; extracting the hidden state feature vector output from the long-range receptive field sequence penetration analysis, and concatenating the hidden state feature vector along the time dimension to generate a time-aware sequence.

[0013] In a preferred embodiment, the specific process of synchronously inputting the topology aggregation path and the time-aware sequence into an adaptive gated fusion unit for multi-order residual aggregation to generate a spatiotemporal joint vector, and then performing deconvolution decoding mapping on the spatiotemporal joint vector to output the dynamic cost overrun probability sequence at a preset time step is as follows: Initialize the spatial forgetting gate parameter and the time updating gate parameter within the adaptive gated fusion unit; input the topology aggregation path into the spatial forgetting gate parameter to parse the spatial attention weight tensor, and input the time-aware sequence into the time updating gate parameter to parse the time attention weight tensor; use the spatial attention weight tensor and the time attention weight tensor to perform tensor inner product fusion on the topology aggregation path and the time-aware sequence, and perform multi-order residual aggregation through a jump connection operator to generate the spatiotemporal joint vector; input the spatiotemporal joint vector into the deconvolution decoding network for high-dimensional feature projection, mapping it to the joint tensor space of the preset time step and the cost overrun interval, and outputting the dynamic cost overrun probability sequence at the preset time step.

[0014] A power grid engineering cost risk prediction system based on time-series data modeling, used to execute the aforementioned power grid engineering cost risk prediction method based on time-series data modeling, includes: a feature reconstruction module, used to acquire continuous material price data and discrete construction node event sequences of the power grid engineering; to perform time-frequency transformation on the continuous material price data to separate high-frequency fluctuation component sequences, and to perform anchor point alignment on the discrete construction node event sequences with a preset construction baseline as a reference, reconstructing the node time-series deformation feature vector; a tensor generation module, used to map the node time-series deformation feature vector to the boundary penalty coefficient of a dynamic time warping matrix, constructing deformation time constraints; to perform time axis distortion and reorganization on the high-frequency fluctuation component sequences through deformation time constraints, generating a cost exposure offset probability tensor containing hysteresis variables; and a graph construction module. The first module sets up a multi-scale sliding window on the time axis to extract cost exposure offset probability tensor fragments corresponding to multiple materials. It calculates the lead-lag cross-entropy between any two cost exposure offset probability tensor fragments, extracts the asymmetric risk transmission features from the lead-lag cross-entropy as network edge weights, and generates a dynamically evolving directed cost risk correlation graph sequence. The second module inputs the directed cost risk correlation graph sequence into a spatial graph convolution kernel to generate a topology aggregation path, and inputs the cost exposure offset probability tensor with lag variables into a temporal convolution kernel to generate a time-aware sequence. The topology aggregation path and the time-aware sequence are synchronously input into an adaptive gated fusion unit for multi-order residual aggregation to generate a spatiotemporal joint vector. The spatiotemporal joint vector is deconvolved and decoded to output a dynamic cost overrun probability sequence at a preset time step.

[0015] The technical effects and advantages of this invention are as follows:

[0016] (1) A method for predicting the cost risk of power grid projects based on time-series data modeling. Before performing the prediction, continuous material price data and discrete construction node event sequences of power grid projects are obtained. The continuous material price data is subjected to time-frequency transformation to separate the high-frequency fluctuation component sequence. The discrete construction node event sequence is reconstructed by anchoring and aligning with a preset construction baseline. The node time-series deformation feature vector is then mapped to the boundary penalty coefficient of the dynamic time warping matrix, and deformation time constraints are constructed to perform time axis distortion and reorganization on the high-frequency fluctuation component sequence. Thus, the problem that existing conventional analysis networks cannot perform low-level alignment operations on spatiotemporally asynchronous misaligned data can be overcome. The model can directly transform the discrete construction progress delay features into time axis distortion mapping operations on continuous material prices, generate a cost exposure offset probability tensor with lagged variables that fits the actual business procurement fluctuation features, and improve the structural integration of cross-modal time-series data processing.

[0017] (2) A power grid engineering cost risk prediction system based on time-series data modeling, which sets a multi-scale sliding window on the time axis to extract cost exposure offset probability tensor segments corresponding to multiple materials, calculates the lead-lag cross-entropy between any two cost exposure offset probability tensor segments to extract asymmetric risk transmission features, and generates a dynamically evolving directed cost risk correlation graph sequence; then, the directed cost risk correlation graph sequence and the cost exposure offset probability tensor with lag variables are respectively input into the spatial graph convolution kernel and the temporal convolution kernel, and multi-order residual aggregation is synchronously executed through an adaptive gated fusion unit to generate a spatiotemporal joint vector and implement deconvolution decoding mapping. Thus, it can improve the shortcomings of conventional networks in being unable to analyze asymmetric chain reactions between materials and easily generating local early warning distortions, and use multi-order residual aggregation to achieve adaptive verification and filtering of global long-range features and local topological features, thereby outputting a high-precision preset time step dynamic cost overrun probability sequence, enhancing the response performance of multi-node engineering data mining in complex environments.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] Figure 1 This is a flowchart of a power grid engineering cost risk prediction method based on time series data modeling according to the present invention;

[0020] Figure 2 This is a three-dimensional schematic diagram of the cost exposure offset probability tensor in an embodiment of the present invention;

[0021] Figure 3 This is a schematic diagram of the three-dimensional dynamic topology of the directed cost-risk association graph sequence in an embodiment of the present invention;

[0022] Figure 4This is a flowchart of a power grid engineering cost risk prediction system based on time series data modeling according to the present invention. Detailed Implementation

[0023] This application provides a method and system for predicting the cost and risk of power grid engineering based on time-series data modeling. This solves the problems that conventional regression models cannot perform nonlinear time alignment and recombination of spatiotemporally misaligned multi-source asynchronous engineering data, and are difficult to analyze asymmetric transmission chain reactions between multiple factors.

[0024] The overall approach of the scheme in this application embodiment is as follows: 1. Obtain continuous material price data and discrete construction node event sequences for power grid engineering. 2. Perform time-frequency transformation on the continuous material price data to separate high-frequency fluctuation component sequences. 3. Perform anchor point alignment on the discrete construction node event sequences using a preset construction baseline as a reference, reconstructing the node time-series deformation feature vector. 4. Map the node time-series deformation feature vector to the boundary penalty coefficient of a dynamic time warping matrix, constructing deformation time constraints. 5. Perform time axis distortion and reorganization on the high-frequency fluctuation component sequences using deformation time constraints, generating a cost exposure offset probability tensor containing hysteresis variables. 6. Set a multi-scale sliding window on the time axis to extract the costs corresponding to multiple materials. The cost exposure offset probability tensor fragments are used to calculate the lead-lag cross-entropy between any two cost exposure offset probability tensor fragments. The asymmetric risk transmission features in the lead-lag cross-entropy are extracted as network edge weights to generate a dynamically evolving directed cost risk correlation graph sequence. The directed cost risk correlation graph sequence is input into a spatial graph convolution kernel to generate a topology aggregation path. The cost exposure offset probability tensor containing lag variables is input into a temporal convolution kernel to generate a time-aware sequence. The topology aggregation path and the time-aware sequence are synchronously input into an adaptive gated fusion unit to perform multi-order residual aggregation, generating a spatiotemporal joint vector. The spatiotemporal joint vector is deconvolutionally decoded and mapped to output a dynamic cost overrun probability sequence at a preset time step.

[0025] Example 1; please refer to Figure 1This invention provides a technical solution: a method for predicting the cost risk of power grid engineering based on time-series data modeling, comprising the following steps: S1, acquiring continuous material price data and discrete construction node event sequences of power grid engineering; performing time-frequency transformation on the continuous material price data to separate high-frequency fluctuation component sequences, and anchoring the discrete construction node event sequences with a preset construction baseline as a reference to reconstruct the node time-series deformation feature vector; S2, mapping the node time-series deformation feature vector to the boundary penalty coefficient of a dynamic time warping matrix to construct deformation time constraints; using the deformation time constraints to perform time axis distortion and reorganization on the high-frequency fluctuation component sequences to generate a cost exposure offset probability tensor containing lag variables; S3, setting on the time axis... A multi-scale sliding window is used to extract cost exposure offset probability tensor fragments corresponding to multiple materials. The lead-lag cross-entropy between any two cost exposure offset probability tensor fragments is calculated, and the asymmetric risk transmission features in the lead-lag cross-entropy are extracted as network edge weights to generate a dynamically evolving directed cost risk correlation graph sequence. S4: The directed cost risk correlation graph sequence is input into a spatial graph convolution kernel to generate a topology aggregation path. The cost exposure offset probability tensor containing lag variables is input into a temporal convolution kernel to generate a time-aware sequence. The topology aggregation path and the time-aware sequence are synchronously input into an adaptive gated fusion unit for multi-order residual aggregation to generate a spatiotemporal joint vector. The spatiotemporal joint vector is deconvolved and decoded to output a dynamic cost overrun probability sequence at a preset time step.

[0026] In this implementation plan, step S1 is used to perform low-level analysis and feature structuring of heterogeneous raw business data in power grid engineering. Continuous material price data reflects daily market conditions for commodities, while discrete construction node event sequences record the stoppage or delay status in the actual construction log. Time-frequency transformation refers to using signal processing algorithms to convert the price line changing over time into a set of waveforms of different frequencies. The high-frequency fluctuation component sequence extracted from this represents drastic price jumps caused by short-term market supply and demand, which are not affected by long-term macroeconomic factors. Anchor point alignment uses the initial project schedule as a reference coordinate system and fills the corresponding actual node events into the coordinate system. The node time sequence deformation feature vector converts the deviation between the planned time and the actual occurrence time into a mathematical vector that can be read by a computer, and is used to directly quantify the degree of delay in project progress.

[0027] Step S2 transforms the physical delays in project progress into a mathematical probability model of the material procurement node shifting on the price fluctuation curve. The dynamic time warping matrix is ​​an algorithmic matrix that handles two time series of different lengths or rhythms; the boundary penalty coefficient limits the maximum tolerance range of this matrix when matching data. The deformation time constraint sets the eigenvectors analyzed in step S1 as the extreme boundary of the algorithm's tolerance. Time axis distortion and reorganization refers to the algorithm, under constraints, forcibly stretching or compressing the originally planned procurement dates into the actual date interval where delays occur. After nonlinear time matching calculation, the output cost exposure offset probability tensor is represented as a multidimensional spatial matrix containing time lag variables, quantifying the probability distribution of a certain material encountering specific high-frequency price jumps within a certain period in the future.

[0028] Step S3 is used to uncover the hidden chain reactions and sequential relationships among price fluctuations of different materials in power grid projects. A multi-scale sliding window is a data truncation range set at different time lengths on the time axis to observe the variable state under different periods. Lead-lag cross-entropy is an algorithmic measure of the information difference and temporal sequence relationship between two probability distributions. Calculating the cross-entropy of any two tensor segments can extract asymmetric risk transmission characteristics, that is, distinguish the difference in the unidirectional transmission strength of price fluctuations of one material leading to price fluctuations of another material. Network edge weights represent the numerical value of this unidirectional transmission strength. The directed cost risk association graph sequence generated by combining node features and network edge weights constitutes a graph data topology that dynamically changes with the project progress, reconstructing the physical path of various material cost risks transferring over time in the business chain.

[0029] Step S4 performs deep learning forward propagation computation in both spatial correlation and temporal memory dimensions to output the final early warning result. The spatial graph convolutional kernel extracts the spatial interaction values ​​between various material nodes along the topological correlation path generated in step S3, forming a topological aggregation path. The temporal convolutional kernel synchronously processes the tensor features generated in step S2 to extract the memory patterns of long-term and short-term price evolution, forming a time-aware sequence. The adaptive gated fusion unit is a neural network component containing a parameter filtering structure. Multi-order residual aggregation relies on cross-layer data transfer mechanisms to prevent gradient vanishing in deep network computation. Both are numerically superimposed within this unit according to feature sensitivity to generate a spatiotemporal joint vector. The deconvolutional decoding mapping performs a dimensionality-upgrading operation, converting the high-dimensional hidden state numerical features back into probability values ​​of the engineering business dimension, outputting a percentage sequence of future cost overruns for specific engineering nodes.

[0030] Specifically, the process of obtaining continuous material price data and discrete construction node event sequences for power grid projects is as follows: Obtain material price flow data from the bulk commodity trading terminal, extract the daily settlement price sequence and bulk transaction volume sequence from the material price flow data, and concatenate the daily settlement price sequence and bulk transaction volume sequence to generate continuous material price data; obtain engineering log text data and external meteorological warning text data from the engineering construction management terminal, and perform natural language entity recognition on the engineering log text data and external meteorological warning text data to extract the shutdown status label sequence; perform discretized one-hot encoding mapping on the shutdown status label sequence according to the timestamp sequence of the engineering log text data and external meteorological warning text data to generate discrete construction node event sequences.

[0031] In this implementation plan, the system accesses the commodity trading platform and the enterprise's engineering management database through an authorized dedicated data interface. During the reading process, a pre-processing data cleaning script is triggered to perform hash desensitization and one-way encryption on fields involving sensitive enterprise information and personnel privacy. After acquiring the material price flow data, the daily settlement price sequence is extracted. With block trade volume sequence ,in, This represents the time step number of the material flow data. The system performs a concatenation operation along the column direction of the data matrix to generate continuous material price data. ; This represents the matrix transpose operation. For the acquired engineering log text data and external weather warning text data, the system inputs them in segments into a pre-trained conditional random field network, parses the text context associations, and outputs the corresponding state entity labels. When a label is determined to belong to a preset construction obstruction category, the timestamp corresponding to that entity is extracted. An initial zero vector with a dimension equal to the preset total length of the discrete time grid is constructed. The extracted timestamps are aligned to the discrete time grid scale. When a work stoppage occurs at the y-th scale, the corresponding element in the initial zero vector is assigned the constant 1, while the remaining elements remain constant 0, generating a discrete construction node event sequence. ,in Indicates the scale number of the discrete-time grid.

[0032] Specifically, the process of performing time-frequency transformation on continuous material price data to separate high-frequency fluctuation component sequences, and anchoring discrete construction node event sequences with a preset construction baseline as a reference to reconstruct the node temporal deformation feature vector is as follows: Empirical mode decomposition (EMD) is performed on continuous material price data to generate multiple intrinsic mode function (IMF) component sequences; instantaneous frequency measurement is performed on multiple IMF component sequences to extract target high-frequency mode function (MMF) component sequences, and these multiple target high-frequency MMF component sequences are fused to generate a high-frequency fluctuation component sequence; actual state transition time nodes of discrete construction node event sequences are extracted, as well as planned milestone time nodes of the preset construction baseline; phase deviation measurement is performed on the actual state transition time nodes and planned milestone time nodes to generate a multidimensional phase deviation sequence, which is then mapped to a node temporal deformation feature vector.

[0033] In this implementation plan, continuous material price data Perform empirical mode decomposition, calculate local extrema and fit upper and lower envelopes. Iteratively filter the original sequence by subtracting the mean of the envelope until multiple intrinsic mode function component sequences satisfying the stationarity condition are extracted. With residual trend term sequence , making ,in Indicates the hierarchical number of the mode decomposition; This represents the total number of levels for the modal components. For each intrinsic mode function component sequence... Perform Hilbert transform to obtain its conjugate signal Calculating instantaneous phase sequence based on arctangent function and the instantaneous phase sequence Obtain the instantaneous frequency sequence by calculating the time derivative. Define the high-frequency screening threshold parameter. This high-frequency screening threshold parameter The method for determining this is to calculate the energy spectral density-weighted average frequency of all intrinsic mode function components over the entire observation period. The system then extracts the instantaneous frequency sequence. Greater than the high-frequency screening threshold parameter The components are used as the target high-frequency mode function component sequence, and are superimposed point by point according to time steps to generate a high-frequency fluctuation component sequence. Subsequently, the discrete construction node event sequence is traversed. The tick marks of the non-zero elements are analyzed to reconstruct the actual state transition time points. ,in This indicates the sequence number of the actual construction event. It extracts the planned milestone time nodes from the preset construction baseline. The phase deviation between the two is calculated to generate a multidimensional phase deviation sequence. The calculation formula is: ,in Represents an exponential function with the natural constant as its base; This represents the node sensitivity decay constant. The phase deviation sequence was determined by performing least-squares fitting on the schedule delay loss records of similar historical projects. Finally, the multidimensional phase deviation sequence was... Sort by number The increasing order is mapped to a contiguous memory space and reconstructed into node temporal deformation feature vectors. .

[0034] Specifically, the process of mapping the nodal temporal deformation feature vectors to the boundary penalty coefficients of the dynamic time warping matrix and constructing deformation time constraints is as follows: Initialize a two-dimensional cost distance grid based on the temporal length of the high-frequency fluctuation component sequence and the temporal length of the preset construction baseline; parse the nonlinear offset coefficients in the nodal temporal deformation feature vectors, input the nonlinear offset coefficients into an asymmetric activation function for nonlinear projection to generate a dynamic penalty weight sequence; map the dynamic penalty weight sequence along the diagonal search path of the two-dimensional cost distance grid to construct the boundary penalty coefficients of the dynamic time warping matrix; and set a global search boundary width limit on the dynamic time warping matrix with the boundary penalty coefficients to generate deformation time constraints.

[0035] In this implementation scheme, the system reads the total number of elements in the aforementioned generated high-frequency fluctuation component sequence. and the total number of scale markings on the pre-set construction baseline. Allocate a space of size in the underlying storage. A two-dimensional floating-point array is used as a two-dimensional cost distance grid. The nonlinear offset coefficients are extracted from the nodal temporal deformation feature vectors and denoted as... , where k represents the dimension index number of the deformation feature. The nonlinear offset coefficients... The input is fed into a preset asymmetric activation function to perform a nonlinear projection calculation, generating a dynamic penalty weight sequence. The calculation formula is: ; The function represents the natural logarithm; e represents the natural constant. Represents the basic scaling constant; This represents the slope control coefficient; This represents the positive asymmetric gain parameter. The method for determining this is to calculate the ratio of the peak price fluctuation caused by delayed material delivery to the peak price during the normal delivery period in three historical phases of similar power grid projects. A dynamic penalty weight sequence is then determined. Then, the coordinate system is projected and mapped along the diagonal of the two-dimensional cost distance grid to construct the boundary penalty coefficient of the dynamic time warping matrix. The formula is u represents the time step index of the high-frequency fluctuation component sequence; v represents the time step index of the preset construction baseline. This represents the weight values ​​after the dynamic penalty weight sequence is interpolated and aligned to the horizontal axis of the grid; This represents the spatial distance penalty index. It is the boundary penalty coefficient. Based on this, set the global search boundary width limit. The boundary determination formula is: This decision formula is then used to generate deformation-time constraints that allow path optimization only within a limited width. Global search boundary width limit. The method for determining it is to calculate the standard deviation sequence of all nonlinear offset coefficients in the node temporal deformation eigenvector and take twice its arithmetic mean.

[0036] Specifically, the process of generating a cost exposure offset probability tensor with lag variables by time-axis distortion and reorganization of the high-frequency fluctuation component sequence under deformation time constraints is as follows: Dynamic programming is performed along the global search boundary width limit defined by the deformation time constraints in a dynamic time warping matrix with boundary penalty coefficients to extract the minimum cumulative distance path; the minimum cumulative distance path is used as a time remapping function to perform nonlinear interpolation resampling on the high-frequency fluctuation component sequence to generate a deformed high-frequency fluctuation component sequence; a Gaussian perturbation operator is introduced at each time step of the deformed high-frequency fluctuation component sequence to perform Monte Carlo probability sampling to generate a multidimensional probability distribution matrix; multiple multidimensional probability distribution matrices are stacked in three dimensions according to the power grid engineering material category dimension to generate a cost exposure offset probability tensor with lag variables.

[0037] In this implementation plan, a boundary penalty coefficient is incorporated. Within the dynamic time warp matrix, dynamic programming calculations are performed according to the range defined by the deformation time constraint. The local distance cost of the mesh nodes is defined as... This represents the difference between the absolute value of the high-frequency fluctuation component sequence at index u and the absolute value of the preset construction baseline at index v. Cumulative distance matrix. The recursive solution formula is as follows ;min represents the minimum value function. Calculate the value reaching the grid endpoint. Then, the combination of grid coordinate sequences with the minimum cumulative distance is extracted by reverse backtracking and used as the minimum cumulative distance path. The path with the minimum cumulative distance. As a time remapping function, spline interpolation resampling is performed on the high-frequency fluctuation component sequence, so that the original time axis follows the path. The coordinate mapping relationship is forced to undergo a distorted mapping, and the output is aligned to the baseline dimension of the high-frequency fluctuation component sequence of deformation. Traversing the high-frequency fluctuation component sequence of deformation Each time step Applying a Gaussian perturbation operator to perform Monte Carlo probability sampling generates a multidimensional probability distribution matrix. The parsing process is as follows ;h represents the simulated batch number of the Monte Carlo sampling; This indicates that the h-th sample follows a mean of 0 and a variance of . Normally distributed random values; The dynamic volatility parameter at the current time step is determined by extracting the moving standard deviation of the high-frequency volatility component sequence within ten consecutive time steps before and after time step u. For multiple different categories of materials in the power grid engineering bill of materials, the aforementioned time remapping and probability sampling calculations are performed in parallel to obtain their respective multidimensional probability distribution matrices. Based on the power grid engineering material category dimension g, all multidimensional probability distribution matrices are orthogonally stitched and stacked in three-dimensional tensors in memory space to generate a cost exposure offset probability tensor containing hysteresis variables. Please see. Figure 2 Combining the aforementioned calculation process with Figure 2 As can be seen from the three-dimensional coordinate system, the horizontal axis of the bottom grid corresponds to the time step mapped to the preset construction baseline, the vertical axis corresponds to the simulated batch of Monte Carlo sampling, and the absolute value of the price fluctuation component after the vertical bearing load sampling. For example... Figure 2 As shown in the figure, the multiple irregular broken line trajectories are not simple translations of the original historical price curves. Specifically, after the system extracts discrete shutdown events and constructs constraints, the algorithm strictly follows the path of minimum cumulative distance. Nonlinear interpolation resampling was performed on continuous material prices, which directly led to Figure 2 The broken line in the graph exhibits significant lateral stretching and compression (i.e., peak misalignment) within a specific time step interval, reconstructing the physical delays in procurement milestones caused by project schedule delays. Furthermore, due to the addition of a Gaussian perturbation operator at each time step, Figure 2 The multiple polylines in the model exhibit probability divergence with defined variance limits on the vertical plane of the same time scale. This physical expansion from a single deterministic temporal polyline to a cluster of three-dimensional batch surfaces is the visual manifestation of tensor stacking. This tensor matrix transforms a single delayed event into a high-dimensional data block containing fluctuation tolerance margins, thus providing a continuous and smooth observation sequence for the temporal causal convolutional layer in the subsequent step S4, avoiding gradient discontinuities in the neural network caused by extreme value misalignment.

[0038] Specifically, the process of setting up a multi-scale sliding window on the time axis, extracting cost exposure offset probability tensor segments corresponding to multiple materials, and calculating the lead-lag cross-entropy between any two cost exposure offset probability tensor segments is as follows: A multi-scale sliding window is generated by configuring a multi-band observation step size. The cost exposure offset probability tensor containing lag variables is dynamically sliced ​​on the time axis using the multi-scale sliding window, extracting cost exposure offset probability tensor segments corresponding to multiple materials. A multi-level time delay operator is introduced between any two cost exposure offset probability tensor segments for misalignment alignment, and the marginal probability density distribution characteristics under misalignment alignment are analyzed. A joint distribution integral calculation is performed on the marginal probability density distribution characteristics to extract the divergence measure variable, and the divergence measure variable is mapped to the lead-lag cross-entropy between any two cost exposure offset probability tensor segments.

[0039] In this implementation scheme, the system allocates multiple sets of incremental time step values ​​in memory as multi-band observation step sizes to construct a multi-scale sliding window set. The system uses this multi-scale sliding window set to perform dynamic slicing calculations on the previously generated cost exposure offset probability tensor on the time axis, extracting cost exposure offset probability tensor fragments corresponding to multiple materials. For any two selected different materials, category indices a and b are assigned respectively, and multi-level time delay operators are configured between the tensor fragment coordinate systems of the two materials. The system performs a time-axis misalignment alignment operation. It employs a kernel density estimation algorithm to analyze the tensor fragments in the misaligned state, outputting continuous edge probability density distribution features, denoted as follows: and a and b represent category indexes for two different types of power grid engineering materials; The step size variable represents the multi-level time delay operator; Represents a continuous random independent variable in the probability integral space; Indicates material Characteristic function of the marginal probability density distribution of tensor fragments; Indicates material Tensor fragments pass through The system then performs a joint distribution integral calculus on the marginal probability density distribution feature function after step-size misalignment and alignment to extract the divergence measure variable. Its calculation formula is expressed as ; This represents the interval for solving the joint integral in the probability density domain; This represents the natural logarithm calculus function. The system iterates through the set multi-level delay steps, selecting the variable that best represents the divergence measure. Obtain the optimal delay state with absolute extremum, and save the numerical mapping of this state as the lead-lag cross-entropy between the two types of materials. The publicly available method for determining the maximum search step size of this multi-level time delay operator is to statistically analyze the number of observation points corresponding to the longest procurement cycle spanned from the commencement of civil construction to the completion and acceptance of core electrical equipment in historical power grid projects of similar scale.

[0040] Specifically, the process of extracting asymmetric risk transmission features from the lead-lag cross-entropy as network edge weights to generate a dynamically evolving directed cost-risk association graph sequence is as follows: Directional gradient analysis is performed on the lead-lag cross-entropy to separate the forward and backward transmission probability components; the forward and backward transmission probability components are input into an asymmetric activation kernel for significance testing, and the topological response variables that pass the significance test are extracted as asymmetric risk transmission features; the entity codes corresponding to multiple materials are extracted as a network node matrix, and the asymmetric risk transmission features are mapped to network edge weights and filled into the network node matrix; Markov state iteration updates are performed on the network node matrix filled with network edge weights along the time axis to generate a dynamically evolving directed cost-risk association graph sequence.

[0041] In this implementation scheme, the system calculates the lead-lag cross-entropy obtained from the preceding calculation. Executing directional gradient analysis yields Where p and q represent the starting material row index and ending material column index in the association matrix architecture, respectively. The system uses gradient difference logic to separate the forward propagation probability component. With reverse propagation probability components The separation formula is expressed as as well as ; This represents a comparison function that takes the minimum value. Indicates from materials Fluctuations cause material The positive propagation probability component of the fluctuation; This represents the reverse propagation probability component, which has the opposite direction. Subsequently, the system inputs the forward propagation probability component into the asymmetric activation kernel to perform a significance test, and extracts the topological response variable that passes the test. As a characteristic of asymmetric risk transmission, the saliency activation mapping formula is: ; Represents an exponential function with the natural constant as its base; This represents the gradient steepness coefficient of the asymmetric activation kernel; This represents the significance test threshold parameter. The publicly disclosed method involves randomly shuffling and rearranging the original continuous material price data to construct a set of alternative sequences that lack temporal causal relationships. The cross-entropy distribution of this alternative sequence set is then calculated, and its 95th percentile value is taken. The system initializes a zero-value network adjacency matrix in the underlying storage, extracts the entity codes corresponding to multiple materials, maps them to the row and column coordinate system of this matrix to construct a network node matrix, and calculates the qualified asymmetric risk transmission characteristics. The network edge weights are redirected and filled into the corresponding node coordinates. Finally, the system performs a Markov state iterative update calculus along the time axis on the network node matrix filled with network edge weights, with the iterative derivation rule being... ; This indicates the time step index number of the Markov state iteration; Indicates the first The state matrix of the dynamically evolving directed cost-risk association graph sequence generated after each iteration update; The parameter representing the time memory retention rate of Markov states is determined by performing a first-order autoregressive equation to fit a series of historical power grid engineering material procurement prices and extracting its autocorrelation coefficient. Please refer to [link to relevant documentation]. Figure 3 , Figure 3 Each three-dimensional spatial section in the diagram corresponds to a sliding window at a specific observation step size. The dark-colored scattered points floating within the section correspond to the mapped coded nodes of power grid engineering entities (such as copper core cables, tower materials, and steel structures). The spatial lines connecting different nodes in the diagram represent the divergence measure mapping extracted after the system performs joint integral calculations on the edge probability density distribution characteristics. This is because the system performs directional gradient analysis (i.e., separates the forward and reverse propagation probability components). Figure 3 The network connections exhibit significant directionality and variations in thickness. The thickness dimension of the connections precisely maps to the topological response variables. The numerical value represents the network edge weights, reflecting the asymmetric risk transmission characteristics. For example, in the diagram, the radius of the line connecting a core metal node to a secondary auxiliary material node is significantly larger than the reverse line, reflecting the strong unidirectional price pressure exerted by the core metal on auxiliary materials in the macro market. Furthermore, following... Figure 3 Observing the evolution of the sliding window along the horizontal axis, as the Markov state iteration progresses, the thickness and direction of the connection between the same pair of nodes undergo dynamic jumps. This directed topological network, which is constantly reconstructed in both temporal evolution and spatial transmission dimensions, serves as the input of the prior physical structure into the spatial graph convolution kernel of step S4. This enables the Laplace smoothing operator to perform multi-hop message aggregation according to the objectively existing chain reaction path, blocking the noise transmission channels between material nodes without causal relationships.

[0042] Specifically, the process of inputting the directed cost-risk association graph sequence into the spatial graph convolution kernel to generate a topological aggregation path, and inputting the cost exposure offset probability tensor containing lagged variables into the temporal convolution kernel to generate a time-aware sequence is as follows: Extract the Laplace smoothing operator within the spatial graph convolution kernel, and use the Laplace smoothing operator to perform spectral domain mapping on the directed cost-risk association graph sequence to analyze its spatial spectral features; perform multi-hop node neighborhood message passing aggregation on the spatial spectral features along the network edge weights of the directed cost-risk association graph sequence to generate a topological aggregation path; activate the dilated causal convolution channel within the temporal convolution kernel, and input the cost exposure offset probability tensor containing lagged variables into the dilated causal convolution channel for long-range receptive field sequence penetration analysis; extract the hidden state feature vector output from the long-range receptive field sequence penetration analysis, and concatenate the hidden state feature vectors along the time dimension to generate a time-aware sequence.

[0043] In this implementation scheme, the system calculates and generates a regularized Laplace smoothing operator based on the adjacency matrix and node degree matrix of the directed cost-risk association graph sequence, and calibrates it as follows: The calculation formula is: ; Indicates the time segment label of the dynamic graph sequence; Represents a unit diagonal matrix; Indicates the first The adjacency matrix of a directed cost-risk correlation graph sequence under a given time section; Indicates and The corresponding diagonal matrix of node degrees. Using... Perform spectral domain mapping parsing on the input graph network structure, and perform multi-hop node neighborhood message passing aggregation along the network edge weights to generate topology aggregation paths. The recursive formula for multi-order aggregation is: ; The network layer labels for spatial graph convolution; This represents the maximum neighborhood hop count limit; A cursor indicating the hop count of the current message passing; Indicates the first The jump distance attenuation coefficient; Indicates the first The spatial spectral characteristic parameter matrix of the layer; This represents the learnable projection matrix of the spatial convolution layer. This represents the linear rectified activation function. The above maximum neighborhood hop count limit... The determination method involves analyzing the length of the longest dependent node connected branch in the critical procurement process of the power grid engineering bill of materials. Using a dilated causal convolutional network layer within the temporal convolution kernel, a long-range receptive field sequence penetration analysis is performed on the input cost exposure offset probability tensor. This involves receiving the cost exposure offset probability tensor input containing hysteresis variables, performing long-range receptive field sequence penetration analysis within the channel, and obtaining the hidden state feature vector. The calculation rule is as follows ; A step cursor representing the temporal convolution; The sliding observation window size represents the one-dimensional temporal convolution kernel; Represents the cursor of the elements inside the convolution kernel; Indicates the cursor The weighting coefficient of perception at the location; The expansion rate parameter is determined by calculating and rounding the logarithmic function based on the number of days in the seasonal fluctuation cycle of the target material's purchase price. Finally, the hidden state feature vectors of the entire sequence are concatenated along the time dimension to generate a time-aware sequence. .

[0044] Specifically, the process of synchronously inputting the topology aggregation path and the time-aware sequence into the adaptive gated fusion unit for multi-order residual aggregation to generate a spatiotemporal joint vector, and then performing deconvolution decoding mapping on the spatiotemporal joint vector to output the dynamic cost overrun probability sequence at the preset time step is as follows: Initialize the spatial forgetting gate parameter and the time updating gate parameter within the adaptive gated fusion unit; input the topology aggregation path into the spatial forgetting gate parameter to parse the spatial attention weight tensor, and input the time-aware sequence into the time updating gate parameter to parse the time attention weight tensor; use the spatial attention weight tensor and the time attention weight tensor to perform tensor inner product fusion on the topology aggregation path and the time-aware sequence, and perform multi-order residual aggregation through the jump connection operator to generate the spatiotemporal joint vector; input the spatiotemporal joint vector into the deconvolution decoding network for high-dimensional feature projection, mapping it to the joint tensor space of the preset time step and the cost overrun interval, and outputting the dynamic cost overrun probability sequence at the preset time step.

[0045] In this implementation scheme, the spatial forgetting gate parameters are initialized in the underlying computation space of the adaptive gated fusion unit. Update gate parameters with time The final layer topology aggregation path output from the preceding sequence is labeled as... topology aggregation path Input space forget gate parameters The computation channel parses and generates spatial attention weight tensors. The calculation logic is as follows: ; This represents a nonlinear squeezing activation function; This represents the number of bias constants for the spatial forgetting gate. Simultaneously, it incorporates the temporally-aware sequence. Input time to update gate parameters The computation channel parses and generates the temporal attention weight tensor. The calculation logic is as follows: ; This represents the number of constant biases for the time update gate. Tensor inner product fusion is performed on the original input sequence using the two acquired attention weight tensors, and multi-order residual aggregation is performed through skip connection operator bypass to generate a spatiotemporal joint vector. The feature fusion calculation formula is as follows: ; This represents the element-wise inner product multiplication operation at the tensor level. This indicates a concatenation operation along the feature channel dimensions; This represents the cross-layer identity mapping weight matrix for multi-order residual paths. Finally, the spatiotemporal joint vector... The input is fed into a deconvolutional decoding network to perform high-dimensional feature projection calculations, mapping it to the joint tensor space of the preset time step and the cost overrun interval, and outputting a dynamic cost overrun probability sequence at the preset time step. The mapping formula is ; This represents the normalized exponential function; This represents a high-dimensional projected tensor convolution kernel for a deconvolutional network layer; This represents the transpose convolution summation operation; This represents the decoding mapping terminal bias parameters. This deconvolution operation directly decodes the underlying abstract hidden layer encoding into a clear percentage sequence of various cost overrun events that will occur at different construction nodes in the power grid project.

[0046] Example 2; please refer to Figure 4A power grid engineering cost risk prediction system based on time-series data modeling is used to execute a power grid engineering cost risk prediction method based on time-series data modeling described in the embodiments. The system includes: a feature reconstruction module for acquiring continuous material price data and discrete construction node event sequences of the power grid engineering; performing time-frequency transformation on the continuous material price data to separate high-frequency fluctuation component sequences; aligning the discrete construction node event sequences with anchor points using a preset construction baseline as a reference to reconstruct the node time-series deformation feature vector; a tensor generation module for mapping the node time-series deformation feature vector to the boundary penalty coefficient of a dynamic time warping matrix to construct deformation time constraints; and performing time axis distortion and reorganization on the high-frequency fluctuation component sequences through deformation time constraints to generate a cost exposure offset probability tensor containing lag variables; and a graph. The construction module is used to set multi-scale sliding windows on the time axis, extract cost exposure offset probability tensor fragments corresponding to multiple materials, calculate the lead-lag cross-entropy between any two cost exposure offset probability tensor fragments, extract the asymmetric risk transmission features in the lead-lag cross-entropy as network edge weights, and generate a dynamically evolving directed cost risk correlation graph sequence. The risk prediction module is used to input the directed cost risk correlation graph sequence into the spatial graph convolution kernel to generate a topology aggregation path, input the cost exposure offset probability tensor with lag variables into the temporal convolution kernel to generate a time-aware sequence, and synchronously input the topology aggregation path and the time-aware sequence into the adaptive gating fusion unit for multi-order residual aggregation to generate a spatiotemporal joint vector. The spatiotemporal joint vector is deconvolved and decoded to output a dynamic cost overrun probability sequence at a preset time step.

[0047] In this implementation scheme, the feature reconstruction module serves as the underlying heterogeneous data access and signal preprocessing execution unit of the system. This module is equipped with a data stream parsing interface, directly calling read commands to obtain continuous material price data from the commodity trading terminal and discrete construction node event sequences from the engineering construction terminal. On the module's computation side, the time-frequency transformation component performs decomposition and frequency domain filtering commands on the continuous material price data, separating out the high-frequency fluctuation component sequence; the synchronously running alignment component extracts the coordinate scale of the preset construction baseline, forcibly projects the occurrence time points of the discrete construction node event sequences onto this baseline scale, and reconstructs and outputs the node temporal deformation feature vector in the underlying register by calculating the offset between the actual time point and the baseline scale.

[0048] The tensor generation module serves as a spatial mapping component for processing asynchronous misalignment features in the time dimension. This module embeds a dynamic programming algorithm engine, receives the node temporal deformation feature vectors output by the feature reconstruction module, and directly writes their values ​​into the corresponding memory address of the dynamic time warping matrix to generate boundary penalty coefficients. Under the set deformation time constraints, the module's resampling unit initiates interpolation calculations, performing nonlinear compression, stretching, distortion, and recombination on the original time axis of the high-frequency fluctuation component sequence. Subsequently, the module calls the tensor operation library to perform high-dimensional encapsulation of the recombined values, allocates fixed-dimensional blocks in the storage space, and generates and caches a cost exposure offset probability tensor containing hysteresis variables.

[0049] The graph construction module is used to establish an asymmetric logical topology among materials in multi-factor concurrent scenarios. The module's dynamic slicer segments the cached cost exposure offset probability tensor according to the configured multi-scale sliding window step size. The cross-entropy calculation unit within the module performs multi-threaded concurrent computation on the segmented tensor fragments to solve for the probability distribution divergence between different materials. The graph topology construction component performs directional analysis on the divergence data to extract asymmetric risk transmission features, which are then used as network edge weights to fill a pre-defined blank node matrix. As the sliding window advances along the time axis, the module continuously overwrites the graph structure state, cyclically outputting a dynamically evolving sequence of directed cost-risk association graphs in memory.

[0050] The risk prediction module undertakes the final spatiotemporal dual-stream forward propagation and decoding mapping tasks. The module's hardware architecture comprises two independent channels: a spatial graph convolution kernel and a temporal convolution kernel. The spatial graph convolution kernel reads the directed cost-risk association graph sequence, executes node neighborhood message passing, and outputs a topology aggregation path; the temporal convolution kernel simultaneously reads the cost exposure offset probability tensor, executes long-range causal convolution, and outputs a time-aware sequence. The module's internal adaptive gating fusion unit uses preset gate parameters to perform tensor inner product and multi-order residual fusion calculations on the two inputs. The decoder component receives the fused spatiotemporal joint vector, performs inverse dimensionality reduction projection operations, restores the high-dimensional spatial features to business values, and finally outputs a dynamic cost overrun probability sequence at a preset time step.

[0051] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0052] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0056] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the cost risk of power grid projects based on time-series data modeling, characterized in that, Includes the following steps: S1. Obtain continuous material price data and discrete construction node event sequences for power grid projects; perform time-frequency transformation on the continuous material price data to separate high-frequency fluctuation component sequences; use a preset construction baseline as a reference to align the discrete construction node event sequences with anchor points and reconstruct the node time-series deformation feature vector. S2. Map the node temporal deformation feature vector to the boundary penalty coefficient of the dynamic time warping matrix to construct deformation time constraints; use the deformation time constraints to perform time axis distortion and recombination on the high-frequency fluctuation component sequence to generate a cost exposure offset probability tensor with lag variables. S3. Set up a multi-scale sliding window on the time axis, extract cost exposure offset probability tensor segments corresponding to multiple materials, calculate the lead-lag cross-entropy between any two cost exposure offset probability tensor segments, extract the asymmetric risk transmission features in the lead-lag cross-entropy as network edge weights, and generate a dynamically evolving directed cost risk association graph sequence. S4. Input the directed cost risk association graph sequence into the spatial graph convolution kernel to generate a topological aggregation path, and input the cost exposure offset probability tensor with hysteretic variables into the time convolution kernel to generate a time-aware sequence. The topology aggregation path and time-aware sequence are synchronously input into the adaptive gating fusion unit for multi-order residual aggregation to generate a spatiotemporal joint vector. The spatiotemporal joint vector is then deconvolved and decoded to output a dynamic cost overrun probability sequence at a preset time step.

2. The method for predicting the cost risk of power grid projects based on time-series data modeling according to claim 1, characterized in that, The specific process for obtaining continuous material price data and discrete construction node event sequences for power grid projects is as follows: Obtain material price flow data from bulk commodity trading terminals, extract daily settlement price series and bulk transaction volume series from the material price flow data, and concatenate the daily settlement price series and bulk transaction volume series to generate continuous material price data; Acquire engineering log text data and external weather warning text data from the engineering construction management terminal, and extract the work stoppage status label sequence from the engineering log text data and external weather warning text data using natural language entity recognition. The work stoppage status label sequence is discretized and mapped using the timestamp sequence of the engineering log text data and the external meteorological warning text data to generate a discrete construction node event sequence.

3. A method for predicting the cost risk of power grid engineering based on time-series data modeling as described in claim 1 or 2, characterized in that, The specific process of performing time-frequency transformation on continuous material price data to separate high-frequency fluctuation components, and using a preset construction baseline as a reference to anchor and align discrete construction node event sequences to reconstruct the node temporal deformation feature vector is as follows: Empirical mode decomposition is performed on continuous material price data to generate multiple intrinsic mode function component sequences; Instantaneous frequency calculation is performed on multiple intrinsic mode function component sequences to extract the target high-frequency mode function component sequence, and multiple target high-frequency mode function component sequences are fused to generate a high-frequency fluctuation component sequence; Extract the actual state transition time nodes of discrete construction node event sequences, and extract the planned milestone time nodes of the preset construction baseline; A multidimensional phase deviation sequence is generated by measuring the phase deviation between the actual state transition time node and the planned milestone time node, and then the multidimensional phase deviation sequence is mapped into a node time-series deformation feature vector.

4. The method for predicting the cost risk of power grid projects based on time-series data modeling according to claim 1, characterized in that, The specific process of mapping the nodal temporal deformation eigenvectors to the boundary penalty coefficients of the dynamic time warping matrix and constructing the deformation time constraints is as follows: The two-dimensional cost distance grid is initialized based on the time length of the high-frequency fluctuation component sequence and the time length of the preset construction baseline; The nonlinear offset coefficients in the temporal deformation feature vector of the nodes are analyzed, and the nonlinear offset coefficients are input into the asymmetric activation function for nonlinear projection to generate a dynamic penalty weight sequence. The boundary penalty coefficients of the dynamic time warping matrix are constructed by mapping the dynamic penalty weight sequence along the diagonal search path of the two-dimensional cost distance grid to the coordinate system. Deformation time constraints are generated by setting a global search boundary width limit on a dynamic time warp matrix with a fused boundary penalty coefficient.

5. A method for predicting the cost risk of power grid engineering based on time-series data modeling as described in claim 1 or 4, characterized in that, The specific process of generating a cost exposure offset probability tensor with lag variables by time-axis distortion and reconstruction of the high-frequency fluctuation component sequence using deformation time constraints is as follows: In a dynamic time warp matrix with a bounded penalty coefficient, dynamic programming is used to find the path with the minimum cumulative distance by following the global search boundary width limit defined by the deformation time constraint. The minimum cumulative distance path is used as the time remapping function to perform nonlinear interpolation resampling on the high-frequency fluctuation component sequence to generate the deformation high-frequency fluctuation component sequence. Gaussian perturbation operator is introduced at each time step of the high-frequency fluctuation component sequence of deformation to perform Monte Carlo probability sampling and generate a multidimensional probability distribution matrix. Multiple multidimensional probability distribution matrices are stacked in three dimensions according to the material category dimension of power grid engineering to generate a cost exposure offset probability tensor with hysteresis variables.

6. The method for predicting the cost risk of power grid projects based on time-series data modeling according to claim 1, characterized in that, The specific process of setting up a multi-scale sliding window on the time axis, extracting cost exposure offset probability tensor segments corresponding to multiple materials, and calculating the lead-lag cross-entropy between any two cost exposure offset probability tensor segments is as follows: Configure multi-band observation step size to generate multi-scale sliding window, and use the multi-scale sliding window to dynamically slice the cost exposure offset probability tensor with lagged variables on the time axis, and extract cost exposure offset probability tensor fragments corresponding to multiple materials. A multi-level time delay operator is introduced between any two cost exposure offset probability tensor fragments for misalignment alignment, and the edge probability density distribution characteristics under the misalignment alignment state are analyzed. The divergence measure variable is extracted by joint distribution integral calculus of the marginal probability density distribution characteristics, and the divergence measure variable is mapped to the lead-lag cross-entropy between any two cost exposure offset probability tensor segments.

7. A method for predicting the cost risk of power grid engineering based on time-series data modeling as described in claim 1 or 6, characterized in that, The specific process of extracting the asymmetric risk transmission features from the lead-lag cross-entropy as network edge weights to generate a dynamically evolving sequence of directed cost-risk association graphs is as follows: Directional gradient analysis is performed on the lead-lag cross-entropy to separate the forward and backward propagation probability components in the lead-lag cross-entropy. The forward and backward propagation probability components are input into an asymmetric activation kernel for significance testing, and the topological response variables that pass the significance test are extracted as asymmetric risk propagation features. Extract the entity codes corresponding to multiple materials as the network node matrix, and map the asymmetric risk transmission features as network edge weights to fill the network node matrix; Markov state iteration updates are performed on the network node matrix filled with network edge weights along the time axis to generate a dynamically evolving sequence of directed cost-risk association graphs.

8. The method for predicting the cost risk of power grid engineering based on time-series data modeling according to claim 1, characterized in that, The specific process of inputting the directed cost risk correlation graph sequence into the spatial graph convolution kernel to generate a topological aggregation path, and inputting the cost exposure offset probability tensor containing hysteresis variables into the time convolution kernel to generate a time-aware sequence is as follows: Extract the Laplacian smoothing operator within the convolution kernel of the spatial graph, and use the Laplacian smoothing operator to perform spectral domain mapping on the directed cost-risk association graph sequence to analyze its spatial spectral features; The network edge weights of the directed cost-risk association graph sequence are used to aggregate the spatial spectrum features through multi-hop node neighborhood message passing to generate a topological aggregation path. Activate the dilated causal convolution channel within the temporal convolution kernel, and input the cost exposure offset probability tensor containing hysteresis variables into the dilated causal convolution channel for long-range receptive field sequence penetration analysis; Extract the latent state feature vector from the long-range receptive field sequence penetration analysis output, and concatenate the latent state feature vector along the time dimension to generate a time-aware sequence.

9. The method for predicting the cost risk of power grid projects based on time-series data modeling according to claim 1, characterized in that, The specific process of synchronously inputting the topology aggregation path and time-aware sequence into the adaptive gated fusion unit for multi-order residual aggregation to generate a spatiotemporal joint vector, and then performing deconvolution decoding and mapping on the spatiotemporal joint vector to output the dynamic cost overrun probability sequence at a preset time step is as follows: Initialize the spatial forgetting gate parameters and the temporal update gate parameters within the adaptive gating fusion unit; The topological aggregation path is input into the spatial forgetting gate parameter to parse the spatial attention weight tensor, and the time-aware sequence is input into the time update gate parameter to parse the time attention weight tensor. The spatial attention weight tensor and the temporal attention weight tensor are used to perform tensor inner product fusion of topological aggregation path and time-aware sequence, and multi-order residual aggregation is performed through jump connection operator to generate spatiotemporal joint vector. The spatiotemporal joint vector is input into the deconvolutional decoding network for high-dimensional feature projection, which maps it to the joint tensor space of the preset time step and the cost overrun interval, and outputs the dynamic cost overrun probability sequence at the preset time step.

10. A power grid engineering cost risk prediction system based on time-series data modeling, used to execute the power grid engineering cost risk prediction method based on time-series data modeling as described in any one of claims 1-9, characterized in that, include: The feature reconstruction module is used to obtain continuous material price data and discrete construction node event sequences for power grid projects. The high-frequency fluctuation component sequence is separated by time-frequency transformation of continuous material price data, and the discrete construction node event sequence is anchored and aligned with the preset construction baseline as a reference to reconstruct the node time-series deformation feature vector. The tensor generation module is used to map the node temporal deformation feature vectors to the boundary penalty coefficients of the dynamic time warping matrix and construct deformation time constraints. Through deformation time constraints, the high-frequency fluctuation component sequence is time-axis distorted and reorganized to generate a cost exposure offset probability tensor containing lag variables. The graph construction module is used to set multi-scale sliding windows on the time axis, extract cost exposure offset probability tensor segments corresponding to multiple materials, calculate the lead-lag cross-entropy between any two cost exposure offset probability tensor segments, extract the asymmetric risk transmission features in the lead-lag cross-entropy as network edge weights, and generate a dynamically evolving directed cost risk association graph sequence. The risk prediction module is used to input the directed cost risk correlation graph sequence into the spatial graph convolution kernel to generate a topological aggregation path, and to input the cost exposure offset probability tensor with lag variables into the time convolution kernel to generate a time-aware sequence. The topology aggregation path and time-aware sequence are synchronously input into the adaptive gating fusion unit for multi-order residual aggregation to generate a spatiotemporal joint vector. The spatiotemporal joint vector is then deconvolved and decoded to output a dynamic cost overrun probability sequence at a preset time step.

Images