New Energy Project Construction Expenditure Estimation System, Its Training Method and Application

Abstract

This invention belongs to the technical field of new energy construction planning. It discloses a system for estimating construction expenditures for new energy projects, its training method, and its application. In this system: a technology layer graph construction module constructs a knowledge spillover graph between technologies; a region layer graph construction module constructs a region-level non-materialized knowledge spillover graph and a region-level materialized technology diffusion graph for each new energy technology; a node feature construction module constructs a technology node feature tensor corresponding to each technology node and a region-technology node feature tensor corresponding to each region-technology node; and a graph neural network prediction module sequentially performs feature linear transformation, graph convolution operation, feature fusion, and prediction operations to obtain predicted construction expenditure values ​​for different new energy technology projects in different regions. This invention, by selecting features closely related to new energy project construction expenditures and performing graph convolution operations on a two-layer structure graph, can achieve rapid prediction of new energy construction expenditures with high prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the technical field of new energy construction planning, and more specifically, relates to a new energy project construction expenditure estimation system and its training method and application. Background Technology

[0002] With the rapid development of new energy technologies such as offshore wind power, onshore wind power, and solar photovoltaic, their capital expenditure (CAPEX) directly affects the planning of new energy projects. When planning new energy projects, estimating the capital expenditure of building a new energy project of a certain scale in a certain region in advance can help assess whether the planned installed capacity is reasonable at the initial stage of project initiation, effectively avoid the risk of budget overruns, and provide accurate data support for subsequent financing arrangements.

[0003] In the evaluation of new energy projects, capital expenditures typically represent the total capital cost required for a power generation technology during the construction phase. This includes equipment procurement, civil construction, installation and commissioning, grid connection, development costs, design, approvals, financing, management, transportation, and other construction-period expenses. For example, the CAPEX of a photovoltaic power plant typically consists of components, inverters, BOS (balance system cost), and soft costs; the CAPEX of onshore wind power typically includes the wind turbine system, towers, foundations, installation, grid connection, and other soft costs; the CAPEX of offshore wind power is generally higher than that of onshore wind power, mainly because offshore foundations, submarine cables, construction vessels, grid connection, and operation and maintenance preparation costs are higher. These costs typically include the costs of foundation structures such as monopiles, jacket foundations, and floating foundations, as well as other soft costs. Due to the numerous expenditure items, expenditure estimation is time-consuming, labor-intensive, and often contains significant errors.

[0004] Therefore, how to quickly estimate the construction expenditure of new energy projects and ensure the accuracy of the estimate is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a new energy project construction expenditure estimation system and its training method and application, the purpose of which is to quickly estimate the construction expenditure of new energy projects and ensure the estimation accuracy.

[0006] To achieve the above objectives, the present invention is proposed.

[0007] According to a first aspect of the present invention, a system for estimating construction expenditures for new energy projects is provided, comprising: The technology layer graph construction module is used to perform the following: taking different new energy technology types as different technology nodes, for any technology node Ti and technology node Tj, analyze the technology similarity between the two and the cross-technology reference strength of Tj to Ti, and perform a weighted sum of technology similarity and cross-technology reference strength to obtain the cross-technology knowledge flow edge weights from Ti to Tj. By combining the cross-technology knowledge flow edge weights of different directions between each pair of nodes, a knowledge spillover graph between technologies is obtained. The regional layer graph construction module is used to perform the following for each new energy technology: Using different regions as different regional nodes, it analyzes the cross-regional reference strength of regional node Cj to regional node Ci for that new energy technology, using this strength as the cross-regional knowledge flow edge weight from Ci to Cj. By combining the cross-regional knowledge flow edge weights of different directions between each pair of nodes, it obtains the regional layer non-materialized knowledge spillover graph for that new energy technology. It also analyzes the transaction strength of Ci selling components related to that new energy technology to Cj, using this strength as the cross-regional transaction flow edge weight from Ci to Cj. By combining the cross-regional transaction flow edge weights of different directions between each pair of nodes, it obtains the regional layer materialized technology diffusion graph for that new energy technology. Different regional nodes under different new energy technologies are different region-technology nodes. The node feature construction module is used to construct the technology node feature tensor and the region-technology node feature tensor. The features of each technology node include the average cost of each component of the corresponding new energy technology, the installed capacity, and the number of technical documents. The features of each region-technology node include the average cost of each component of the corresponding new energy technology in the corresponding region, the installed capacity, the number of technical documents, and the sales revenue. The graph neural network prediction module includes a linear embedding layer, a technology layer graph convolutional unit corresponding to the inter-technology knowledge spillover graph, a region layer graph convolutional unit corresponding to the region layer non-materialized knowledge spillover graph and the region layer materialized technology diffusion graph, a cross-layer fusion unit, and a prediction head. The linear embedding layer is used to map the feature tensor of each technology node to the initial embedding representation of the corresponding technology node in the inter-technology knowledge spillover graph, and to map the feature tensor of each region-technology node to the initial embedding representation of the corresponding region-technology node in the region layer non-materialized knowledge spillover graph and the region layer materialized technology diffusion graph. The technology layer graph convolutional unit and the region layer graph convolutional unit are used to perform neighborhood aggregation on their corresponding graph structures to update the embedding representation of each node. The cross-layer fusion unit is used to fuse the updated embedding representations of all nodes to form a fused feature. The prediction head is used to predict the construction expenditure of different new energy technology projects in each region based on the fused feature.

[0008] According to a second aspect of the present invention, a training method for a new energy project construction expenditure estimation system is provided, the new energy project construction expenditure estimation system comprising a technology layer graph construction module, a regional layer graph construction module, a node feature construction module, and a graph neural network prediction module; the training method includes: Input new energy information from different historical periods into the new energy project construction expenditure estimation system; The technology layer graph construction module constructs a knowledge spillover graph between technologies for the corresponding time period; the region layer graph construction module constructs a region-level non-materialized knowledge spillover graph and a region-level materialized technology diffusion graph for the corresponding time period; and the node feature construction module constructs a technology node feature tensor and a region-technology node feature tensor for the corresponding time period. The graph neural network prediction module sequentially performs feature linear transformation, graph convolution operation, feature fusion and prediction operation to obtain the predicted construction expenditure values ​​of different new energy technology projects in different regions. The training loss is calculated based on the training loss function, and the network parameters of the graph neural network prediction module are updated by backpropagation. The optimization is iteratively performed until convergence, resulting in a well-trained new energy project construction expenditure estimation system. The training loss function includes the error function between the expenditure prediction value and the actual expenditure value.

[0009] According to a third aspect of the present invention, a method for estimating the construction expenditure of a new energy project is provided, comprising: The new energy information during the construction period is input into the new energy project construction expenditure estimation system mentioned above; the new energy project construction expenditure estimation system includes a technology layer graph construction module, a regional layer graph construction module, a node feature construction module, and a graph neural network prediction module; The technology layer graph construction module constructs a knowledge spillover graph between technologies during the construction period; the region layer graph construction module constructs a region-level non-materialized knowledge spillover graph and a region-level materialized technology diffusion graph during the construction period; and the node feature construction module constructs a technology node feature tensor and a region-technology node feature tensor during the construction period. The graph neural network prediction module sequentially performs feature linear transformation, graph convolution operation, feature fusion, and prediction operation to obtain the predicted construction expenditure values ​​for different new energy technology projects in different regions.

[0010] According to a fourth aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the training method of the above-mentioned new energy project construction expenditure estimation system or to implement the above-mentioned new energy project construction expenditure estimation method.

[0011] In summary, compared with the prior art, the technical solutions conceived in this invention have the following main advantages: 1. The new energy project construction expenditure estimation system proposed in this invention selects features closely related to the construction expenditure of new energy projects, including component costs and installed capacity. These features directly affect the expenditure on construction equipment. In addition, it also includes the number of technical documents and sales revenue. These features essentially reflect the development status of various new energy technologies in different regions, which affects the cost of constructing a certain new energy project in the corresponding region, such as design costs, installation costs, and management costs. The better the development of new energy technologies in a region, the lower the expenditure per unit installed capacity for constructing new energy projects. The two are naturally correlated. Therefore, selecting the above-mentioned features closely related to the construction expenditure of new energy projects as input features can improve the prediction accuracy of the system. On the other hand, this invention also constructs a technology-level inter-technology knowledge spillover graph, a region-level non-materialized knowledge spillover graph, and a region-level materialized technology diffusion graph. The inter-technology knowledge spillover graph reflects the degree of technical knowledge correlation between different categories of new energy technologies. The region-level non-materialized knowledge spillover graph reflects the dissemination relationship of non-physical new energy technology knowledge between different regions. The region-level materialized technology diffusion graph reflects the diffusion relationship of physical new energy technologies between different regions. The technological and regional relationships form a heterogeneous architecture that can complement each other. These relationships consider the interdependence between different new energy technologies and different regions, which inevitably affects the expenditure on different new energy projects in different regions. Therefore, this invention constructs a technological knowledge spillover graph reflecting the interdependence between different new energy technologies and different regions, a regional non-materialized knowledge spillover graph, and a regional materialized technology diffusion graph. It also selects node features closely related to new energy project construction expenditures to form a technological and regional structure graph. By performing graph convolution operations, the graph structure information and the node's own feature information can be fused, providing more comprehensive structured features for prediction. This enables rapid prediction of new energy construction expenditures with high accuracy. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the structure of a new energy project construction expenditure estimation system according to an embodiment of the present invention.

[0013] Figure 2 This is a flowchart illustrating the construction structure of the technical layer map construction module and the regional layer map construction module in one embodiment of the present invention.

[0014] Figure 3 This is a schematic diagram of a two-layer heterogeneous structure of the corresponding technology node domain and the region-technology node domain in one embodiment of the present invention.

[0015] Figure 4 This is a schematic diagram illustrating cross-layer heterogeneous information interaction between the technical layer graph convolutional unit and the regional layer graph convolutional unit in one embodiment of the present invention.

[0016] Figure 5 This is a flowchart of feature fusion and prediction in one embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0018] Example 1 This invention provides a system for estimating construction expenditures for new energy projects, such as... Figure 1 The diagram shown is a structural schematic of a new energy project construction expenditure estimation system according to an embodiment of the present invention, which includes a technology layer graph construction module, a regional layer graph construction module, a node feature construction module, and a graph neural network prediction module. Figure 2 The diagram shown is a flowchart of the technical layer map construction module and the regional layer map construction module in one embodiment of the present invention. The following is in conjunction with... Figure 1 and Figure 2 A detailed description of the expenditure estimation system for this new energy project is provided.

[0019] The technology layer graph construction module is used to construct a knowledge spillover graph between technologies. Specifically, it includes: taking different new energy technology types as different technology nodes, for any technology node Ti and technology node Tj, analyzing the technology similarity between the two and the cross-technology reference strength of technology node Tj to technology node Ti, and performing a weighted sum of technology similarity and cross-technology reference strength to obtain the cross-technology knowledge flow edge weights from technology node Ti to technology node Tj. Combining the cross-technology knowledge flow edge weights of different directions between each pair of nodes, the knowledge spillover graph between technologies is obtained.

[0020] Specifically, new energy technologies can be divided into wind power and photovoltaics, and wind power can be further divided into onshore wind power and offshore wind power.

[0021] Specifically, when analyzing different new energy technologies, one can retrieve technical literature corresponding to the new energy technology from publicly available databases. For example, one can search patent databases, using patents corresponding to each new energy technology as technical analysis indicators. Alternatively, one can search journals, papers, and other literature. Patent searches are simpler and more targeted; therefore, the following explanation uses patents as technical documents.

[0022] Taking patents as an analytical indicator, the process of constructing a knowledge spillover graph between technologies is as follows.

[0023] (1) Different new energy technologies are used as different technology nodes.

[0024] (2) Analyze the relevant patents for each technology node and determine the set of classification codes for the corresponding patents for each technology node; calculate the number of intersection classification codes and the number of union classification codes between the classification code sets of technology node Ti and technology node Tj, and calculate the ratio of the number of intersection classification codes to the number of union classification codes. Then, normalization is performed to obtain the technology similarity between technology node Ti and technology node Tj. .

[0025] Specifically, the patent identifier can be standardized first, and the IPC or Derwent classification code set can be extracted to characterize the sub-technology coverage of each technology.

[0026] For technology node Ti and technology node Tj, let their classification code sets be respectively and Then the ratio It can be represented as: .

[0027] Comparison value Normalization is performed to obtain the normalized technical similarity. .

[0028] (3) Analyze the relevant patents for each technology node and calculate the number of times technology node Tj cites the patents of technology node Ti. Then, normalization is performed to obtain the cross-technology reference strength of technology node Tj referencing technology node Ti. .

[0029] Specifically, if a patent at technology node Tj references a patent at technology node Ti, then knowledge is considered to flow from technology node Ti to technology node Tj. Let... The number of times technology node Tj cites the patent of technology node Ti. Normalization is performed to obtain the normalized cross-technology reference strength of technology node Tj referencing technology node Ti. .

[0030] (4) Regarding technical similarity and cross-technology citation strength We perform a weighted summation to obtain the cross-technology knowledge flow edge weights from technology node Ti to technology node Tj. .

[0031] Specifically, border rights The calculation formula is: ; In the formula, α is the weighting coefficient, which can be, for example, 0.5.

[0032] (5) By combining the weights of cross-technology knowledge flow edges with different directions between each pair of nodes, a knowledge spillover graph between technologies is obtained.

[0033] This invention constructs a knowledge spillover graph between technologies, using new energy technologies as nodes. It calculates the cross-technology knowledge flow edge weights between different nodes based on the similarity between technologies and the intensity of knowledge flow. This allows the cross-technology knowledge flow edge weights to simultaneously reflect the degree of technological proximity and the intensity of knowledge transfer between different new energy technologies. As a result, the graph neural network adaptively allocates the contribution of adjacent node information according to the strength of technological association during the information aggregation process, thereby enhancing the model's ability to learn about cross-technology knowledge spillover, changes in technology maturity, and the evolution of construction expenditure.

[0034] The regional layer graph construction module is used to construct a regional layer non-materialized knowledge spillover graph and a regional layer materialized technology diffusion graph for each new energy technology. Specifically, constructing the regional layer non-materialized knowledge spillover graph for each new energy technology involves: using different regions as different regional nodes, analyzing the cross-regional reference strength of regional node Cj to regional node Ci for that new energy technology as the cross-regional knowledge flow edge weight from regional node Ci to regional node Cj, and combining the cross-regional knowledge flow edge weights of different directions between each pair of nodes to obtain the corresponding regional layer non-materialized knowledge spillover graph for the new energy technology. Constructing the regional layer materialized technology diffusion graph for each new energy technology involves: using different regions as different regional nodes, analyzing the transaction strength of regional node Ci selling components related to that new energy technology to regional node Cj as the cross-regional transaction flow edge weight from regional node Ci to regional node Cj, and combining the cross-regional transaction flow edge weights of different directions between each pair of nodes to obtain the corresponding regional layer materialized technology diffusion graph for the new energy technology. Different regional nodes under different new energy technologies are considered as different regional-technology nodes.

[0035] Taking patents as an analytical indicator, the process of constructing a regional layer non-material knowledge spillover graph for any new energy technology is as follows.

[0036] (1) Different regions are selected as different regional nodes. Among them, regions with better development of new energy technologies are selected.

[0037] (2) Analyze the target new energy technology related patents for each regional node. The target new energy technology is the new energy technology currently being analyzed. Calculate the number of times regional node Cj cites the patents of regional node Ci. The data is then normalized to obtain the cross-regional reference strength of the target new energy technology from region node Cj to region node Ci, and this strength is used as the cross-regional knowledge flow edge weight from region node Ci to region node Cj. .

[0038] Specifically, the normalization process can be performed using maximum value normalization.

[0039] (3) By combining the cross-regional knowledge flow edge weights with different directions between each pair of nodes, a non-material knowledge spillover graph of the regional layer corresponding to the target new energy technology is obtained.

[0040] For each new energy technology, a corresponding regional layer non-material knowledge spillover graph is constructed.

[0041] This invention constructs a regional-level non-materialized knowledge spillover graph for each new energy technology. Using regions as nodes, it calculates the cross-regional knowledge flow edge weights between different nodes based on the intensity of knowledge flow between regions. This transforms the technological knowledge propagation relationships of the same new energy technology across different regions into graph-structured edge weights. Since technical information reflects regional technological accumulation, engineering scheme maturity, and equipment selection experience, it has an objective correlation with new energy project construction expenditures. Therefore, it can provide the model with structured technical features related to construction expenditure estimation.

[0042] The specific process for constructing a regional-level physical technology diffusion map for any new energy technology is as follows.

[0043] (1) Different regions are used as different regional nodes.

[0044] (2) Analyze the transaction value related to the target new energy technology for each regional node. The target new energy technology is the new energy technology currently being analyzed. Calculate the transaction value of regional node Ci selling components related to the target new energy technology to regional node Cj. Then, normalization is performed to obtain the cross-regional transaction flow edge weights from region node Ci to region node Cj. .

[0045] Specifically, the normalization process can be performed using maximum value normalization.

[0046] For example, if region i sells new energy-related components to region j, then a directed edge is constructed from i to j. Let... Let i be the set of transaction records from the selling region i to the importing region j. Let r be the transaction value of transaction record r. Then the original edge weights of the transaction flow from i to j can be expressed as: ; Subsequently, the original edge weights of the transaction flow are normalized to their maximum values ​​to obtain the cross-regional transaction flow edge weights pointing from region node Ci to region node Cj. : .

[0047] (3) By combining the cross-regional transaction flow edge rights with different directions between each pair of nodes, the regional layer materialization technology diffusion map corresponding to the target new energy technology is obtained.

[0048] For each new energy technology, a corresponding regional physicochemical technology diffusion map is constructed.

[0049] This invention constructs a regional-level materialized technology diffusion map for each new energy technology, using regions as nodes. It calculates the cross-regional transaction flow edge weights between different nodes based on the intensity of transaction flows between regions. This transforms the transaction flow relationships of the same new energy technology across different regions into graph-structured edge weights. Since the intensity of transaction flows reflects the supply path of equipment components, the availability of technical equipment, and engineering support conditions, and is objectively related to the asset investment in new energy projects, it helps the model learn the differences in expenditures for constructing new energy projects in different regions.

[0050] The regional-level non-materialized knowledge spillover graph constructed above is a cross-technology relationship graph between different technologies. The regional-level non-materialized knowledge spillover graph and the regional-level materialized technology diffusion graph are cross-regional relationship graphs for the same technology, respectively. The above-mentioned multi-type structural graphs can characterize the impact on the construction expenditure of new energy projects from different dimensions of innovation knowledge flow and equipment-driven technology diffusion. The different structural graphs complement each other and can unify the differences in technical status reflected by cross-technology knowledge transfer, cross-regional patent knowledge dissemination, and cross-regional equipment component flow into structured features that can be learned by graph neural networks. This allows the model to simultaneously utilize information on technology association, regional association, and trade supply association when estimating construction expenditure, thereby improving the ability to represent the differences in investment in different regions and different new energy technology assets.

[0051] The node feature construction module is used to construct the technology node feature tensor corresponding to each technology node and the region-technology node feature tensor corresponding to each region-technology node. The features of each technology node include the average cost of each component of the corresponding new energy technology, the installed capacity, and the number of technical documents. The features of each region-technology node include the average cost of each component of the corresponding new energy technology in the corresponding region, the installed capacity, the number of technical documents, and the sales revenue.

[0052] Specifically, different new energy technology projects involve different components.

[0053] For example, the components involved in the construction of photovoltaic power generation projects include: photovoltaic modules, inverters, brackets, combiner boxes, box-type transformers, grid connection equipment, cables, monitoring equipment, and energy storage interface equipment.

[0054] For example, the components involved in the construction of onshore wind power projects include: wind turbine blades, hubs, gearboxes, generators, converters, main bearings, yaw systems, pitch systems, nacelles, towers, foundation structures, cables, and grid connection equipment.

[0055] For example, the components involved in the construction of offshore wind power projects include: offshore wind turbines, blades, hubs, gearboxes, generators, converters, nacelles, towers, monopile foundations, jacket foundations, offshore substations, submarine cables, collector lines, landing cables, and grid connection equipment.

[0056] In one embodiment, for each new energy technology, a candidate set of components can be determined first, and then a preset number of components most relevant to the project construction expenditure can be selected from the candidate set based on the feature selection method with the maximum information coefficient to participate in the construction of the node feature tensor.

[0057] Feature selection using the Maximum Information Coefficient (MIC) measures the strength of the association between each feature and the target variable, thereby selecting the most relevant features and eliminating redundant or irrelevant features. Feature selection can reduce redundant input before graph learning.

[0058] In this invention, a knowledge spillover graph between technologies is constructed, with nodes representing technology nodes. Furthermore, a regional-level non-materialized knowledge spillover graph and a regional-level materialized technology diffusion graph are constructed for each new energy technology. Different regional nodes under different new energy technologies are called regional-technology nodes. For example, region A under the wind power technology dimension, region B under the wind power technology dimension, and region A under the photovoltaic technology dimension are all different regional-technology nodes. Therefore, the number of technology nodes is the same as the number of types of new energy technologies involved in the analysis. Assuming there are 3 types of new energy technologies involved in the analysis, the number of technology nodes is 3. The number of regional-technology nodes is equal to the number of types of new energy technologies involved in the analysis multiplied by the number of regions involved in the analysis. Assuming there are 3 types of new energy technologies involved in the analysis and 5 regions involved in the analysis, the number of regional-technology nodes is 15.

[0059] For a knowledge spillover graph between technologies, the characteristics of each technology node include the average cost, installed capacity, and number of technical documents for each component of the corresponding new energy technology. In this case, data statistics only distinguish between technology types, not regions, representing statistical data across the entire region. For example, the average cost of a component of any new energy technology represents the average cost of that component across the entire region during the analysis period; the installed capacity of any new energy technology represents the total installed capacity of projects related to that new energy technology across the entire region during the analysis period; and the number of technical documents for any new energy technology represents the number of technical documents related to that new energy technology across the entire region during the analysis period. By statistically analyzing the characteristics of each technology node during the analysis period, the node feature tensor of each technology node is obtained. .

[0060] For the regional-level non-materialized knowledge spillover graph and the regional-level materialized technology diffusion graph, the characteristics of each regional-technology node include the average cost of each component of the corresponding new energy technology in the corresponding region, the installed capacity, the number of technical documents, and the sales revenue. In this case, the feature statistics of different regional-technology nodes must distinguish between both regions and technology types. For example, the average cost of each component of any new energy technology in any region represents the average cost of that component in that region during the analysis period; the installed capacity of any new energy technology in any region represents the total installed capacity of that new energy technology project in that region during the analysis period; and the number of technical documents for any new energy technology in any region represents the number of technical documents related to that new energy technology in that region during the analysis period. By statistically analyzing the characteristics of each regional-technology node during the analysis period, the node feature tensor of each regional-technology node is obtained. .

[0061] Subsequently, a graph neural network prediction module is required.

[0062] The graph neural network prediction module includes a linear embedding layer, a technology layer graph convolutional unit, a region layer graph convolutional unit, a cross-layer fusion unit, and a multilayer perceptron prediction head. The linear embedding layer maps the feature tensor of each technology node to the initial embedding representation of the corresponding technology node in the inter-technology knowledge spillover graph, and maps the feature tensor of each region-technology node to the initial embedding representation of the corresponding region-technology node in the region-level non-materialized knowledge spillover graph and the region-level materialized technology diffusion graph. The technology layer graph convolutional unit performs neighborhood aggregation on the inter-technology knowledge spillover graph to update the embedding representation of each node. The region layer graph convolutional unit performs neighborhood aggregation on the region-level non-materialized knowledge spillover graph and the region-level materialized technology diffusion graph to update the embedding representation of each node. The cross-layer fusion unit fuses the updated embedding representations of all nodes to form a fused feature. The multilayer perceptron prediction head predicts the construction expenditure of different new energy technology projects in each region based on the fused feature.

[0063] This invention constructs a two-layer heterogeneous graph neural network, which includes two coupled node domains: a technology node domain and a region-technology node domain. For example... Figure 3 The diagram shown is a schematic representation of a two-layer heterogeneous structure of a corresponding technology node domain and a region-technology node domain in one embodiment of the present invention. A detailed description follows.

[0064] Technology node domain.

[0065] Node feature tensors of each technical node First, an independent linear embedding layer is used to map to a shared latent space, obtaining the initial embedding representation of each technology node. At this point, each technology node in the inter-technology knowledge spillover graph has an initial embedding representation, which is input into the subsequent technology layer graph convolution units for graph convolution operations in graph structure form.

[0066] The technical layer graph convolutional unit contains a graph convolutional network (GCN) that performs GCN aggregation on the input technical layer structure graph. For any node i in any structure graph r, its neighborhood aggregation can be represented as: ; In the formula, This represents the GCN aggregation update result of node i in the structure graph r. Let r be the set of neighbors of node i in the structure graph. Let be the edge weight from node i to node j in the structure graph r. Let be the embedding representation of neighbor node j. Let r be the trainable weight matrix corresponding to the structure graph r.

[0067] In one embodiment, the technical layer graph convolutional unit further includes a GraphSAGE network. After performing GCN aggregation, a GraphSAGE-style update method is used to fuse the node's own representation with the neighborhood aggregation representation, specifically as follows: ; In the formula, This represents the GraphSAGE update result for node i.

[0068] Through the above aggregation method, the updated embedded representation of each technology node can be obtained, denoted as... The update embedding of this technology node can reflect the inherent cost drivers and cross-technology knowledge spillover effects of various new energy technologies. All of these factors will directly affect the construction expenditure of new energy projects.

[0069] Geographic-Technology Node Domain.

[0070] Node feature tensors of each region-technology node The initial embedding representations of each region-technology node are obtained by mapping independent linear embedding layers to a shared latent space. At this point, each region-technology node in the region-layer non-materialized knowledge spillover graph and the region-layer materialized technology diffusion graph has an initial embedding representation, which is input into the subsequent region-layer graph convolution unit for graph convolution operation in the form of a graph structure.

[0071] In one embodiment, the regional layer graph convolutional unit includes a graph convolutional network (GCN) that performs GCN aggregation on the input regional layer structure graph. Its aggregation form is the same as that of the GCN aggregation form in the technical layer.

[0072] In one embodiment, the regional layer graph convolutional unit further includes a GraphSAGE network. After performing GCN aggregation, a GraphSAGE-style update method is used to fuse the node's own representation with the neighborhood aggregated representation. Its aggregation form is the same as the GraphSAGE aggregation form in the technical layer.

[0073] Through the above aggregation method, the updated embedded representation of each region-technology node can be obtained, denoted as... The updated embedding of this region-technology node can reflect the comprehensive status of the target region, including the accumulation of new energy technologies, cross-regional technology knowledge input, and the intensity of equipment component transactions. All of these factors directly affect the construction expenditure of new energy projects.

[0074] In one embodiment, the technology layer graph convolutional unit and the region layer graph convolutional unit can also perform cross-layer heterogeneous information interaction, such as... Figure 4 The diagram illustrates cross-layer heterogeneous information interaction between a technology layer graph convolutional unit and a region layer graph convolutional unit in one embodiment of the present invention. Specifically, the embedded representation of the corresponding technology node in the technology layer is mapped to the region-technology node of the technology in different regions, thereby transmitting cross-technology knowledge spillover features to the region-technology node corresponding to the same technology in the region layer; and the representation of the same technology aggregated in each region node in the region layer is back-converged to the corresponding node in the technology layer, thereby feeding back the patent knowledge dissemination and equipment transaction diffusion features in different regions to the corresponding technology node in the technology layer; the two control the information transmission intensity between the upper and lower layers through cross-layer connection weights, so that the common features of technology and the regional differences can be integrated.

[0075] In one embodiment, after performing graph aggregation, the technical layer graph convolutional unit and the regional layer graph convolutional unit can continue to perform operations such as layer normalization, nonlinear activation, dropout, and residual connections to stabilize the training process and retain the original node information, and finally output an updated embedding representation.

[0076] like Figure 5The diagram shows a flowchart of feature fusion and prediction in one embodiment of the present invention. After obtaining the updated embedding representation of each node, the updated embedding representation of each region-technology node and the update of each technology node are concatenated by a cross-layer fusion unit to form a fused feature. The fused feature is then passed through a multilayer perceptron prediction head to output the construction expenditure of different new energy technology projects k in each region c. .

[0077] In summary, the above-mentioned new energy project construction expenditure estimation system selects features closely related to new energy project construction expenditures, including component costs and installed capacity, which directly affect construction equipment expenditures. Additionally, it includes the number of technical documents and sales revenue, which essentially reflect the development status of various new energy technologies in different regions. These features influence the soft costs of constructing a certain new energy project in that region, such as design costs, installation costs, and management costs. The better the development of new energy technologies in a region, the lower the soft costs of constructing new energy projects; there is a natural correlation between the two. Therefore, selecting these features closely related to new energy project construction expenditures as input features can improve the system's prediction accuracy. On the other hand, this invention also constructs a technology-level inter-technology knowledge spillover diagram, a regional-level non-materialized knowledge spillover diagram, and a regional-level materialized technology diffusion diagram. The inter-technology knowledge spillover diagram reflects the degree of technical knowledge association between different new energy technology categories. The regional-level non-materialized knowledge spillover diagram reflects the new energy non-physical technology knowledge dissemination relationship between different regions. The regional-level materialized technology diffusion diagram reflects the new energy physical technology diffusion relationship between different regions. The technology-level relationship and the regional-level relationship form a heterogeneous architecture, which can complement each other. The above relationships take into account the degree of dependence between different new energy technologies and different regions, which will inevitably affect the expenditure on building different new energy projects in different regions. Therefore, this invention constructs a technology-level knowledge spillover graph reflecting the degree of dependence between different new energy technologies and different regions, a region-level non-materialized knowledge spillover graph, and a region-level materialized technology diffusion graph. It also selects node features closely related to the construction expenditure of new energy projects to form a technology-level structure graph and a region-level structure graph. By performing graph convolution operations, the graph structure information and the feature information of the nodes themselves can be fused, providing more comprehensive structured features for prediction. This enables rapid prediction of new energy construction expenditure with high prediction accuracy.

[0078] Example 2 This invention also provides a training method for a new energy project construction expenditure estimation system, used to train the trainable parameters of the graph neural network prediction module in the new energy project construction expenditure estimation system described above. The specific training method is as follows: Input new energy information from different historical periods into the new energy project construction expenditure estimation system; The technology layer graph construction module constructs a knowledge spillover graph between technologies for the corresponding time period; the region layer graph construction module constructs a region-level non-materialized knowledge spillover graph and a region-level materialized technology diffusion graph for the corresponding time period; and the node feature construction module constructs a technology node feature tensor and a region-technology node feature tensor for the corresponding time period. The graph neural network prediction module sequentially performs feature linear transformation, graph convolution operation, feature fusion and prediction operation to obtain the predicted construction expenditure values ​​of different new energy technology projects in different regions. The training loss is calculated based on the training loss function, and the network parameters of the graph neural network prediction module are updated by backpropagation. The optimization is iteratively performed until convergence, resulting in a well-trained new energy project construction expenditure estimation system. The training loss function includes the error function between the expenditure prediction value and the actual expenditure value.

[0079] In one embodiment, the training loss function further includes a theoretical expenditure lower bound loss function. : ; In the formula, N represents the number of training samples. This represents the projected construction expenditure for new energy technology-related projects in region c during time period t. Let denot the lower bound of the theoretical expenditure of new energy technology-related projects (k).

[0080] In this embodiment, considering the theoretical expenditure lower bound loss function above, if the predicted value Greater than or equal to the theoretical expenditure lower bound This indicates that the prediction results do not violate physical feasibility, and the theoretical lower bound loss is 0. If the predicted value... Less than the theoretical expenditure lower bound If the prediction result violates physical feasibility, then the model is penalized based on the deviation between the two, thereby suppressing the occurrence of results that are lower than the theoretically feasible cost during the prediction process.

[0081] In one embodiment, the training loss function further includes a learning curve consistency loss. : ; ; In the formula, This represents the reference cost of the learning curve for constructing new energy technology k-related projects in region c during time period t. The actual expenditure value for new energy technology-related projects in region c, the starting historical period used for training. To construct the installed capacity of new energy technology-related projects in region c, which is the starting historical period used for training. The learning index is k, which corresponds to the new energy technology.

[0082] In this embodiment, considering the above-mentioned learning curve consistency loss, the constraint information that the cost of new energy technologies decreases with the increase of cumulative installed capacity can be introduced during the model training process. This makes the predicted construction expenditure results not only dependent on the data fitting error, but also conform to the objective cost evolution law formed by the large-scale construction of new energy technologies, the maturity of equipment manufacturing, and the accumulation of engineering implementation experience, thereby improving the reliability of the model's prediction of construction expenditure.

[0083] Therefore, considering the above three error terms, the theoretical cost lower bound constraint, and the learning curve consistency constraint, the training loss function can be expressed as: ; in, Used to measure prediction error Used to constrain the predicted value to be no less than the lower bound of the theoretical cost. Used to regularize long-term learning behavior and These are the weight coefficients for the corresponding constraint terms.

[0084] Example 3 This invention also relates to a method for estimating the construction expenditure of new energy projects, which includes: Input the new energy information during the construction period into the new energy project construction expenditure estimation system described above; the new energy project construction expenditure estimation system includes a technology layer graph construction module, a regional layer graph construction module, a node feature construction module, and a graph neural network prediction module; The technology layer graph construction module constructs a knowledge spillover graph between technologies during the construction period; the region layer graph construction module constructs a region-level non-materialized knowledge spillover graph and a region-level materialized technology diffusion graph during the construction period; and the node feature construction module constructs a technology node feature tensor and a region-technology node feature tensor during the construction period. The graph neural network prediction module sequentially performs feature linear transformation, graph convolution operation, feature fusion, and prediction operation to obtain the predicted construction expenditure values ​​for different new energy technology projects in different regions.

[0085] Furthermore, hard constraints can be applied to the construction expenditure forecasts to obtain physically consistent final forecasts. : .

[0086] Example 4 The present invention also relates to an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0087] The electronic device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory.

[0088] Example 5 The following analysis takes three new energy technologies—offshore wind power, onshore wind power, and solar photovoltaic—as examples, and selects five major new energy market regions for analysis. Candidate input features include component cost, installed capacity, patent records, and component trade data.

[0089] The selected comparison models included single-factor learning curves, two-factor learning curves, Extreme Gradient Boosting Machine (XGboost), Gradient Boosting Machine (GBM), Support Vector Machine (SVM), Random Forest (RF), Multilayer Perceptron (MLP), Transformer, Long Short-Term Memory Network (LSTM), and Bidirectional Gated Recurrent Neural Network (BiGRU). Model hyperparameters were determined through grid search. After training each model, the CAPEX prediction accuracy for various regions and technologies was evaluated on the test set. Evaluation metrics included Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Percentage Error (MAPE).

[0090] The comparison results are shown in Table 1 below.

[0091] Table 1

[0092] The results of the embodiments show that, from the perspective of model structure, traditional time series models mainly rely on the autocorrelation characteristics of historical CAPEX sequences, making it difficult to utilize patent knowledge spillover, regional diffusion relationships, and inter-technology correlation information; shallow machine learning models can utilize multi-dimensional explanatory variables, but usually treat samples from different technologies and regions as independent samples, making it difficult to express non-Euclidean graph structure dependencies; deep time series prediction models can characterize time dependencies, but still struggle to explicitly model the technology layer and regional layer relationship networks. Compared with the above models, the method of this invention characterizes the knowledge correlation between new energy technologies through a technology layer multi-relationship knowledge spillover graph, and characterizes the cross-regional diffusion mechanism through a regional layer non-materialized knowledge spillover graph and materialized technology diffusion graph. By constructing the above relationship graphs, the patent knowledge dissemination relationship and equipment trade diffusion relationship between different technologies and regions can be transformed into learnable non-Euclidean structure features, enabling the model to adaptively identify the source of key technologies, equipment supply paths, and cross-technology correlation strength during node aggregation, thereby improving the model's prediction accuracy.

[0093] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" are intended to illustrate the present invention and are not intended to limit the present invention.

[0094] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A system for estimating construction expenditures for new energy projects, characterized in that, include: The technology layer graph construction module is used to perform the following: taking different new energy technology types as different technology nodes, for any technology node Ti and technology node Tj, analyze the technology similarity between the two and the cross-technology reference strength of Tj to Ti, and perform a weighted sum of technology similarity and cross-technology reference strength to obtain the cross-technology knowledge flow edge weights from Ti to Tj. By combining the cross-technology knowledge flow edge weights of different directions between each pair of nodes, a knowledge spillover graph between technologies is obtained. The regional layer graph construction module is used to perform the following for each new energy technology: Using different regions as different regional nodes, it analyzes the cross-regional reference strength of regional node Cj to regional node Ci for that new energy technology, using this strength as the cross-regional knowledge flow edge weight from Ci to Cj. By combining the cross-regional knowledge flow edge weights of different directions between each pair of nodes, it obtains the regional layer non-materialized knowledge spillover graph for that new energy technology. It also analyzes the transaction strength of Ci selling components related to that new energy technology to Cj, using this strength as the cross-regional transaction flow edge weight from Ci to Cj. By combining the cross-regional transaction flow edge weights of different directions between each pair of nodes, it obtains the regional layer materialized technology diffusion graph for that new energy technology. Different regional nodes under different new energy technologies are different region-technology nodes. The node feature construction module is used to construct the technology node feature tensor and the region-technology node feature tensor. The features of each technology node include the average cost of each component of the corresponding new energy technology, the installed capacity, and the number of technical documents. The features of each region-technology node include the average cost of each component of the corresponding new energy technology in the corresponding region, the installed capacity, the number of technical documents, and the sales revenue. The graph neural network prediction module includes a linear embedding layer, a technology layer graph convolutional unit corresponding to the inter-technology knowledge spillover graph, a region layer graph convolutional unit corresponding to the region layer non-materialized knowledge spillover graph and the region layer materialized technology diffusion graph, a cross-layer fusion unit, and a prediction head. The linear embedding layer is used to map the feature tensor of each technology node to the initial embedding representation of the corresponding technology node in the inter-technology knowledge spillover graph, and to map the feature tensor of each region-technology node to the initial embedding representation of the corresponding region-technology node in the region layer non-materialized knowledge spillover graph and the region layer materialized technology diffusion graph. The technology layer graph convolutional unit and the region layer graph convolutional unit are used to perform neighborhood aggregation on their corresponding graph structures to update the embedding representation of each node. The cross-layer fusion unit is used to fuse the updated embedding representations of all nodes to form a fused feature. The prediction head is used to predict the construction expenditure of different new energy technology projects in each region based on the fused feature.

2. The new energy project construction expenditure estimation system as described in claim 1, characterized in that, The technology layer graph construction module is used to construct a knowledge spillover graph between technologies based on publicly available patents in the patent database; the specific process includes: Different types of new energy technologies are used as different technology nodes; Analyze the relevant patents for each technology node to determine the set of classification codes for the corresponding patents at each technology node; calculate the number of intersection and union classification codes between the classification code sets of technology nodes Ti and Tj, and calculate the ratio of the number of intersection classification codes to the number of union classification codes. Then, normalization is performed to obtain the technology similarity between technology node Ti and technology node Tj. ; Analyze the relevant patents for each technology node and calculate the number of citations of the patents of technology node Ti by technology node Tj. Then, normalization is performed to obtain the cross-technology reference strength of technology node Tj referencing technology node Ti. ; Technical similarity and cross-technology citation strength We perform a weighted summation to obtain the cross-technology knowledge flow edge weights from technology node Ti to technology node Tj. ; By combining the weights of cross-technology knowledge flow edges with different directions between each pair of nodes, a knowledge spillover graph between technologies is obtained.

3. The new energy project construction expenditure estimation system as described in claim 1, characterized in that, The regional layer map construction module constructs a regional layer non-materialized knowledge spillover map for each new energy technology based on publicly available patents in the patent database; the specific process includes: Different regions are used as different regional nodes; Analyze the target new energy technology-related patents for each regional node, where the target new energy technology is the current new energy technology being analyzed, and calculate the number of citations of the patents of regional node Ci by regional node Cj. The data is then normalized to obtain the cross-regional reference strength of the target new energy technology from region node Cj to region node Ci, and this strength is used as the cross-regional knowledge flow edge weight from region node Ci to region node Cj. ; By combining the cross-regional knowledge flow edge weights with different directions between each pair of nodes, a non-materialized knowledge spillover graph corresponding to the target new energy technology at the regional level is obtained.

4. The new energy project construction expenditure estimation system as described in claim 1, characterized in that, The regional layer map construction module constructs a regional layer physical technology diffusion map for each new energy technology based on transaction data; The specific process includes: Different regions are used as different regional nodes; Analyze the transaction value related to the target new energy technology for each regional node, where the target new energy technology is the one currently being analyzed, and calculate the transaction value of regional node Ci selling components related to the target new energy technology to regional node Cj. Then, normalization is performed to obtain the cross-regional transaction flow edge weights from region node Ci to region node Cj. ; By combining the cross-regional transaction edge rights with different directions between each pair of nodes, a regional layer materialization technology diffusion map corresponding to the target new energy technology is obtained.

5. The new energy project construction expenditure estimation system as described in claim 1, characterized in that, The technology layer graph convolutional unit includes a graph convolutional network and a GraphSAGE network. The inter-technology knowledge spillover graph is aggregated by graph convolutional network and then inductively aggregated by GraphSAGE network to obtain the updated embedding representation of each technology node. The region layer graph convolutional unit also includes a graph convolutional network and a GraphSAGE network. The region layer non-materialized knowledge spillover graph and the region layer materialized technology diffusion graph are aggregated by graph convolutional network and then inductively aggregated by GraphSAGE network to obtain the updated embedding representation of each region-technology node.

6. A training method for a new energy project construction expenditure estimation system, characterized in that, The new energy project construction expenditure estimation system is the new energy project construction expenditure estimation system as described in any one of claims 1 to 5, which includes a technology layer graph construction module, a regional layer graph construction module, a node feature construction module, and a graph neural network prediction module; the training method includes: Input new energy information from different historical periods into the new energy project construction expenditure estimation system; The technology layer graph construction module constructs a knowledge spillover graph between technologies for the corresponding time period; the region layer graph construction module constructs a region-level non-materialized knowledge spillover graph and a region-level materialized technology diffusion graph for the corresponding time period; and the node feature construction module constructs a technology node feature tensor and a region-technology node feature tensor for the corresponding time period. The graph neural network prediction module sequentially performs feature linear transformation, graph convolution operation, feature fusion and prediction operation to obtain the predicted construction expenditure values ​​of different new energy technology projects in different regions. The training loss is calculated based on the training loss function, and the network parameters of the graph neural network prediction module are updated by backpropagation. The optimization is iteratively performed until convergence, resulting in a well-trained new energy project construction expenditure estimation system. The training loss function includes the error function between the expenditure prediction value and the actual expenditure value.

7. The training method for the new energy project construction expenditure estimation system as described in claim 6, characterized in that, The training loss function also includes: Theoretical expenditure lower bound loss function : ； In the formula, N represents the number of training samples. This represents the projected construction expenditure for new energy technology-related projects in region c during time period t. Let denot the lower bound of the theoretical expenditure of new energy technology-related projects (k).

8. The training method for the new energy project construction expenditure estimation system as described in claim 6 or 7, characterized in that, The training loss function also includes: Learning curve consistency loss : ； ； In the formula, N represents the number of training samples. This represents the projected construction expenditure for new energy technology-related projects in region c during time period t. This represents the reference cost of the learning curve for constructing new energy technology k-related projects in region c during time period t. The actual expenditure value for new energy technology-related projects in region c, the starting historical period used for training. To construct the installed capacity of new energy technology-related projects in region c, which is the starting historical period used for training. The installed capacity of new energy technology-related projects to be constructed in region c during time period t. The learning index is k, which corresponds to the new energy technology.

9. A method for estimating construction expenditures for new energy projects, characterized in that, include: The new energy information during the construction period is input into the new energy project construction expenditure estimation system as described in any one of claims 1 to 5; the new energy project construction expenditure estimation system includes a technology layer graph construction module, a regional layer graph construction module, a node feature construction module, and a graph neural network prediction module; The technology layer graph construction module constructs a knowledge spillover graph between technologies during the construction period; the region layer graph construction module constructs a region-level non-materialized knowledge spillover graph and a region-level materialized technology diffusion graph during the construction period; and the node feature construction module constructs a technology node feature tensor and a region-technology node feature tensor during the construction period. The graph neural network prediction module sequentially performs feature linear transformation, graph convolution operation, feature fusion, and prediction operation to obtain the predicted construction expenditure values ​​for different new energy technology projects in different regions.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the training method for the new energy project construction expenditure estimation system as described in any one of claims 6 to 8, or the new energy project construction expenditure estimation method as described in claim 9.

Images