Power grid project cost evaluation auxiliary method and system based on multi-modal feature fusion

Abstract

The invention relates to the technical field of project cost review auxiliary methods, in particular to a power grid project cost review auxiliary method and system based on multi-modal feature fusion, and the method comprises the steps: obtaining project quantity list information, drawing information and historical cost data of a power grid project construction project; constructing an engineering quantity list knowledge graph based on the engineering quantity list information and the drawing information; constructing a historical project cost data graph based on the historical cost data; according to the project quantity list knowledge graph and the historical project cost data graph, generating a fusion feature vector through a multi-modal feature fusion network; predicting the future project cost based on the fused feature vector; and outputting the future project cost prediction result and the cost review suggestion, and through multi-modal feature fusion, the project quantity list, the drawing information and the historical cost data are comprehensively utilized, so that the data utilization rate and the comprehensiveness of information extraction are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering cost review assistance methods, in particular to an engineering cost review assistance method and system for power grids based on multi-modal feature fusion. Background Art

[0002] The review of power grid engineering cost is a key link in power engineering construction, directly affecting the cost control and economic benefits of projects. With the continuous expansion of the power grid scale and the improvement of project complexity, traditional manual review methods are difficult to meet the current requirements. In recent years, with the development of information technology, some auxiliary review systems have begun to be applied in the field of power grid engineering cost review.

[0003] Existing power grid engineering cost review assistance systems are mainly based on historical data analysis and simple statistical models. These systems usually adopt single-modal data input, such as only using historical cost data or bill of quantities information. Although this method improves the review efficiency to a certain extent, there are still many deficiencies. First, single-modal data input cannot comprehensively reflect the complexity of the project, resulting in limitations in the accuracy and reliability of the review results. Second, traditional statistical models are difficult to capture the non-linear relationships and long-term dependencies in engineering costs, making the prediction results often deviate greatly from the actual situation.

[0004] In addition, existing systems are inefficient in processing large-scale and high-dimensional engineering data and are difficult to adapt to the increasing data volume. At the same time, these systems lack an effective mechanism for identifying and automatically correcting abnormal items, and reviewers still need to invest a lot of time in manual review and adjustment. More importantly, existing systems generally lack in-depth exploration of the correlations between different types of data and cannot make full use of the complementarity of multi-source information, thus limiting the comprehensiveness and accuracy of the review results.

[0005] Facing these problems, the industry urgently needs an engineering cost review assistance method and system for power grids that can fuse multi-modal data and make full use of advanced machine learning technologies. Such a method should be able to comprehensively consider various types of information such as bill of quantities, drawing information, and historical cost data, deeply explore the complex relationships between the data, and improve the accuracy and reliability of cost prediction. At the same time, it should also have the ability to automatically identify abnormal items and self-correct, so as to reduce the workload of reviewers and improve the review efficiency. Summary of the Invention

[0006] The present invention precisely proposes an innovative solution to the above technical problems. By introducing multi-modal feature fusion technology and combining advanced deep learning algorithms, the present invention aims to construct a comprehensive, efficient, and intelligent auxiliary system for grid engineering cost review. This system can not only effectively integrate multi-source information but also automatically identify and process abnormal items, providing more reliable and comprehensive decision-making support for reviewers.

[0007] The present invention proposes an auxiliary method for grid engineering cost review based on multi-modal feature fusion, including:

[0008] An acquisition step, including:

[0009] Obtain the bill of quantities information, drawing information, and historical cost data of the grid engineering construction project;

[0010] A processing step, including:

[0011] Based on the bill of quantities information and drawing information, construct a bill of quantities knowledge graph;

[0012] Based on the historical cost data, construct a historical project cost data graph;

[0013] According to the bill of quantities knowledge graph and the historical project cost data graph, generate a fusion feature vector through a multi-modal feature fusion network;

[0014] Based on the fusion feature vector, predict the future project cost;

[0015] An output step, including:

[0016] Output the future project cost prediction result and cost review suggestions.

[0017] Preferably, the acquisition step specifically includes:

[0018] Perform text preprocessing on the bill of quantities information, including word segmentation, stop word removal, and normalization;

[0019] Use a pre-trained BERT model to map the preprocessed text to a high-dimensional vector space.

[0020] Preferably, the step of constructing the bill of quantities knowledge graph specifically includes:

[0021] Encode the bill of quantities text vector into a graph structure;

[0022] Use a graph attention network to process the graph structure to obtain a bill of quantities knowledge graph representation vector.

[0023] Preferably, the step of constructing the historical project cost data graph specifically includes:

[0024] Perform outlier handling, missing value handling, and duplicate record removal on the historical cost data;

[0025] Construct the preprocessed historical cost data into a graph structure;

[0026] Use a graph attention network to process the graph structure to obtain a graph representation vector of the historical project cost data.

[0027] Preferably, the multi-modal feature fusion network includes:

[0028] Two independent encoders, respectively used to process the bill of quantities knowledge graph and the graph of historical project cost data;

[0029] A cross-modal attention module for fusing features of different modalities.

[0030] Preferably, the processing steps of the cross-modal attention module include:

[0031] Calculate the attention score between the representation vector of the bill of quantities knowledge graph and the representation vector of the graph of historical project cost data;

[0032] Generate a fused feature vector based on the attention score.

[0033] Preferably, the steps of predicting the future project cost specifically include:

[0034] Construct a hybrid network model based on CNN and LSTM;

[0035] Input the fused feature vector into the hybrid network model to obtain the prediction result of the future project cost.

[0036] Preferably, it further includes a cost deviation analysis step:

[0037] Calculate the cost deviation based on the prediction result of the future project cost and the historical cost data;

[0038] Judge whether there are abnormal items according to a preset threshold.

[0039] Preferably, it further includes an automatic correction step:

[0040] Construct a temporal convolutional network based on the industry expert knowledge base;

[0041] Use the temporal convolutional network to analyze the audited quotes similar to the current project to judge whether the automatic correction mechanism is triggered;

[0042] If the automatic correction mechanism is triggered, correct the prediction result according to the expert rules.

[0043] A power grid project cost review assistance system based on multi-modal feature fusion for implementing the described method, comprising:

[0044] An acquisition module, configured to acquire the bill of quantities information, drawing information, and historical cost data of a power grid project construction project;

[0045] A preprocessing module, configured to preprocess the bill of quantities information and historical cost data;

[0046] A knowledge graph construction module, configured to construct a bill of quantities knowledge graph and a historical project cost data graph;

[0047] A feature fusion module, configured to generate a fused feature vector through a multi-modal feature fusion network;

[0048] A prediction module, configured to predict future project costs based on the fused feature vector;

[0049] A review module, configured to output the prediction result of future project costs and cost review suggestions.

[0050] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0051] Firstly, through multi-modal feature fusion, the present invention realizes the comprehensive utilization of the bill of quantities, drawing information, and historical cost data, greatly improving the data utilization rate and the comprehensiveness of information extraction. This fusion is not just a simple data splicing, but through a deep learning network, it realizes the deep association and complementarity between different modal data, thus capturing complex features and patterns that cannot be reflected by single-modal data.

[0052] Secondly, the graph structure representation and graph neural network processing method adopted by the present invention cleverly solve the problem that traditional methods are difficult to process high-dimensional complex data. By converting the bill of quantities and historical cost data into a graph structure, the system can better capture the similarities and correlations between projects, laying a solid foundation for subsequent feature extraction and cost prediction.

[0053] In addition, the cross-modal attention mechanism introduced by the present invention is an important innovation point. This mechanism can automatically learn the importance weights between different modal data, realizing the dynamic fusion of data. This not only improves the expressive ability of the model but also enhances the adaptability of the system to different types of engineering projects.

[0054] Another significant advantage of the present invention lies in its automatic anomaly detection and correction mechanism. By combining expert knowledge and machine learning algorithms, the system can intelligently identify potential abnormal items and automatically correct them according to historical data and industry rules. This greatly reduces the workload of reviewers and improves the reliability of review results at the same time.

[0055] Finally, the method of the present invention has strong scalability and adaptability. As new data continues to accumulate, the system can continuously learn and optimize, and its performance will improve over time. This self-improving feature enables the system to adapt to the dynamic changes in the field of grid project cost review and maintain long-term practical value.

[0056] Generally speaking, by innovatively integrating multi-modal data and advanced deep learning technologies, the present invention effectively solves many problems existing in the existing grid project cost review assistance system. It not only significantly improves the accuracy and reliability of cost prediction, but also greatly enhances the review efficiency, providing a powerful tool for cost control and decision-making support in grid project construction. With the wide application of this system, it is expected to promote the development of the entire grid project cost review field towards a more intelligent and refined direction. Brief Description of the Drawings

[0057] Figure 1 is the overall method flowchart of the present invention.

[0058] Figure 2 is the detailed logic block diagram of the preprocessing module of the present invention.

[0059] Figure 3 is the logic block diagram of the atlas construction module of the present invention.

[0060] Figure 4 is the logic block diagram of the feature fusion module of the present invention.

[0061] Figure 5 is the logic block diagram of the prediction module of the present invention.

[0062] Figure 6 is the logic block diagram of the prediction module of the present invention. Detailed Embodiments

[0063] Please refer to Figure 1-6 , the present invention provides a grid project cost review assistance method and system based on multi-modal feature fusion. The method mainly includes the following steps:

[0064] First, obtain the bill of quantities information, drawing information, and historical cost data of the grid project construction project. In a preferred embodiment of the present invention, the obtaining step can be completed by docking with the project management system or manual entry. Preferably, the bill of quantities information includes project name, project content, material specifications, quantities, etc.; the drawing information includes design drawings, construction drawings, etc.; the historical cost data includes historical cost records of similar projects.

[0065] Next, based on the obtained bill of quantities information and drawing information, a knowledge graph of the bill of quantities is constructed. The present invention uses a graph structure to represent the bill of quantities information, and this method can better capture the structured information in the data. Specifically, each item in the bill of quantities can be used as a node in the graph, and the relationship between items as an edge. For example, the graph structure can be represented by the formula:

[0066] G=(V, E),

[0067] where V is the set of nodes and E is the set of edges.

[0068] When constructing the graph structure, the following formula can be used to calculate the edge weights between nodes:

[0069]

[0070] where, w i j is the edge weight between node i and node j, sin(v i , v j ) is the similarity between node i and node j, and N(i) is the set of neighbor nodes of node i.

[0071] Then, based on the historical cost data, a historical project cost data graph is constructed. The purpose of this step is to transform the historical cost data into structured data that can be processed by machine learning algorithms. In an embodiment of the present invention, each historical project can be used as a node in the graph, and the similarity relationship between projects as an edge. The edge weight can be determined according to the similarity between projects. For example, the following formula can be used to calculate the similarity between projects:

[0072]

[0073] where, p i and p j are two projects, f i and f j are the corresponding feature vectors, and σ is the bandwidth parameter of the Gaussian kernel. In practical applications, the value of σ can be adjusted according to the specific data distribution, and usually the median of the Euclidean distance of the feature vectors in the dataset can be selected as the initial value.

[0074] Next, based on the constructed bill of quantities knowledge graph and historical project cost data graph, a fused feature vector is generated through a multi-modal feature fusion network. This step is one of the core innovations of the present invention. The multi-modal feature fusion network can effectively integrate information from different data sources, improving the comprehensiveness and accuracy of the model. In a preferred embodiment of the present invention, the multi-modal feature fusion network adopts the structure of a graph attention network (GAT). The core idea of GAT is to learn the importance weights between nodes through a self-attention mechanism. Specifically, the attention coefficient can be calculated using the following formula:

[0075]

[0076] where h i and h j are the feature vectors of nodes i and j respectively, W is a weight matrix, a is a weight vector, ∥ represents the vector concatenation operation. LeakyReLU is the activation function, and its slope parameter is usually set to 0.2.

[0077] Finally, based on the generated fused feature vector, the future project cost is predicted. In an embodiment of the present invention, a deep neural network model can be used for prediction. For example, a multi-layer perceptron (MLP) can be used as the prediction model:

[0078] y = f(W n (...f(W 2 f(W 1 x + b 1 ) + b 2 )...) + b n ),

[0079] where x is the input fused feature vector, W i and b i are the weight matrix and bias vector of the i-th layer respectively, f is the activation function (such as ReLU), and y is the predicted project cost.

[0080] After the prediction is completed, the method will output the future project cost prediction result and the cost review suggestion. These outputs can provide important reference information for project cost reviewers, helping them make more accurate and objective review decisions.

[0081] Next, the present invention further processes the bill of quantities information. Specifically, text preprocessing is performed on the bill of quantities information, including word segmentation, stop word removal, and normalization. The purpose of this step is to transform the original text data into a form that is more easily processed by machine learning algorithms. In a preferred embodiment of the present invention, word segmentation can adopt the maximum matching algorithm based on a dictionary, stop word removal can use a predefined stop word list, and normalization processing includes operations such as unifying case and deleting special characters.

[0082] For example, for the original text "install one 10kV power distribution cabinet", after preprocessing, a word sequence such as ["install", "10kV", "power distribution cabinet", "one"] may be obtained. This preprocessing method can effectively extract the key information in the text and remove the noise data that is not important for cost review.

[0083] After completing the text preprocessing, the method uses a pre-trained BERT model to map the preprocessed text into a high-dimensional vector space. BERT is a powerful pre-trained language model that can capture the deep semantic information of the text. In the present invention, we use the output of the [CLS] token of BERT as the representation of the entire sentence. Specifically, the output of BERT can be represented by the following formula:

[0084] h = BERT(x),

[0085] where x is the input word sequence and h is the output high-dimensional vector representation. Usually, the output dimension of BERT is 768.

[0086] This processing method of the present invention has significant advantages. First, through text preprocessing, we can effectively extract the key information in the bill of quantities and remove the irrelevant noise data. Second, using the BERT model can transform the text into high-quality vector representations, which can capture the deep semantic information of the text and provide a solid foundation for subsequent feature fusion and cost prediction.

[0087] When constructing the bill of quantities knowledge graph, the present invention adopts an innovative method. First, the bill of quantities text vectors are encoded into a graph structure. The purpose of this step is to transform the linear text data into a graph structure that can represent complex relationships. In an embodiment of the present invention, the following method can be used to construct the graph:

[0088] 1. Take each engineering project as a node in the graph.

[0089] 2. Calculate the similarity between nodes. If the similarity exceeds a preset threshold (e.g., 0.7), then add an edge between these two nodes.

[0090] The similarity between nodes can be calculated using cosine similarity:

[0091]

[0092] wherein, v i and v j are the BERT vectors corresponding to node i and node j respectively.

[0093] Next, use the graph attention network to process the constructed graph structure to obtain the knowledge graph representation vector of the bill of quantities. The core idea of the graph attention network is to learn the importance weights between nodes through the self-attention mechanism. In the present invention, we adopt the multi-head attention mechanism, which can learn the relationships between nodes from multiple perspectives. Specifically, the output of the k-th attention head can be expressed as:

[0094]

[0095] wherein, is the attention coefficient, W (k) is the weight matrix, and σ is a non-linear activation function (such as ReLU).

[0096] Finally, the outputs of multiple attention heads are concatenated to obtain the final node representation.

[0097] This method based on the graph structure and the graph attention network has multiple advantages. First, it can effectively capture the complex relationships between engineering projects, which are difficult to represent in traditional linear models. Second, through the attention mechanism, the model can automatically learn the importance weights between different projects, so as to better understand the overall structure of the bill of quantities. Finally, the multi-head attention mechanism allows the model to learn from multiple perspectives, further improving the richness and robustness of the representation.

[0098] Through the above steps, the present invention can transform complex bill of quantities information into a high-quality knowledge graph representation, laying a solid foundation for subsequent multi-modal feature fusion and cost prediction. This method can not only improve the accuracy of cost prediction, but also provide more intuitive and comprehensive information support for cost review personnel.

[0099] The method of the present invention adopts a series of innovative steps when processing historical cost data. First, perform outlier processing, missing value processing, and duplicate record deduplication on the historical cost data. The purpose of this step is to improve the data quality and ensure the accuracy of subsequent analysis.

[0100] In terms of outlier processing, in a preferred embodiment of the present invention, a quartile-based method is adopted. Specifically, calculate the quartiles Q 1 , Q 2 (median) and Q 3 , and then define the inner limit as:

[0101] [Q 1 -1.5×IQR, Q 3 +1.5×IQR],

[0102] where IQR = Q 3 - Q 1 is the interquartile range. Data points falling outside this range are considered potential outliers and require further analysis or processing.

[0103] For handling missing values, this method adopts different strategies according to the characteristics of the missing data. For data with missing completely at random (MCAR), records containing missing values can be directly deleted. For data with missing at random (MAR) or missing not at random (MNAR), the multiple imputation method can be used for filling. The basic idea of the multiple imputation method is to generate multiple possible complete data sets using the observed data, then analyze these data sets separately, and finally merge the results.

[0104] In terms of removing duplicate records, the present invention adopts a fast deduplication algorithm based on hashing. A hash value is calculated for each record, and then a hash table is used to quickly detect and delete duplicate records. This method can efficiently complete the deduplication operation on large-scale data sets.

[0105] After the above preprocessing, this method constructs the preprocessed historical cost data into a graph structure. In this graph, each node represents a historical project, and the edges represent the similarity relationships between projects. The weight of the edge can be determined by the similarity of project features. For example, the similarity between two projects i and j can be calculated using the following formula:

[0106]

[0107] where f i and f j are the feature vectors of projects i and j respectively, and σ is the bandwidth parameter of the Gaussian kernel. In practical applications, σ can be selected as the optimal value through cross-validation.

[0108] Next, the present invention uses a graph attention network to process the constructed graph structure to obtain a graph representation vector of the historical project cost data. The core idea of the graph attention network is to learn the importance weights between nodes through the self-attention mechanism. In an embodiment of the present invention, a multi-head attention mechanism is adopted, and the calculation of each attention head is as follows:

[0109]

[0110] where h i and h jis the feature vector of nodes i and j, and W (k) is the weight matrix of the k-th attention head, and a (k) is the attention vector, and σ is a non-linear activation function (such as ReLU). Finally, the outputs of multiple attention heads are concatenated to obtain the final node representation:

[0111] Finally, the outputs of multiple attention heads are concatenated to obtain the final node representation:

[0112]

[0113] where K is the number of attention heads, which can usually be set to 8 or 16.

[0114] The multi-modal feature fusion network of the present invention is the core part of the whole method. The network includes two independent encoders, which are respectively used to process the bill of quantities knowledge graph and the historical project cost data graph, and a cross-modal attention module, which is used to fuse the features of different modalities.

[0115] In a preferred embodiment of the present invention, both of the two independent encoders adopt the graph convolutional network (GCN) structure. The basic idea of GCN is to propagate and update the features of nodes through the adjacency relationship of the graph. Specifically, the calculation of the l-th layer of GCN can be expressed as:

[0116]

[0117] where is the adjacency matrix with self-loops added, is 's degree matrix, H (l) is the node feature matrix of the l-th layer, W (l) is a learnable weight matrix, and σ is a non-linear activation function.

[0118] The cross-modal attention module is another innovation point of the present invention. The purpose of this module is to effectively fuse the information from different modalities. In an embodiment of the present invention, the processing steps of the cross-modal attention module include:

[0119] First, calculate the attention scores between the representation vectors of the bill of quantities knowledge graph and the historical project cost data graph. Specifically, for the representation vector h q of the bill of quantities knowledge graph k and the representation vector h

[0120] e qk = v T tanh(W q h q + Wk h k ),

[0121] where W q 、W k and v are learnable parameters.

[0122] Then, the attention scores are normalized to obtain attention weights:

[0123]

[0124] Finally, based on the attention weights, a fused feature vector is generated:

[0125] h fused = ∑ k α qk h k .

[0126] This cross-modal attention mechanism can effectively capture the correlations between different modalities, thereby generating more informative fused features.

[0127] After obtaining the fused feature vector, the method of the present invention uses a hybrid network model based on CNN and LSTM for future project cost prediction. This hybrid model can simultaneously capture the spatial and temporal features of the data, thereby improving the prediction accuracy.

[0128] Specifically, first, a one-dimensional convolutional neural network (1DCNN) is used to extract the local patterns of the fused features:

[0129] h cnn = CNN(h fused ),

[0130] Then, the output of the CNN is fed into the LSTM network to learn the long-term dependencies:

[0131] h t = LSTM(h cnn , h t-1 ),

[0132] Finally, a fully connected layer is used to obtain the final prediction result:

[0133] y = Wh t + b,

[0134] where W and b are the weights and biases of the fully connected layer, respectively.

[0135] In this way, the method of the present invention can make full use of the advantages of multi-modal data, effectively fuse the bill of quantities information and historical cost data, so as to achieve more accurate project cost prediction. This can not only improve the efficiency of cost review, but also provide more reliable decision-making support for reviewers. After the method of the present invention completes the future project cost prediction, it also includes an important step of cost deviation analysis. The purpose of this step is to evaluate the reliability of the prediction results and identify possible abnormal items.

[0136] Specifically, this method first calculates the cost deviation based on the future project cost prediction results and historical cost data. In a preferred embodiment of the present invention, the relative error method is used to calculate the cost deviation:

[0137]

[0138] Among them, the historical mean refers to the average cost of historical projects similar to the current project.

[0139] After calculating the deviation, this method determines whether there are abnormal items according to a preset threshold. The selection of the threshold is crucial for the effect of anomaly detection. In the practice of the present invention, through a large number of experiments and expert experience, it is found that setting the threshold to 15% can achieve good results. That is to say, if the cost deviation of a certain project exceeds 15%, it will be marked as a potential abnormal item.

[0140] Preferably, this method also introduces the concept of an adaptive threshold. Considering that different types of engineering projects may have different degrees of cost fluctuations, the adaptive threshold can be dynamically adjusted according to the distribution of historical data. For example, the following formula can be used to calculate the adaptive threshold:

[0141] Threshold = μ + k·σ

[0142] Among them, μ is the mean of historical deviations, σ is the standard deviation of historical deviations, and k is an adjustable parameter, which can usually be set to 2 or 3. This method can better adapt to the characteristics of different projects and improve the accuracy of anomaly detection.

[0143] The method of the present invention also includes an innovative automatic correction step. The purpose of this step is to automatically correct the prediction results based on expert knowledge and historical data after discovering potential anomalies, so as to improve the accuracy and reliability of the final prediction.

[0144] First, this method constructs a temporal convolutional network (TCN) based on the industry expert knowledge base. The advantage of TCN is that it can effectively capture long-term dependencies, and at the same time, its computational efficiency is higher than that of traditional recurrent neural networks. In an embodiment of the present invention, the basic structure of TCN is as follows:

[0145] 1. Causal Convolution Layer: Ensure that the model does not utilize future information.

[0146] 2. Dilated Convolution: Capture long-term dependencies by increasing the receptive field.

[0147] 3. Residual Connection: Facilitate the training of deeper networks.

[0148] The computation of a key layer in TCN can be expressed as:

[0149]

[0150]

[0151] where * represents the convolution operation, W f and b f are learnable parameters, d is the dilation rate, and x t is the input sequence.

[0152] Next, the present method uses the constructed TCN to analyze the reviewed quotes similar to the current project and determine whether to trigger the automatic correction mechanism. The judgment criterion can be based on the statistical distribution of the prediction deviation. For example, if the deviation of the current prediction exceeds the 95th percentile of the deviation distribution of historical similar projects, the automatic correction mechanism is triggered.

[0153] If the automatic correction mechanism is triggered, the present method corrects the prediction result according to the expert rules. The expert rules here can be based on industry standards, historical experience, or rules learned by a machine learning model. For example, a simple correction rule may be:

[0154] Corrected predicted value = α · original predicted value + (1 - α) · historical mean

[0155] where α is a weight coefficient between 0 and 1, which can be dynamically adjusted according to the magnitude of the prediction deviation. The larger the deviation, the smaller α, that is, more inclined to the historical mean.

[0156] Preferably, the method of the present invention also introduces a feedback mechanism. After each automatic correction is completed, the system records the correction result and the actual cost data. This information is used to continuously optimize the TCN model and the correction rules, so that the system can continuously learn and improve.

[0157] Finally, the present invention also provides a power grid project cost review assistance system based on multi-modal feature fusion corresponding to the above method. The system includes multiple functional modules, and each module is responsible for a specific step in the method.

[0158] Specifically, the system includes:

[0159] Acquisition Module 1, which is used to acquire the bill of quantities information, drawing information, and historical cost data of power grid engineering construction projects. This module can be docked with other systems through interfaces or provide a user interface for manual input.

[0160] Preprocessing Module 2, responsible for preprocessing the bill of quantities information and historical cost data. This includes word segmentation, stop word removal, and normalization of text data, as well as outlier handling, missing value filling, and deduplication of historical data.

[0161] Knowledge Graph Construction Module 3, used to construct the bill of quantities knowledge graph and the historical project cost data graph. This module realizes the conversion from text and numerical data to graph structure.

[0162] Feature Fusion Module 4, which implements a multi-modal feature fusion network, including two independent encoders and a cross-modal attention module. This is the core part of the system, responsible for integrating information from different sources.

[0163] Prediction Module 5, which predicts future project costs based on the fused feature vectors. This module realizes a hybrid network model based on CNN and LSTM.

[0164] Review Module 6, responsible for outputting the prediction results of future project costs and cost review suggestions. This module also includes functions such as cost deviation analysis and automatic correction.

[0165] These modules work closely together to jointly implement the power grid project cost review assistance method of the present invention. The modular design of the system not only improves the maintainability and scalability of the code but also enables the system to flexibly respond to different application scenarios and requirements.

[0166] Generally speaking, the method and system provided by the present invention, through the innovative fusion of multi-modal data and the use of advanced deep learning technologies, provide a powerful auxiliary tool for power grid project cost review. It can not only improve the efficiency and accuracy of the review but also provide more comprehensive and reliable decision-making support for reviewers. With the continuous learning and optimization of the system, its performance is expected to be further improved, making greater contributions to the cost control of power grid engineering construction.

[0167] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A power grid project cost review auxiliary method based on multi-modal feature fusion, characterized in that: include: The acquisition steps include: Obtain bill of quantities information, drawing information and historical cost data for power grid construction projects; Processing steps include: Based on the bill of quantities information and drawing information, construct a bill of quantities knowledge graph; Based on the historical cost data, construct a historical engineering cost data graph; According to the bill of quantities knowledge graph and the historical engineering cost data graph, generating a fused feature vector through a multimodal feature fusion network; Based on the fused feature vector, predicting future project cost; Output steps include: Output the future project cost forecast results and cost review suggestions.

2. The method according to claim 1, characterized in that The acquisition step specifically includes: Performing text preprocessing on the bill of quantities information, including word segmentation, stop word removal and standardization; Use the pre-trained BERT model to map the pre-processed text into a high-dimensional vector space.

3. The method according to claim 1, characterized in that: The steps of constructing the bill of quantities knowledge graph specifically include: Encode the bill of quantities text vector into a graph structure; The graph structure is processed using a graph attention network to obtain a bill of quantities knowledge graph representation vector.

4. The method according to claim 1, characterized in that The steps of constructing the historical engineering cost data graph specifically include: Perform outlier processing, missing value processing and duplicate record removal on the historical cost data; Construct the preprocessed historical cost data into a graph structure; The graph structure is processed using a graph attention network to obtain a graph representation vector of historical engineering cost data.

5. The method according to claim 1, characterized in that The multimodal feature fusion network includes: Two independent encoders, one for processing the bill of quantities knowledge graph and the other for processing the historical construction cost data graph; A cross-modal attention module is used to fuse features from different modalities.

6. The method according to claim 5, characterized in that The processing steps of the cross-modal attention module include: Calculate the attention score between the bill of quantities knowledge graph representation vector and the historical engineering cost data graph representation vector; Based on the attention scores, a fused feature vector is generated.

7. The method according to claim 1, characterized in that The steps of predicting future project costs specifically include: Build a hybrid network model based on CNN and LSTM; The fused feature vector is input into the hybrid network model to obtain a future project cost prediction result.

8. The method according to claim 1, characterized in that It also includes cost deviation analysis steps: Calculate the cost deviation based on the future project cost forecast result and historical cost data; Determine whether there are any abnormal items based on the preset threshold.

9. The method according to claim 1, characterized in that: Also includes automatic correction steps: Build a temporal convolutional network based on the industry expert knowledge base; Using the temporal convolutional network to analyze reviewed quotations similar to the current project, and determine whether to trigger an automatic correction mechanism; If the automatic correction mechanism is triggered, the prediction results will be corrected according to the expert rules.

10. A power grid engineering cost review auxiliary system based on multimodal feature fusion that implements the method described in any one of claims 1 to 9, characterized in that: include: An acquisition module is used to obtain the bill of quantities information, drawing information and historical cost data of power grid construction projects; A preprocessing module, used for preprocessing the bill of quantities information and historical cost data; Graph construction module, used to construct the knowledge graph of bill of quantities and historical engineering cost data graph; A feature fusion module, used to generate a fused feature vector through a multimodal feature fusion network; A prediction module, used for predicting future project cost based on the fused feature vector; The review module is used to output future project cost forecast results and cost review suggestions.

Images