Construction engineering cost prediction method based on deep learning

Abstract

The application discloses a building engineering cost prediction method based on deep learning and belongs to the technical field of building engineering cost prediction. The method comprises the following steps: constructing a cost time sequence knowledge graph of a target building engineering, wherein the graph comprises historical cost data of multiple engineering nodes, material price fluctuation sequences and construction progress mapping relationships; performing spatiotemporal feature aggregation processing on the cost time sequence knowledge graph through a deep learning encoder to generate a cost dynamic feature tensor; calling a pre-trained engineering cost prediction network to perform multitask decoding processing on the cost dynamic feature tensor to generate a basic cost prediction interval and a key cost-sensitive node distribution; performing market fluctuation elasticity correction processing on the basic cost prediction interval to generate a corrected cost prediction interval; and generating a cost control strategy scheme according to the key cost-sensitive node distribution, wherein the scheme comprises material procurement time sequence optimization parameters and fund allocation rhythm adjustment parameters.

Description

Technical Field

[0001] This invention relates to the field of construction project cost prediction technology, specifically a construction project cost prediction method based on deep learning. Background Technology

[0002] Construction project cost forecasting is a core component of project investment decision-making and cost control. Traditional cost forecasting methods rely on quota pricing and static coefficient adjustments. After decomposing the project into sub-items, they obtain an overall cost estimate based on historical experience or recent data from similar projects through linear regression, exponential smoothing, or simple time series extrapolation. This method treats each project node as an isolated object, failing to integrate the dynamic coupling relationship between construction progress and material market fluctuations, resulting in forecasts lagging behind market changes. Existing data-driven cost forecasting technologies attempt to introduce machine learning models, using macroeconomic indicators such as material price indices and labor rates as input features, and outputting a single cost forecast value through fully connected networks or shallow time series models. These methods face two main drawbacks. First, the feature construction process isolates the project structure topology, construction sequence, and the time-series evolution of costs at each node, failing to capture the spatial dependencies between project nodes and their temporal linkage effects, and omitting information on cost transfer and accumulation between sub-items due to factors such as schedule coordination and resource allocation. On the other hand, existing methods only output deterministic cost values ​​or fixed-width confidence intervals, failing to differentiate the volatility sensitivity of different cost components or identify the key nodes with the greatest impact on the overall cost. This makes it impossible for the prediction intervals to reflect the asymmetric inflation effect caused by high-volatility and low-volatility categories of market materials. The resulting cost control recommendations are difficult to deeply bind to the cost structure and procurement sequence of specific project nodes, lacking a quantitative basis for adjusting material procurement lead times and fund disbursement schedules.

[0003] One of the core issues to be addressed in cost forecasting is how to integrate the spatial neighborhood relationships of project nodes, the time series patterns of pricing deviations, and the external shocks of material price fluctuations into a unified feature representation space, forming a representation that reflects the cost evolution mechanism throughout the entire construction process. Another pressing issue is how, after obtaining the initial cost range, to make targeted adjustments based on the differences in market volatility intensity for different cost sub-items such as materials, labor, and machinery, and, based on the automatic identification of cost fluctuation-sensitive nodes, to derive quantifiable procurement timing adjustment parameters and fund disbursement parameters, thus achieving a closed loop from forecasting to control strategies. Summary of the Invention

[0004] This paper proposes a deep learning-based method for predicting construction project costs to address the problems in existing technologies, such as fragmented cost prediction features leading to a lack of spatiotemporal linkage information, inability of prediction intervals to adapt to market fluctuations, and difficulty in generating node-level control strategies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a deep learning-based method for predicting construction project costs. This method includes: constructing a time-series knowledge graph of the cost of a target construction project, which contains historical cost data for multiple project nodes, material price fluctuation sequences, and construction progress mapping relationships; performing spatiotemporal feature aggregation processing on the time-series knowledge graph using a deep learning encoder to generate a dynamic cost feature tensor for the target construction project; calling a pre-trained construction cost prediction network to perform multi-task decoding processing on the dynamic cost feature tensor to generate a basic cost prediction interval and a distribution of key cost-sensitive nodes for the target construction project; performing market fluctuation elasticity correction processing on the basic cost prediction interval to generate a corrected cost prediction interval; and generating a cost control strategy based on the distribution of key cost-sensitive nodes, which includes material procurement timing optimization parameters and fund disbursement rhythm adjustment parameters. In this way, the structured cost information, market fluctuation information and construction progress information are fully integrated by using a cost time series knowledge graph. Through end-to-end deep spatiotemporal feature learning, the basic cost prediction range and cost-sensitive node distribution that can reflect the coupling effect of multiple factors are obtained. After market fluctuation elasticity correction, cost prediction results and dynamic control schemes that are more in line with the real-time market environment and have strategic guidance significance are given.

[0006] Preferably, the construction of the cost time-series knowledge graph of the target construction project includes: obtaining a set of sub-item project nodes of the target construction project, each sub-item project node corresponding to a cost accounting unit, the cost accounting unit containing the original pricing data of labor cost sub-items, material cost sub-items and machinery cost sub-items; extracting the pricing deviation sequence of each cost accounting unit at multiple historical time nodes, the pricing deviation sequence being the difference sequence between the actual pricing data and the quota benchmark data; establishing a mapping association edge between the pricing deviation sequence and the corresponding material market price based on the material category identifier in the material price fluctuation sequence, the mapping association edge having timestamp attributes and regional attributes; integrating the comparison deviation between the planned progress node and the actual progress node in the construction progress mapping relationship, adjusting the weight of the mapping association edge for time-series alignment, and generating a cost time-series knowledge graph containing node features, association edges and time attributes. Therefore, the coupling relationship between the actual cost deviation of the project, the fluctuation of the material market, and the deviation of the construction schedule is explicitly expressed as a knowledge graph, so that the subsequent spatiotemporal feature aggregation can more accurately capture the transmission path and key influencing factors of cost fluctuations.

[0007] As a preferred technical solution of the present invention, a deep learning encoder is used to perform spatiotemporal feature aggregation processing on the cost time-series knowledge graph and generate a cost dynamic feature tensor. Specifically, this includes: using a graph attention encoding layer to perform spatial neighborhood feature aggregation on each engineering node of the cost time-series knowledge graph, wherein the graph attention encoding layer calculates the attention allocation weight of the set of its neighboring nodes for each engineering node, and performs weighted summation on the feature vectors of the neighboring nodes according to the attention allocation weight to generate the spatial aggregation feature of each engineering node; using a temporal convolutional encoding layer to perform multi-scale sliding window extraction processing on the temporal features of each engineering node, wherein the temporal convolutional encoding layer contains multiple dilated convolution kernels, each dilated convolution kernel capturing the short-term fluctuation pattern and long-term trend pattern of the pricing deviation sequence with different receptive fields to generate the temporal evolution feature of each engineering node; performing cross-fusion processing on the feature dimension of the spatial aggregation feature and the temporal evolution feature to obtain the spatiotemporal fusion vector of each engineering node; and performing tensor stacking processing on all the spatiotemporal fusion vectors of the engineering nodes according to the topological order of the cost time-series knowledge graph to generate the cost dynamic feature tensor. When performing tensor stacking, the topological order is determined according to the execution order of each project node in the construction progress sequence. Project nodes at the same construction level are arranged according to the numbering order of the sub-projects to preserve the structural logic in the project progress process.

[0008] Furthermore, when the graph attention encoding layer calculates the attention allocation weights for the set of neighboring nodes for each engineering node, these weights are obtained by normalizing the concatenated vector of the neighboring node feature vectors and the current node feature vector after calculation through a single-layer feedforward network. The graph attention encoding layer also employs a multi-head attention mechanism to concatenate or average the outputs of multiple attention heads, enhancing the robustness and richness of the feature representation. Through these processes, the generated cost dynamic feature tensor not only characterizes the cost evolution patterns of each engineering node but also incorporates the mutual influence between upstream and downstream nodes and multi-scale market fluctuations during construction, laying a feature foundation for subsequent predictions.

[0009] The method of calling the pre-trained engineering cost prediction network to perform multi-task decoding processing on the cost dynamic feature tensor preferably includes: inputting the cost dynamic feature tensor into the shared feature extraction backbone of the engineering cost prediction network, which uses a multi-level residual contraction structure to perform noise suppression and key feature enhancement processing on the cost dynamic feature tensor to generate a compressed feature vector; inputting the compressed feature vector branch into a first task decoding head, which generates the basic cost prediction interval based on a quantile regression mechanism, the basic cost prediction interval containing cost prediction values ​​corresponding to multiple preset quantiles; simultaneously, inputting the compressed feature vector branch into a second task decoding head, which calculates the contribution sensitivity coefficient of each engineering node to the basic cost prediction interval through layer-by-layer deconvolution operations, and filters out engineering nodes that exceed a preset sensitivity threshold from all engineering nodes based on the contribution sensitivity coefficient, and marks the spatial location corresponding to the filtered engineering nodes as the key cost sensitive node distribution. In this process, the second task decoding head reconstructs the compressed feature vector into a reconstructed feature map with the same dimensions as the cost dynamic feature tensor space through deconvolution. Then, it normalizes the feature response value of each engineering node on this reconstructed feature map and uses it as the contribution sensitivity coefficient for that engineering node. Through this multi-task decoding architecture, it can output a basic cost prediction range with confidence information and locate engineering parts that are highly sensitive to cost fluctuations, providing a basis for differentiated cost control.

[0010] The process of applying market volatility elasticity correction to the basic cost forecast interval and generating a corrected cost forecast interval preferably includes: obtaining the price fluctuation amplitude and frequency of each material category in the material price fluctuation sequence within the current forecast window; calculating the price fluctuation intensity index of each material category based on the price fluctuation amplitude and frequency; elastically stretching the corresponding material cost sub-items in the basic cost forecast interval based on the price fluctuation intensity index of each material category to generate a material cost elasticity correction sub-interval; obtaining the change rate of the labor cost adjustment index and the machinery cost adjustment index within the current forecast window; linearly adjusting the labor cost sub-item in the basic cost forecast interval based on the change rate of the labor cost adjustment index; linearly adjusting the machinery cost sub-item in the basic cost forecast interval based on the change rate of the machinery cost adjustment index to generate a labor cost correction sub-interval and a machinery cost correction sub-interval; and superimposing and merging the material cost elasticity correction sub-interval, the labor cost correction sub-interval, and the machinery cost correction sub-interval to generate the corrected cost forecast interval. The elastic stretching process further includes: dividing each material category into a high-volatility category set and a low-volatility category set based on the price volatility intensity index of each material category within the current prediction window; the price volatility intensity index of the material categories in the high-volatility category set is greater than a preset volatility threshold, while the price volatility intensity index of the material categories in the low-volatility category set is less than or equal to the preset volatility threshold; for each material category in the high-volatility category set, multiplying the interval width of the corresponding material cost sub-interval in the basic cost prediction interval by a first stretching coefficient to generate a high-volatility material elastic sub-interval, where the first stretching coefficient is positively correlated with the price volatility intensity index; for each material category in the low-volatility category set, multiplying the interval width of the corresponding material cost sub-interval in the basic cost prediction interval by a second stretching coefficient to generate a low-volatility material elastic sub-interval, where the second stretching coefficient is a fixed constant and less than the minimum value of the first stretching coefficient; and weighting all high-volatility material elastic sub-intervals and all low-volatility material elastic sub-intervals according to the proportion of each material category in the cost composition to obtain the material cost elastic correction sub-interval. This approach fully considers the differentiated impact of market fluctuations in different material categories, allowing the revised cost forecast range to expand reasonably when encountering sharp fluctuations in material prices, while narrowing when the market is stable, thereby improving the realism and practicality of the forecast range.

[0011] Based on the distribution of key cost-sensitive nodes, a cost control strategy is generated, preferably including: extracting the sensitivity coefficient and node type identifier of each sensitive node in the key cost-sensitive node distribution, wherein the node type identifier includes labor-intensive nodes, material-intensive nodes, and machinery-intensive nodes; for labor-intensive nodes whose sensitivity coefficient is higher than a preset labor sensitivity threshold, generating material procurement timing optimization parameters, wherein the material procurement timing optimization parameters include the procurement lead time adjustment value and the batch quantity of each material category; for material-intensive nodes whose sensitivity coefficient is higher than a preset material sensitivity threshold, generating fund disbursement rhythm adjustment parameters, wherein the fund disbursement rhythm adjustment parameters include the fund disbursement ratio adjustment value and disbursement time offset for each stage; and coordinating and aligning the material procurement timing optimization parameters and the fund disbursement rhythm adjustment parameters according to the construction progress sequence to generate the cost control strategy. This allows for precise procurement plans and fund allocation suggestions tailored to the cost structure characteristics and sensitivity levels of different nodes, reducing the risk of cost overruns from the source.

[0012] The generation of material procurement timing optimization parameters for labor-intensive nodes further includes: obtaining the construction schedule of the labor-intensive nodes and extracting the demand time nodes and quantities for each material category; calculating the procurement delay tolerance for each material category based on the historical price peak distribution of the corresponding materials in the demand time nodes and price fluctuation sequences, where the procurement delay tolerance is the number of days that can be offset between the demand time node and the historical price trough; determining the procurement lead time adjustment value for each material category based on the procurement delay tolerance, i.e., the number of days to extend or shorten the original procurement lead time; determining the batch quantity for batch procurement based on the ratio of the demand quantity for each material category to the upper limit of single procurement capacity, rounded up; and associating and packaging the procurement lead time adjustment value and batch quantity for each material category by material category to generate the material procurement timing optimization parameters. By utilizing the matching of historical price trough distribution with construction demand time points, material procurement costs can be effectively reduced while ensuring construction progress.

[0013] The generation of adjustment parameters for the funding disbursement rhythm of material-intensive nodes further includes: extracting the funding plan table for material-intensive nodes, obtaining the planned disbursement amount and planned disbursement time node for each construction stage from the funding plan table; calculating the funding demand elasticity coefficient for each construction stage based on the ratio of material cost to labor cost in each construction stage, wherein the funding demand elasticity coefficient is positively correlated with the material cost ratio; calculating the fund disbursement ratio adjustment value for each construction stage based on the fund demand elasticity coefficient for each construction stage, wherein the adjustment ratio adjustment value is the difference between the planned disbursement amount multiplied by the elasticity adjustment factor and the original planned disbursement amount; calculating the disbursement time offset for each construction stage based on the material supply market settlement cycle for each construction stage, i.e., the number of calendar days that the planned disbursement time node is advanced or delayed; and arranging and combining the fund disbursement ratio adjustment value and the disbursement time offset for each construction stage according to the construction stage sequence to generate the fund disbursement rhythm adjustment parameters. This method coordinates the fund disbursement rhythm with the material market price cycle and the actual construction progress, smoothing cash flow pressure and avoiding the risk of settlement at high prices while meeting funding needs.

[0014] For the joint training process of the deep learning encoder and the engineering cost prediction network, this invention provides the following scheme: Obtain historical cost time-series knowledge graph samples of multiple historical building projects and actual final settlement cost data of each historical building project; use each historical cost time-series knowledge graph sample as training input data and the corresponding actual final settlement cost data as training label data; sequentially pass the training input data through the deep learning encoder and the engineering cost prediction network for forward propagation calculation to obtain the training prediction cost intervals corresponding to each historical building project; calculate the composite loss function value between each training prediction cost interval and the corresponding training label data, where the composite loss function value includes a weighted sum of the interval coverage width criterion loss component and the interval average width loss component; update the first network weight parameters of the deep learning encoder and the second network weight parameters of the engineering cost prediction network according to the composite loss function value using a backpropagation algorithm until the composite loss function value converges to below a preset threshold, thus obtaining the trained deep learning encoder and engineering cost prediction network. This training method balances the reasonable coverage of the prediction interval to the actual cost with the efficient narrowing of the interval itself, enabling the model to achieve an optimal balance between prediction reliability and accuracy, thereby providing high-quality parametric prediction capabilities for subsequent applications.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0016] A cost time-series knowledge graph is adopted as the data organization form. The historical cost data, pricing deviation sequences, material price fluctuation sequences, and construction progress mapping relationships of sub-item project nodes are uniformly expressed by node features, mapping association edges with timestamps and regional attributes, and edge weights adjusted for construction progress comparison deviations. Based on this knowledge graph, the graph attention encoding layer in the deep learning encoder performs spatial neighborhood feature aggregation on each project node. The multi-head attention mechanism is used to assign weights to adjacent nodes to generate spatial aggregation features. At the same time, the temporal convolutional encoding layer captures the short-term fluctuation patterns and long-term trend patterns of the pricing deviation sequence with multiple convolution kernels with different dilation coefficients to generate temporal evolution features. The two are then cross-fused in the feature dimension to form a spatiotemporal fusion vector, which is stacked into a cost dynamic feature tensor. This processing method enables the model to simultaneously learn the spatial dependencies between engineering nodes, the multi-scale time-varying patterns of pricing data for each node, and the corrective effect of construction progress deviations on the correlation strength. The cost dynamic feature tensor can characterize the linkage and transmission mechanism of cost changes throughout the entire construction process, overcoming the information isolation caused by traditional methods that extract features from only a single node or a single time segment. A pre-trained engineering cost prediction network is invoked to perform multi-task decoding on the cost dynamic feature tensor. The first task decoding head generates a basic cost prediction interval containing multiple preset quantile prediction values ​​based on a quantile regression mechanism. The second task decoding head calculates the sensitivity coefficient of each engineering node's contribution to the basic cost interval through layer-by-layer deconvolution operations, filtering out the distribution of key cost-sensitive nodes. Based on this, the price fluctuation amplitude and frequency of each material category in the material price fluctuation sequence within the current forecast window are obtained. A price fluctuation intensity index is calculated, and the material categories are divided into a high-volatility category set and a low-volatility category set. The width of the material cost sub-interval corresponding to the high-volatility category is multiplied by a first stretching coefficient positively correlated with the fluctuation intensity, while the width of the low-volatility category is multiplied by a smaller fixed second stretching coefficient, generating a material cost elasticity correction sub-interval. Simultaneously, the corresponding sub-items are linearly adjusted according to the change rates of the labor cost adjustment index and the machinery cost adjustment index, respectively. The results are then merged to obtain the corrected cost forecast interval. This correction mechanism allows the coverage width of the forecast interval to adapt to the asymmetry of market fluctuations with different cost compositions. The forecast uncertainty of high-volatility material categories is directionally expanded within the interval, while the low-volatility portion remains compact. The corrected cost forecast interval is more sensitive to market price anomalies.

[0017] Based on the identification results of the distribution of key cost-sensitive nodes, nodes are categorized into labor-intensive, material-intensive, and machinery-intensive nodes. For labor-intensive high-sensitivity nodes, the procurement delay tolerance for each material category is calculated, and the adjustment value for procurement lead time and the number of batches for phased procurement are determined, generating material procurement timing optimization parameters. For material-intensive high-sensitivity nodes, the adjustment value for the proportion of fund disbursement at each construction stage and the disbursement time offset are calculated based on the capital demand elasticity coefficient, generating fund disbursement rhythm adjustment parameters. These two types of parameters are then aligned in a coordinated manner according to the construction progress sequence to form a cost control strategy. This scheme directly originates from the node-level sensitivity information output by the prediction model, quantitatively linking cost prediction with proactive control measures. The parameterized adjustment of material procurement rhythm and fund disbursement rhythm can effectively suppress the transmission of cost fluctuations at key nodes, enhancing the operability of cost management. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of a deep learning-based method for predicting construction project costs.

[0020] Figure 2 This is a flowchart of the construction of a cost time-series knowledge graph and the generation of cost dynamic feature tensors;

[0021] Figure 3 This is a schematic diagram of the structure of an engineering cost prediction network;

[0022] Figure 4 This is a flowchart of the joint training process between a deep learning encoder and an engineering cost prediction network;

[0023] Figure 5 This is a schematic diagram of the feature response of the multi-scale dilated convolution kernel in the temporal convolutional coding layer;

[0024] Figure 6 It is the convergence curve of the composite loss function during the training process. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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, 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.

[0026] See Figure 1 This invention provides a deep learning-based method for predicting construction project costs, comprising: constructing a time-series knowledge graph of the cost of a target construction project, the time-series knowledge graph containing historical cost data of multiple project nodes, material price fluctuation sequences, and construction progress mapping relationships; performing spatiotemporal feature aggregation processing on the time-series knowledge graph using a deep learning encoder to generate a dynamic cost feature tensor of the target construction project; calling a pre-trained construction cost prediction network to perform multi-task decoding processing on the dynamic cost feature tensor to generate a basic cost prediction interval and a distribution of key cost-sensitive nodes for the target construction project; performing market fluctuation elasticity correction processing on the basic cost prediction interval to generate a corrected cost prediction interval; and generating a cost control strategy scheme based on the distribution of key cost-sensitive nodes, the cost control strategy scheme including material procurement time-series optimization parameters and fund disbursement rhythm adjustment parameters.

[0027] Example 1:

[0028] In specific implementation, please refer to Figure 2This embodiment provides a process for constructing a time-series knowledge graph of construction costs and generating dynamic cost feature tensors in a deep learning-based method for predicting construction project costs. When constructing the time-series knowledge graph of the target construction project, a set of sub-item project nodes is obtained. Each sub-item project node corresponds to a cost accounting unit, which includes the original pricing data for labor costs, material costs, and machinery costs. From each cost accounting unit, a pricing deviation sequence at multiple historical time points is extracted. The pricing deviation sequence is the difference sequence between the actual pricing data and the quota benchmark data. Then, based on the material category identifier in the material price fluctuation sequence, a mapping relationship is established between the pricing deviation sequence and the corresponding material market price. Each mapping relationship has a timestamp attribute and a geographic attribute. The timestamp attribute records the collection time of the market price data, and the geographic attribute records the source region of the market price data. After establishing the mapping relationships, the weights of the mapping relationships are adjusted for temporal alignment by integrating the discrepancies between planned and actual progress nodes in the construction progress mapping relationship, generating a cost time-series knowledge graph containing node features, associated edges, and time attributes. The temporal alignment adjustment works as follows: when the discrepancy between the planned and actual progress nodes indicates a delay or advancement in the construction period for a certain project node, the weight values ​​of the mapping relationships associated with that project node are adjusted accordingly. The magnitude of the weight adjustment is positively correlated with the absolute value of the discrepancy, ensuring that associated edges with discrepancies in the time dimension are represented differently in subsequent processing.

[0029] After obtaining the cost time-series knowledge graph, a deep learning encoder is used to perform spatiotemporal feature aggregation on the cost time-series knowledge graph to generate a dynamic feature tensor of the target building project's cost. The deep learning encoder consists of two core modules: a graph attention coding layer and a temporal convolutional coding layer, as well as a feature cross-fusion module and a tensor stacking module located after them.

[0030] The graph attention encoding layer aggregates spatial neighborhood features for each engineering node in the cost time-series knowledge graph. For each engineering node, the graph attention encoding layer calculates the attention assignment weights for its neighboring node set. These weights are obtained by concatenating the feature vectors of the neighboring nodes and the current node's feature vector, processing them through a single-layer feedforward network, and then normalizing them using the softmax function. Taking node i as an example, its neighboring node set is denoted as... The feature vector of the current node i is The eigenvector of neighboring node j is Attention weights are assigned between node i and its neighboring node j. The calculation formula is:

[0031]

[0032] in, The linear transformation matrix is ​​shared by all nodes and has dimensions of . , The dimension of the input feature vector. The transformed feature dimensions; This represents a vector concatenation operation; Here are the learnable weight vectors of a single-layer feedforward network, with dimension 1. ; The activation function is a linear rectifier with leakage, and the negative slope is set to 0.2. The function is an exponential function; the denominator is the exponential summation of the original attention coefficients corresponding to all neighboring nodes k of node i. The graph attention encoding layer employs a multi-head attention mechanism, independently performing the above attention calculations. Second-rate, The value is 4, and each calculation represents one attention head. Each attention head uses an independent linear transformation matrix. and weight vector , After obtaining the attention assignment weights of node i to its neighboring node j from the output of each attention head, the feature vectors of the neighboring nodes are weighted and summed according to the attention assignment weights to generate the spatial aggregated feature vector of node i under that attention head. . Indicates the first The spatial aggregated feature vector of node i under each attention head has a dimension of d'; For the first The attention weights assigned to node i by each attention head to its neighboring node j, with values ​​ranging from 0 to 1, and the sum of the attention weights assigned to all neighboring nodes of node i is 1; For the first The linear transformation matrix used by each attention head has a dimension of . ; Let be the feature vector of the adjacent node j, with dimension d; This indicates that the sum of all adjacent nodes j of node i is calculated. Then... The spatial aggregation feature vectors obtained from each attention head are concatenated to form the final spatial aggregation feature of node i. The dimension of the final spatial aggregation feature is .

[0033] The temporal convolutional coding layer performs multi-scale sliding window extraction on the temporal features of each project node. The temporal convolutional coding layer contains multiple parallel dilated convolutional kernel branches, with three dilated convolutional kernels of size 3 and dilation rates of 1, 2, and 4, respectively. For the pricing deviation sequence corresponding to the project node, it is arranged in chronological order as a one-dimensional temporal signal and input into the three dilated convolutional kernel branches. The convolutional kernel with a dilation rate of 1 captures short-term fluctuation patterns in the pricing deviation sequence using receptive fields that are adjacent to each other; the convolutional kernel with a dilation rate of 2 captures medium-term trends using receptive fields spaced one time step apart; and the convolutional kernel with a dilation rate of 4 captures long-term trend patterns using receptive fields spaced three time steps apart. Each dilated convolutional kernel branch outputs a feature sequence of the same length as the input sequence, with zero-padding to maintain the temporal dimension. The output feature sequences of the three dilated convolutional kernel branches are concatenated along the channel dimension to generate the temporal evolution feature vector of the project node, the dimension of which is the sum of the number of output channels of the three branches.

[0034] After obtaining the spatial aggregation features and temporal evolution features of each engineering node, the spatial aggregation features and temporal evolution features are cross-fused along the feature dimension. The cross-fusion process is as follows: for each engineering node, the spatial aggregation feature vector and the temporal evolution feature vector are added element-wise. Before addition, the two feature vectors are scaled separately using a learnable scaling factor, initially set to 0.5, which is automatically optimized and updated during the joint training of the deep learning encoder. The result after addition is the spatiotemporal fusion vector of that engineering node.

[0035] The spatiotemporal fusion vectors of all project nodes are stacked as tensors according to the topological order of the cost time-series knowledge graph to generate a cost dynamic feature tensor. The topological order is determined by the execution order of each project node in the construction progress timeline, and project nodes at the same construction level are arranged according to the numbering order of the sub-items. Specifically, the tensor stacking method is as follows: the spatiotemporal fusion vectors of each project node are used as row vectors and arranged from top to bottom according to the topological order to form a two-dimensional tensor. The number of rows in the tensor is the total number of project nodes, and the number of columns is the dimension of the spatiotemporal fusion vector. This two-dimensional tensor is the final output cost dynamic feature tensor, which contains the fusion feature expressions of all project nodes in both spatial dependency and temporal evolution dimensions.

[0036] See Figure 5The figure shows the feature response curves corresponding to three different dilation rates of convolutional kernels extracted by the temporal convolutional coding layer in Example 1. The horizontal axis represents the time step, ranging from 0 to 500, and the vertical axis represents the feature response value, reflecting the feature intensity at the corresponding time step. Dilation rate 1, as shown in the figure, represents a short-term fluctuation pattern. The curve is marked with solid dots, exhibiting frequent and large-amplitude oscillations with obvious periodicity, capturing the subtle short-term fluctuations in the pricing deviation sequence. Dilation rate 2 represents a medium-term trend. The curve is marked with dashed squares, showing relatively gentle changes and a longer period, with fluctuation amplitudes between the short and long term, reflecting the medium-term pattern of the pricing deviation sequence. Dilation rate 4 represents a long-term trend. The curve is marked with a diamond-shaped dashed line, exhibiting slower and smoother fluctuations compared to the previous two, reflecting the long-term evolution trend of material prices and cost deviations. The three curves maintain synchronous changes over time, with the short-term fluctuation curve frequently interspersed above the medium-term and long-term trend curves, demonstrating the hierarchical structure of multi-scale temporal features.

[0037] Example 2:

[0038] In specific implementation, please refer to Figure 3 The system takes the cost dynamic feature tensor generated by the deep learning encoder as input and calls the pre-trained engineering cost prediction network to perform multi-task decoding processing. The engineering cost prediction network consists of three parts: a shared feature extraction backbone, a first-task decoding head, and a second-task decoding head. The shared feature extraction backbone receives the cost dynamic feature tensor and outputs a compressed feature vector. The compressed feature vector is then distributed to the first-task decoding head and the second-task decoding head to generate the basic cost prediction interval and the distribution of key cost-sensitive nodes, respectively.

[0039] The shared feature extraction backbone employs a multi-level residual contraction structure to perform noise suppression and key feature enhancement on the cost dynamic feature tensor. The cost dynamic feature tensor is first input into a two-dimensional convolutional layer with a kernel size of 3×3, a stride of 1, padding of 1, and 64 output channels, generating an initial feature map. This initial feature map is then fed into three cascaded residual contraction modules. Each residual contraction module contains a residual branch and a soft-thresholding branch. The residual branch consists of two cascaded convolutional layers: the first convolutional layer has a kernel size of 3×3, a stride of 1, padding of 1, and 64 output channels, followed by batch normalization and a linear rectified activation function; the second convolutional layer also has a kernel size of 3×3, a stride of 1, padding of 1, and 64 output channels, followed by batch normalization but without an activation function. The soft thresholding branch takes the absolute value of the feature map output by the residual branch and then performs global average pooling to obtain a 64-dimensional vector. This vector is passed through a fully connected layer to output a 64-dimensional vector, which is then passed through a sigmoid activation function to obtain a scaling factor vector in the range (0,1). The scaling factor vector is multiplied by the mean of the absolute value vector after global average pooling to obtain the soft thresholding parameter vector. , The dimension is 64, where the first... element Indicates channel The threshold, Specifically, it is calculated from the global average pooling result and the learned scaling factor. Then, for each channel of the feature map output by the residual branch... Noise suppression is achieved by applying a soft threshold function. The formula for calculating the soft threshold is as follows:

[0040]

[0041] in, Indicates the position in the residual branch output feature map ,aisle The original feature response value at that location, Here, is the height coordinate of the feature map, and w is the width coordinate of the feature map. For a sign function, when Output 1 when Output -1 when Output 0 when the time is right; This is an operation that takes the larger of the two values; For the corresponding channel The soft threshold parameter; The features are after soft thresholding. This soft thresholding operation adaptively shrinks response values ​​with absolute values ​​below a threshold to zero, thereby suppressing background noise and preserving key features with larger amplitudes. The soft-thresholded feature map is element-wise added to the input feature map of the residual branch to complete the residual connection. The result of the addition is used as the output of the current residual shrinking module. All three cascaded residual shrinking modules perform the above processing, maintaining a total of 64 output channels. After the last residual shrinking module, a global average pooling layer is applied to compress the spatial dimension to 1×1, resulting in a 64-dimensional feature vector, which is the compressed feature vector.

[0042] After the compressed feature vector is generated, it is input into the first task decoder and the second task decoder, respectively. The first task decoder generates the basic cost prediction interval based on a quantile regression mechanism. The first task decoder consists of a fully connected layer with an input dimension of 64 and an output dimension of [missing information]. , This is the preset number of quantiles. In this embodiment, The value is 5, and the preset quantiles are respectively , , , , Fully connected layer output The scalar value, the first Output values That is, at the quantile The following are the predicted project costs. These five predicted values ​​constitute the basic cost prediction range, among which... As a reference for the lower bound of the interval As a reference for the upper bound of the interval, the median As the central estimate, the pre-trained engineering cost prediction network optimizes the first task decoder head during the training phase using a quantile loss function, which takes the form of: ,in , For indicator functions, This represents the actual cost, of which The residual between the actual value and the predicted value, when hour ,otherwise During the inference phase, the basic cost prediction interval is generated directly using the trained network parameters.

[0043] The second task decoding head calculates the sensitivity coefficient of each engineering node's contribution to the basic cost prediction interval through layer-by-layer deconvolution operations, and filters out key cost-sensitive nodes from all engineering nodes. After receiving the compressed feature vector, the second task decoding head first maps the 64-dimensional vector to a 512-dimensional vector through a fully connected layer, and then reshapes it into a tensor of size 4×4×32. The height and width of this tensor represent the initial resolution of the spatial dimension. Subsequently, it passes through three deconvolution layers in sequence. Each deconvolution layer has a kernel size of 4×4, a stride of 2, and padding of 1, with 32, 16, and 1 output channels respectively, and is equipped with batch normalization and linear rectified activation functions. The activation function is used after the first two deconvolution layers, and not after the last deconvolution layer. After three layers of deconvolution operations, a single-channel reconstructed feature map is generated. The spatial size of the reconstructed feature map is consistent with the spatial dimension of the cost dynamic feature tensor (i.e., the number of rows and columns after the total number of engineering nodes are arranged in space). Each spatial position on the reconstructed feature map corresponds to an engineering node. On this reconstructed feature map, the pixel values ​​at each engineering node are normalized to generate a contribution sensitivity coefficient. The normalization process involves summing the absolute values ​​of all pixel values ​​on the entire reconstructed feature map to obtain a total value. ,in This represents the total number of project nodes. To reconstruct the corresponding engineering nodes on the feature map The pixel value; then the project node Contribution sensitivity coefficient , Furthermore, the sum of the contribution sensitivity coefficients of all project nodes is 1. This contribution sensitivity coefficient reflects the degree to which each project node contributes to the overall uncertainty of the basic cost prediction interval.

[0044] After obtaining the contribution sensitivity coefficients of all project nodes, a preset sensitivity threshold is used for filtering. The preset sensitivity threshold is set to all. The average contribution sensitivity coefficient of each engineering node Add twice the standard deviation That is, threshold For the contribution sensitivity coefficient Exceed For each engineering node, its spatial location is marked as a key cost-sensitive node distribution. The final output key cost-sensitive node distribution is a binary mask image with the same dimension as the cost dynamic feature tensor space, where the marked positions are 1 and the rest are 0.

[0045] Example 3:

[0046] In practice, the basic cost prediction interval is adjusted for market fluctuations to generate a revised cost prediction interval. The basic cost prediction interval contains cost prediction values ​​corresponding to multiple preset quantiles output by the first task decoding head. These prediction values ​​collectively constitute a prediction interval for the total cost, which can be decomposed into three parts based on cost composition: material costs, labor costs, and machinery costs. The decomposition method is as follows: based on the ratio of the original pricing data of the labor cost, material cost, and machinery cost sub-items of each cost accounting unit in the set of sub-items of the target building project to the total cost of the corresponding unit, the total cost prediction value corresponding to each quantile is split into material cost prediction values, labor cost prediction values, and machinery cost prediction values ​​according to the same ratio, thus obtaining the material cost sub-interval, labor cost sub-interval, and machinery cost sub-interval in the basic cost prediction interval.

[0047] Obtain the price fluctuation amplitude and frequency of each material category in the material price fluctuation sequence within the current forecast window. The current forecast window is defined as a specified time interval, backwards from the current moment, with a length of 90 calendar days. Price fluctuation amplitude is the difference between the highest and lowest market prices for a single material category within the current forecast window, and price fluctuation frequency is the ratio of the number of times the price direction of a single material category changes within the current forecast window to the number of trading days included in the current forecast window. Calculate the price volatility intensity index for each material category based on the price fluctuation amplitude and price fluctuation frequency using the following formula:

[0048]

[0049] in, Indicates material category The price volatility intensity index has a value range of [0,1]. For material categories Price fluctuation range within the current forecast window, in yuan; This represents the maximum price fluctuation range for all material categories within the current forecast window. This represents the minimum price fluctuation range for all material categories within the current forecast window. At that time, let the normalized magnitude term be 0.5; The amplitude adjustment factor is set to 0.65. This value is based on attribution analysis of cost fluctuations in multiple historical projects. The analysis shows that price fluctuation amplitude explains approximately 65% ​​of the uncertainty in material costs, with the remaining 35% attributed to fluctuation frequency. Therefore, [the value is set as follows]. The value is 0.65; For material categories The price fluctuation frequency within the current forecast window is the ratio of the number of times the price direction changes to the number of trading days contained in the current forecast window, with a range of [0,1].

[0050] Based on the price volatility intensity index of each material category within the current forecast window, each material category is divided into a high-volatility category set and a low-volatility category set. The preset volatility threshold is set to 0.55, based on the following criteria: in the full historical project material price volatility data, material categories with a price volatility intensity index higher than 0.55 account for more than 80% of the total cost deviation; therefore, 0.55 is used as the dividing point. When material categories... of At that time, the material categories Classified into the high volatility category set; when the material category of At that time, the material categories It is classified into the low volatility category.

[0051] For each material category in the high-volatility category set, the width of its corresponding material cost sub-interval within the basic cost prediction interval is multiplied by a first stretching coefficient to generate a high-volatility material elasticity sub-interval. The first stretching coefficient is positively correlated with the price volatility intensity index, and is specifically calculated as follows: for a given material category in the high-volatility category set... Its first tensile coefficient ,in The stretching factor is set to 1.8. This value is based on statistics of the relationship between historically high-volatility material price fluctuations and the expansion of the cost range. For every 0.1 increase in the price volatility index, the reasonable cost range needs to expand by approximately 18%. Therefore, the stretching factor is set to 1.8. The material categories within the basic cost prediction range are then... The corresponding material cost sub-interval is denoted as The central value of the interval , interval half width Then the elastic subinterval of the high-volatility material is .

[0052] For each material category in the low-volatility category set, the width of its corresponding material cost sub-interval within the basic cost prediction interval is multiplied by a second stretching coefficient to generate a low-volatility material elastic sub-interval. The second stretching coefficient is a fixed constant set to 1.12, and it must be less than the minimum value of the first stretching coefficient for all high-volatility materials. This setting is based on the fact that while low-volatility materials have lower market risk, they still possess some unpredictability. Historical data shows that the final settlement price of low-volatility materials exceeds the original interval by an average of approximately 12%. Therefore, the second stretching coefficient is set to 1.12. Furthermore, since the minimum value of the first stretching coefficient for high-volatility materials appears in... When it is slightly greater than 0.55, its value is Clearly, 1.12 is less than 1.99. For a specific material category within a set of low-volatility categories... Its basic material cost sub-interval is central value and half width Similarly, the elastic sub-interval of low-fluctuation materials is .

[0053] All high-volatility material cost elasticity sub-intervals and all low-volatility material cost elasticity sub-intervals are weighted and summed according to the proportion of each material category in the cost composition to generate a material cost elasticity correction sub-interval. The proportion weight is determined based on the ratio of the total material cost of each material category in the sub-item project to the total material cost, and the sum of the proportion weights of all material categories is 1. During the weighted summation, the lower bound of the interval is also weighted and summed to obtain the lower bound of the material cost elasticity correction sub-interval. The upper bound of the material cost elasticity correction sub-interval is obtained by weighted summation over the upper bound of the interval. ,Right now , ,in Iterate through all material categories. For material categories The proportion and weight, and These are the lower and upper bounds of the corresponding elastic sub-intervals.

[0054] After obtaining the flexible adjustment sub-interval for material costs, adjustment sub-intervals for labor costs and machinery costs are generated. The rate of change of the labor cost adjustment index and the machinery cost adjustment index within the current forecast window is obtained. The labor cost adjustment index is determined by the construction market labor cost guidance price index regularly published by the local engineering cost management department, and the machinery cost adjustment index is determined by the adjustment coefficient of the machinery shift cost quota. The rate of change within the current forecast window is then calculated. The difference between the labor cost adjustment index at the start date and the labor cost adjustment index at the end date of the current forecast window is divided by the index value at the start date. The difference between the machinery cost adjustment index at the start date and the machinery cost adjustment index at the end date of the current forecast window is divided by the index value at the start date. Based on the rate of change of the labor cost adjustment index, the labor cost sub-items in the basic cost forecast interval are linearly adjusted proportionally to generate a labor cost correction sub-interval. If the basic labor cost sub-interval is... The corrected lower bound is... Upper Realm Based on the rate of change of the machinery cost adjustment index, the machinery cost sub-items in the basic cost forecast interval are linearly adjusted proportionally to generate a machinery cost correction sub-interval. If the basic machinery cost sub-interval is... The corrected lower bound is... Upper Realm .

[0055] The material cost adjustment sub-intervals, labor cost adjustment sub-intervals, and machinery cost adjustment sub-intervals are overlaid and merged to generate the revised cost forecast interval. The overlay and merging method is as follows: the lower bounds of the three sub-intervals are added together to obtain the lower bound of the revised cost forecast interval. The upper bound of the corrected cost prediction interval is obtained by adding the upper bounds of the three sub-intervals. The revised cost forecast range is The revised cost forecast interval can still be adjusted synchronously with reference to the center position of the basic cost forecast interval and the relative offset of each interval before and after the revision. That is, the predicted values ​​of each quantile of the basic cost forecast interval are linearly transformed according to the revision range of material cost, labor cost and machinery cost respectively, and then summed again to obtain a complete revised quantile prediction sequence, which is used for the formulation of subsequent cost control strategy.

[0056] Example 4:

[0057] In practical implementation, a cost control strategy is generated based on the distribution of key cost-sensitive nodes. The distribution of key cost-sensitive nodes is described by a binary mask image output by the second task decoding head, where the marked positions correspond to engineering nodes exceeding a preset sensitivity threshold. First, the sensitivity coefficient and node type identifier of each sensitive node are extracted from the key cost-sensitive node distribution. The sensitivity coefficient is generated by the second task decoding head, and the sensitivity coefficient of each sensitive node is the normalized value obtained after normalizing the feature response value at the corresponding position on the reconstructed feature map. Node type identifiers are attribute labels pre-assigned to each sub-item of the project node. These attribute labels are determined based on the proportion of each cost sub-item in the cost accounting unit. If the labor cost sub-item accounts for more than 50% of the total cost of the cost accounting unit, the node is marked as a labor-intensive node; if the material cost sub-item accounts for more than 50%, the node is marked as a material-intensive node; if the machinery cost sub-item accounts for more than 50%, the node is marked as a machinery-intensive node. If no single cost sub-item accounts for more than 50%, the node type is determined by the highest proportion. After extracting the sensitivity coefficient and node type identifier for each sensitive node, material procurement timing optimization parameters and fund disbursement rhythm adjustment parameters are generated for both labor-intensive and material-intensive nodes.

[0058] For nodes with sensitivity coefficients higher than the preset sensitivity threshold for labor-intensive nodes, material procurement timing optimization parameters are generated. The preset sensitivity threshold for labor-intensive nodes is set to 0.08. This value is set based on the analysis of the distribution of sensitivity coefficients of engineering nodes in all historical projects. The mean sensitivity coefficient of labor-intensive nodes is 0.05, and the standard deviation is 0.02. The threshold is set to the mean plus 1.5 times the standard deviation, i.e., 0.05 + 1.5 × 0.02 = 0.08, to filter out labor-intensive nodes with significantly high sensitivity. For each selected labor-intensive node, the construction schedule table for that node is obtained. The construction schedule table is part of the project construction organization design document, recording the demand time nodes and demand quantities for each material category of the corresponding sub-item project. The demand time nodes are expressed in calendar dates, and the demand quantities are expressed in the physical units of measurement of the corresponding material category. The demand time nodes and demand quantities for each material category are extracted from the construction schedule table.

[0059] Based on the demand time points and the historical price peak distribution of corresponding materials in the material price fluctuation sequence, the procurement delay tolerance for each material category is calculated. The material price fluctuation sequence is the material market price sequence used when constructing the cost time series knowledge graph. The historical price peak distribution is the set of time positions of local maxima in this sequence. The criterion for determining local maxima is: within a window of 7 sampling points before and after the current time point, the current price is the highest, and the price increase exceeds 15% of the previous trough price. The procurement delay tolerance is the number of days that can be offset between the demand time point and the historical price trough. The historical price trough is the time point corresponding to the lowest price between two adjacent peaks in the historical price peak distribution, and the most recent trough before the demand time point. If there is no trough before the demand time point that meets the conditions, the lowest price point within 30 calendar days before the demand time point is taken. The procurement delay tolerance is then calculated. When using the demand time point as the base date, the negative value of the number of calendar days tracing back to the historical price trough indicates the number of days that can be offset for advance procurement. Conversely, a positive value is used if the demand time point can be postponed to the next trough. However, the overall tolerance for material procurement delays is limited to the range [-60, +30] days. When the calculated result exceeds the lower limit of -60 days, it is taken as -60 days; when it exceeds the upper limit of 30 days, it is taken as 30 days. The basis for setting these upper and lower limits is that the general lead time for procurement of building materials should not exceed 60 days, while a demand delay of no more than 30 days is an acceptable range of flexibility for construction organization.

[0060] The adjustment value for the procurement lead time of each material category is determined based on the procurement delay tolerance for each category. The procurement lead time adjustment value is the corrected number of days to extend or shorten the original procurement lead time. The original procurement lead time is determined by the standard procurement cycle in the construction organization design, which is the standard number of days from the issuance of the purchase order to the delivery of materials to the site. Procurement lead time adjustment value The calculation method is as follows: when the procurement delay tolerance When it is negative, This means purchasing in advance, extending the original lead time for procurement by an amount equal to... Day; when procurement delay tolerance When it is a positive value, This means delaying procurement, shortening the original procurement lead time by an amount equal to... The adjusted procurement lead time is the original procurement lead time plus the adjusted procurement lead time value.

[0061] The batch quantity for each material category is determined by rounding up the ratio of the required quantity to the maximum single-purchase capacity. The maximum single-purchase capacity is determined by the supplier's minimum batch size limit or the project's storage capacity, and is the pre-set maximum single-purchase quantity for each material category. The required quantity for each material category is used as the basis for determining the batch quantity. Divide by the maximum capacity of a single purchase To obtain the ratio Quantity of each batch of purchases To Round up, that is .when hour, The lead time adjustment values ​​for each material category are associated and packaged with the batch purchase quantities according to the material category to generate material procurement timing optimization parameters. The associated package is in the form of a data structure with the material category identifier as the key. For each material category, the data structure contains the lead time adjustment value (unit: days) and the batch purchase quantity (unit: batch).

[0062] For material-intensive nodes with sensitivity coefficients exceeding a preset material sensitivity threshold, adjustment parameters for fund disbursement schedules are generated. The preset material sensitivity threshold is set to 0.10, based on the statistical analysis of the sensitivity coefficients of material-intensive nodes across all historical projects. The mean is 0.06, and the standard deviation is 0.025. Adding 1.6 times the standard deviation to the mean yields 0.10, thus identifying material-intensive nodes with significantly high sensitivity. For each selected material-intensive node, its funding plan is extracted. The funding plan is part of the project financial plan document, recording the planned disbursement amounts and timeframes for each construction phase of the corresponding sub-item project. The division of construction phases aligns with the phase divisions in the construction progress mapping relationship. The planned disbursement amounts for each construction phase are obtained from the funding plan. and planned disbursement timeline ,in This refers to the construction phase number.

[0063] Calculate the capital demand elasticity coefficient for each construction stage based on the ratio of material costs to labor costs in each stage. (Material cost ratio) Construction phase Total cost of all materials and construction phases The ratio of total costs; the proportion of labor costs Construction phase Total labor costs and construction phase The ratio of total cost. The elasticity of capital demand is positively correlated with the proportion of material costs. The formula for calculating the elasticity of capital demand is:

[0064]

[0065] in, Indicates the construction stage The elasticity of demand for funds, with a value range of [value missing]. However, in actual engineering, it usually falls within the range of [0.2, 3.0]. Construction phase The percentage of material costs is between 0 and 1. Construction phase The percentage of labor costs is between 0 and 1. The proportional adjustment factor is set to 0.8. Its value is based on fitting historical project financial data, which found that the average sensitivity coefficient between capital elasticity and the proportion of material costs is 0.8. To prevent the denominator from being zero, the smoothing constant is set to 0.01, ensuring that the denominator is always greater than zero. (Construction phase) Capital demand elasticity coefficient The larger the value, the more sensitive the stage is to adjustments in fund allocation, requiring greater adjustments when formulating the funding plan.

[0066] The adjusted fund disbursement ratio for each construction stage is calculated based on the fund demand elasticity coefficient for each stage. The adjusted fund disbursement ratio is the difference between the planned disbursement amount multiplied by the elasticity adjustment factor and the original planned disbursement amount. Elasticity adjustment factor. The decision is based on the elasticity coefficient of funding demand and the overall funding situation of the project. ,in This represents the average of the capital demand elasticity coefficients across all construction phases. The response coefficient is set to 0.5, meaning that when the elasticity coefficient at a certain stage deviates from the average by 1 unit, the payment ratio will be adjusted by 50%. This applies to the construction stage. Adjustment value of fund allocation ratio The calculation method is as follows: adjusted amount Then, the fund allocation ratio is adjusted. . It can be a positive or negative value. A positive value indicates that the amount to be allocated needs to be increased, while a negative value indicates that the amount to be allocated can be reduced. The unit is the same as the planned amount to be allocated.

[0067] Calculate the payment time offset for each construction phase based on the material supply market settlement cycle. Material supply market settlement cycle This is a standard payment period for materials suppliers, typically derived from historical transaction data in the materials supply market, and expressed in calendar days. Common values ​​are 30, 45, or 60 days. (For the construction phase...) Disbursement time offset The calculation method is as follows: if the material cost of this construction stage accounts for a certain percentage... Greater than the proportion of labor costs ,but Set as "Days" indicates that the disbursement time is shifted after the planned disbursement date to match material settlement payment terms and utilize the floating time of funds; if ,but Set as The number of days indicates the offset of the disbursement time from the planned disbursement time, to ensure timely payment of labor costs. When the calculated offset exceeds the range of [-30, +30] days, it is truncated according to the upper and lower limits, with the truncation value being -30 days or +30 days. The truncation is based on not exceeding the reasonable offset range of the monthly measurement.

[0068] Adjustment of fund disbursement ratios for each construction phase Offset from disbursement time The permutations and combinations are arranged according to the construction phase sequence to generate parameters for adjusting the fund disbursement schedule. The permutations and combinations are presented as a data structure indexed by the construction phase number. For each construction phase… The data structure includes the fund allocation ratio adjustment value (unit: yuan) and the allocation time offset (unit: days).

[0069] After generating the material procurement timing optimization parameters and fund disbursement rhythm adjustment parameters, the material procurement timing optimization parameters and fund disbursement rhythm adjustment parameters are coordinated and aligned according to the construction progress sequence to generate a cost control strategy. The coordination and alignment process is implemented as follows: using the execution order of engineering nodes in the construction progress sequence as the horizontal axis, the adjustment values ​​of the procurement lead time for each material category and the quantity of batch procurement are mapped to the corresponding demand time nodes. At the same time, the adjustment values ​​of the fund disbursement ratio for each construction stage and the disbursement time offset are mapped to the corresponding planned disbursement time nodes. Then, the time conflicts and cash flow gaps between the two are checked. The specific checking rules are as follows: for any construction stage, if the delivery time caused by the adjusted material procurement lead time is more than 3 days later than the planned start time of the construction stage, a material supply warning entry is generated, and the quantity of batch procurement for the affected material category is increased by 1 batch as a buffer; if the adjusted fund disbursement arrival time is more than 5 days later than the adjusted procurement payment time, a fund shortage warning entry is generated, and the disbursement time offset of the corresponding stage is adjusted forward by 5 days. After the aforementioned conflict checks and automatic adjustments, the finalized material procurement timing optimization parameters and fund disbursement rhythm adjustment parameters are integrated into a comprehensive cost control strategy plan. This plan includes a time-ordered material procurement execution plan and a fund disbursement execution plan. The material procurement execution plan lists the lead time, batches, and procurement timelines and quantities for each material category; the fund disbursement execution plan lists the fund disbursement amount, disbursement timelines, and disbursement ratios for each construction phase.

[0070] Example 5:

[0071] In specific implementation, please refer to Figure 4 The joint training process of the deep learning encoder and the engineering cost prediction network involves four stages: training data preparation, forward propagation calculation, composite loss function calculation, and backpropagation parameter update. The deep learning encoder includes a graph attention encoding layer, a temporal convolutional encoding layer, a feature cross-fusion module, and a tensor stacking module, with the structure of each module consistent with that described in Example 1. The engineering cost prediction network includes a shared feature extraction backbone, a first task decoding head, and a second task decoding head, with the structure of each module consistent with that described in Example 2.

[0072] In the training data preparation phase, historical cost time-series knowledge graph samples and actual final settlement cost data for each historical building project were acquired. The construction method for the historical cost time-series knowledge graph samples is the same as that used in Example 1: for each historical building project, the set of sub-item project nodes was obtained, the pricing deviation sequence for each cost accounting unit at multiple historical time nodes was extracted, and a mapping relationship between the pricing deviation sequence and the corresponding material market price was established based on the material category identifier in the material price fluctuation sequence. Furthermore, the weights of the mapping relationship were adjusted for time-series alignment by integrating the comparison deviation between planned and actual progress nodes in the construction progress mapping relationship, thus generating historical cost time-series knowledge graph samples. The actual final settlement cost data for each historical building project is the final audited and confirmed total final settlement cost amount for that historical building project, expressed in RMB. Data from a total of N historical building projects is collected, where N is an integer not less than 500. The historical cost time-series knowledge graph samples are used as training input data, and the actual final settlement cost data corresponding to each historical building project are used as training label data. The training input data and training label data are randomly divided into training and validation sets in an 8:2 ratio.

[0073] During the forward propagation computation phase, a batch of training input data from the training set is sequentially passed through a deep learning encoder and an engineering cost prediction network to obtain the training predicted cost intervals for each historical building project. A batch contains B samples, where B is 32. For each sample in the batch, the deep learning encoder receives samples from the historical cost time-series knowledge graph and outputs a cost dynamic feature tensor. The generation process of the cost dynamic feature tensor is consistent with that described in Example 1. The engineering cost prediction network receives the cost dynamic feature tensor, generates a compressed feature vector through shared feature extraction, and then the first task decoder generates a basic cost prediction interval based on a quantile regression mechanism. This basic cost prediction interval serves as the training predicted cost interval during training, containing predicted values ​​corresponding to K preset quantiles, where K is 5 and the preset quantiles are 0.05, 0.25, 0.50, 0.75, and 0.95. The second task decoder also performs forward propagation during the training phase, generating contribution sensitivity coefficients for each project node to assist gradient calculation but not directly participating in the generation of the training predicted cost interval.

[0074] After obtaining the predicted cost ranges, calculate the composite loss function value between each predicted cost range and the corresponding training label data. The training label data is a single-valued scalar. This represents the actual construction cost. The training-predicted cost interval consists of predicted values ​​at 5 quantiles, denoted as... , , , , The composite loss function value is composed of the weighted sum of the interval coverage width criterion loss component and the interval average width loss component. The formula for calculating the composite loss function is:

[0075]

[0076] in, The value is the composite loss function value, in square yuan or yuan, depending on the unit of the loss components; The loss balancing weight coefficient is set to 0.7. This value is based on the following criteria: on the validation set... Perform a grid search from 0.1 to 0.9 with a step size of 0.1. The optimal score is achieved by combining the interval coverage and the average interval width on the validation set. The overall score is defined as the weighted sum of the coverage probability and the normalized interval width. This represents the number of quantiles, with a value of 5. That is, at the quantile The predicted project cost is as follows; Let quantile loss function be the loss component of the interval coverage width criterion. , representing the average quantile loss predicted for all quantiles, this loss component guides the prediction interval to accurately cover the actual value; the average interval width loss component is... The difference between the 95th percentile predicted value and the 5th percentile predicted value represents the width of the prediction interval. This loss component guides the prediction interval to be as narrow as possible while covering the actual value. and The relative importance of the two loss components is controlled separately. This indicates that 70% weight is assigned to coverage accuracy and 30% weight is assigned to interval width penalty.

[0077] After calculating the composite loss function value, the first network weight parameters of the deep learning encoder and the second network weight parameters of the engineering cost prediction network are updated using the backpropagation algorithm based on the composite loss function value. The first network weight parameters include the linear transformation matrix of the graph attention encoding layer. and attention weight vector The first layer contains the kernel weights and biases of each dilated convolutional kernel in the temporal convolutional coding layer, as well as the learnable scaling coefficients in the feature cross-fusion module. The second network weight parameters include the weights and biases of all convolutional and fully connected layers in the shared feature extraction backbone, the weights and biases of the fully connected layers in the first task decoding head, and the weights and biases of the fully connected and deconvolutional layers in the second task decoding head. Backpropagation is used to calculate the composite loss function value. The gradients of the weight parameters of each first network and each weight parameter of each second network are used to update the parameters using the Adam optimizer. The initial learning rate of the Adam optimizer is set to 0.001. The momentum decay coefficient is set to 0.9. The squared gradient decay coefficient was set to 0.999, and the numerical stability constant ε was set to... Each batch completes one parameter update, and one iteration of the entire training set is counted as one training epoch. Training is performed for multiple epochs until the composite loss function value converges to below a preset threshold. The preset threshold is set to 0.005, based on the following criteria: In historical construction cost prediction tasks, when the composite loss function value is below 0.005, the average coverage of the prediction interval exceeds 90%, and the average width of the interval accounts for less than 15% of the actual cost, meeting the requirements for cost prediction accuracy in practical engineering applications. After each training epoch, the composite loss function value on the validation set is calculated. If the composite loss function value on the validation set does not reach a new low in five consecutive training epochs, an early stopping mechanism is triggered, and the model parameters corresponding to the lowest loss on the validation set are taken as the final result, resulting in the trained deep learning encoder and the construction cost prediction network. The first network weight parameters of the trained deep learning encoder and the second network weight parameters of the construction cost prediction network are saved for subsequent use when calling the pre-trained construction cost prediction network.

[0078] See Figure 6 In the graph, the horizontal axis represents the training epochs, ranging from 0 to 500, and the vertical axis represents the composite loss function value L, ranging from approximately -0.05 to 0.55. The solid line with dots represents the composite loss function value on the training set, and the dashed line with squares represents the composite loss function value on the validation set. As the number of training epochs increases, both the training and validation composite loss function values ​​show a significant decreasing trend. Initially, both are around 0.5, and as training progresses, the loss value gradually decreases. After approximately 200 training epochs, the loss function value drops below 0.1 and tends to stabilize in subsequent training, with reduced fluctuations. Finally, when the number of training epochs approaches 500, both curves reach low and similar values, indicating that the model achieves good fit on both the training and validation sets without significant overfitting. The trend of the curve is consistent with the expectation that the composite loss function value decreases with the training rounds during the joint training process in Example 5. This verifies that the training process of the deep learning encoder and the engineering cost prediction network is effective, and the composite loss function value successfully converges to below the preset threshold, thus realizing the training of a high-precision model for predicting construction engineering costs.

[0079] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A deep learning-based method for predicting construction project costs, characterized in that, The method includes: Construct a time-series knowledge graph of the cost of the target construction project. The time-series knowledge graph includes historical cost data of multiple project nodes, material price fluctuation sequences, and construction progress mapping relationships. The cost time-series knowledge graph is subjected to spatiotemporal feature aggregation processing by a deep learning encoder to generate the cost dynamic feature tensor of the target building project. The pre-trained engineering cost prediction network is invoked to perform multi-task decoding processing on the cost dynamic feature tensor to generate the basic cost prediction interval and key cost-sensitive node distribution of the target building project. The basic cost prediction range is adjusted for market volatility to generate a revised cost prediction range. Based on the distribution of key cost-sensitive nodes, a cost control strategy is generated, which includes material procurement timing optimization parameters and fund disbursement rhythm adjustment parameters.

2. The deep learning-based construction project cost prediction method according to claim 1, characterized in that, The construction of the time-series knowledge graph of the target building project's cost includes: Obtain the set of sub-item project nodes of the target construction project. Each sub-item project node corresponds to a cost accounting unit. The cost accounting unit includes the original pricing data of labor cost sub-items, material cost sub-items and machinery cost sub-items. Extract the pricing deviation sequence of each cost accounting unit at multiple historical time points. The pricing deviation sequence is the difference sequence between the actual pricing data and the quota benchmark data. Based on the material category identifier in the material price fluctuation sequence, a mapping relationship is established between the pricing deviation sequence and the corresponding material market price. The mapping relationship has timestamp and geographic attributes. By integrating the comparison deviation between planned progress nodes and actual progress nodes in the construction progress mapping relationship, the weights of the mapping associated edges are adjusted for temporal alignment, generating a cost time-series knowledge graph that includes node features, associated edges, and time attributes.

3. The deep learning-based construction project cost prediction method according to claim 2, characterized in that, The step of performing spatiotemporal feature aggregation processing on the cost time-series knowledge graph through a deep learning encoder to generate the cost dynamic feature tensor of the target building project includes: A graph attention coding layer is used to aggregate spatial neighborhood features of each engineering node in the cost time series knowledge graph. The graph attention coding layer calculates the attention allocation weight of the set of its neighboring nodes for each engineering node, and performs a weighted summation of the feature vectors of the neighboring nodes according to the attention allocation weight to generate the spatial aggregation features of each engineering node. A temporal convolutional coding layer is used to extract the temporal features of each engineering node using a multi-scale sliding window. The temporal convolutional coding layer contains multiple dilated convolutional kernels, and each dilated convolutional kernel captures the short-term fluctuation pattern and long-term trend pattern of the pricing deviation sequence with different receptive fields, thereby generating the temporal evolution features of each engineering node. The spatial aggregation features and the temporal evolution features are cross-fused along the feature dimension to generate a spatiotemporal fusion vector for each engineering node. The spatiotemporal fusion vectors of all project nodes are stacked into tensors according to the topological order of the cost time-series knowledge graph to generate the cost dynamic feature tensor.

4. The deep learning-based construction project cost prediction method according to claim 3, characterized in that, When the spatiotemporal fusion vectors of all engineering nodes are processed by tensor stacking according to the topological order of the cost time series knowledge graph, the topological order is determined according to the execution order of each engineering node in the construction progress time series, and the engineering nodes of the same construction level are arranged according to the numbering order of the sub-items of the project.

5. The deep learning-based construction project cost prediction method according to claim 3, characterized in that, When the graph attention encoding layer calculates the attention allocation weights for the set of neighboring nodes for each engineering node, the attention allocation weights are obtained by normalizing the concatenated vector of the feature vectors of the neighboring nodes and the feature vector of the current node after calculation by a single-layer feedforward network and the result by the softmax function. Furthermore, the graph attention encoding layer adopts a multi-head attention mechanism to concatenate or average the output results of multiple attention heads.

6. The deep learning-based construction project cost prediction method according to claim 1, characterized in that, The pre-trained engineering cost prediction network is invoked to perform multi-task decoding processing on the cost dynamic feature tensor to generate the basic cost prediction interval and key cost-sensitive node distribution of the target construction project, including: The cost dynamic feature tensor is input into the shared feature extraction backbone of the engineering cost prediction network. The shared feature extraction backbone uses a multi-level residual contraction structure to perform noise suppression and key feature enhancement on the cost dynamic feature tensor to generate a compressed feature vector. The compressed feature vector branch is input to the first task decoding head of the engineering cost prediction network. The first task decoding head generates the basic cost prediction interval based on the quantile regression mechanism. The basic cost prediction interval contains the cost prediction values ​​corresponding to multiple preset quantiles. The compressed feature vector branch is input to the second task decoding head of the engineering cost prediction network. The second task decoding head calculates the contribution sensitivity coefficient of each engineering node to the basic cost prediction interval through layer-by-layer deconvolution operation. Based on the contribution sensitivity coefficient, engineering nodes that exceed the preset sensitivity threshold are selected from all engineering nodes. The spatial location corresponding to the selected engineering nodes is marked as the key cost sensitive node distribution.

7. The deep learning-based construction project cost prediction method according to claim 6, characterized in that, When the second task decoding head calculates the contribution sensitivity coefficient of each project node to the basic cost prediction interval through layer-by-layer deconvolution operation, it reconstructs the compressed feature vector into a reconstructed feature map with the same dimension as the cost dynamic feature tensor space through deconvolution, and normalizes the feature response value of each project node on the reconstructed feature map as the contribution sensitivity coefficient of that project node.

8. The deep learning-based construction project cost prediction method according to claim 6, characterized in that, The basic cost forecast interval is adjusted for market volatility to generate a revised cost forecast interval, including: Obtain the price fluctuation amplitude and frequency of each material category in the material price fluctuation sequence within the current prediction window, and calculate the price fluctuation intensity index of each material category based on the price fluctuation amplitude and the fluctuation frequency; Based on the price fluctuation intensity index of each material category, the corresponding material cost sub-item in the basic cost prediction interval is elastically stretched to generate an elastic correction sub-interval for material costs. Obtain the rate of change of the labor cost adjustment index and the machinery cost adjustment index within the current prediction window. Based on the rate of change of the labor cost adjustment index, make a linear proportional adjustment to the labor cost sub-item in the basic cost prediction interval. Based on the rate of change of the machinery cost adjustment index, make a linear proportional adjustment to the machinery cost sub-item in the basic cost prediction interval, and generate a labor cost correction sub-interval and a machinery cost correction sub-interval. The material cost elastic correction sub-interval, labor cost correction sub-interval, and machinery cost correction sub-interval are superimposed and merged to generate the corrected cost prediction interval.

9. The deep learning-based construction project cost prediction method according to claim 8, characterized in that, The step of elastically stretching the corresponding material cost sub-items in the basic cost prediction interval based on the price fluctuation intensity index of each material category to generate an elastically corrected material cost sub-interval includes: Based on the price volatility intensity index of each material category within the current prediction window, each material category is divided into a high volatility category set and a low volatility category set. Differentiated elastic stretching is then applied to the material cost sub-intervals in the basic cost prediction intervals corresponding to each material category in the high volatility category set and the low volatility category set, respectively. Finally, all the stretched material category sub-intervals are weighted and summed according to the proportion of each material category in the cost composition to generate the material cost elastic correction sub-interval.

10. The deep learning-based construction project cost prediction method according to claim 9, characterized in that, The step of applying differentiated elastic stretching to the material cost sub-intervals within the basic cost prediction intervals corresponding to each material category in the high-volatility category set and the low-volatility category set includes: For each material category in the set of high volatility categories, the interval width of the corresponding material cost sub-interval in the basic cost prediction interval is multiplied by a first stretching coefficient to generate a high volatility material elasticity sub-interval. The first stretching coefficient is positively correlated with the price volatility intensity index. For each material category in the set of low-fluctuation categories, the interval width of the corresponding material cost sub-interval in the basic cost prediction interval is multiplied by the second stretching coefficient to generate a low-fluctuation material elastic sub-interval. The second stretching coefficient is a fixed constant and is less than the minimum value of the first stretching coefficient. Wherein, the price volatility intensity index of the material category in the high volatility category set is greater than a preset volatility threshold, and the price volatility intensity index of the material category in the low volatility category set is less than or equal to the preset volatility threshold.

Images