Workshop market value automatic prediction method based on deep learning

Abstract

The invention discloses a plant market value automatic prediction method based on deep learning, and the method comprises the following steps: collecting multi-source heterogeneous data, and forming a standardized data set; constructing a space diagram structure of the plant plot, and generating a space adjacency matrix; constructing a multi-modal feature vector set based on the standardized data set, and forming a node feature representation set; inputting the node feature representation set and the spatial adjacency matrix into an improved space-time diagram neural network model to obtain a joint embedding representation matrix; constructing a reinforcement learning controller, and optimizing the improved space-time diagram neural network model; outputting an estimation result set based on the optimized improved space-time diagram neural network model; and on the basis of historical error feedback and market change conditions, an online updating mechanism of the improved space-time diagram neural network model is triggered regularly or in real time. According to the invention, the improved space-time diagram neural network is adopted to realize automatic prediction of the factory market value.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and real estate information technology, and in particular to an automatic prediction method for the market value of factory buildings based on deep learning. Background Technology

[0002] In the current process of valuing factory buildings, common methods mainly rely on manual valuation and rule-based calculation. Manual valuation depends on the experience of valuers, which is inefficient and highly subjective. Rule-based engines, on the other hand, are mostly based on limited historical transaction data or land transfer information, and lack comprehensive utilization of business registration, policy announcements, location and supporting facilities, and factory building images, resulting in insufficient accuracy of valuation results.

[0003] Some studies have attempted to use deep learning for price prediction, but most are limited to structured data modeling, lacking the integration of textual and image information, and thus failing to comprehensively reflect the multidimensional influencing factors of factory building market value. Existing spatiotemporal graph neural networks are mainly applied in the transportation or energy sectors, with limited use in industrial real estate valuation, lacking modeling methods for factory building spatial relationships and policy dynamics. Furthermore, existing models are mostly statically trained with fixed parameters, making them difficult to adapt to market fluctuations and policy adjustments, and prone to accuracy degradation when migrated across regions.

[0004] Therefore, how to provide an automatic prediction method for the market value of factory buildings based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an automatic prediction method for the market value of factory buildings based on deep learning. This invention utilizes multi-source heterogeneous data fusion, an improved spatiotemporal graph neural network, and reinforcement learning dynamic optimization technology. It describes in detail the steps of predicting factory building value through spatial adjacency modeling, multimodal feature encoding, and spatiotemporal feature interaction fusion. It has the advantages of comprehensive data coverage, high valuation accuracy, and strong adaptability to changes in market policies.

[0006] An automatic prediction method for the market value of factory buildings based on deep learning according to an embodiment of the present invention includes the following steps: Collect heterogeneous data from multiple sources and perform data cleaning, missing value imputation, timestamp alignment and spatial coordinate unification to form a standardized dataset; Construct a spatial graph structure for factory plots, and assign weights to edges based on geographical distance, industry similarity, and policy overlap to generate a spatial adjacency matrix; A multimodal feature vector set is constructed based on a standardized dataset, and then normalized and aligned to form a set of node feature representations. By inputting the node feature representation set and the spatial adjacency matrix into the improved spatiotemporal graph neural network model, a joint embedding representation matrix is ​​obtained; Based on the joint embedding representation matrix, a reinforcement learning controller is constructed to optimize the improved spatiotemporal graph neural network model; The optimized and improved spatiotemporal graph neural network model is applied to the valuation task, and a set of valuation results is output. Based on historical error feedback and market changes, an improved spatiotemporal graph neural network model is updated online periodically or in real time.

[0007] Optionally, the step of collecting multi-source heterogeneous data and performing data cleaning, missing value imputation, timestamp alignment, and spatial coordinate unification to form a standardized dataset specifically includes: The data collected includes multi-source heterogeneous data such as land transfer data, factory transaction data, enterprise registration information, policy announcement text data, location and supporting facility data, and factory image data. The land transfer data includes fields such as plot number, transfer time, planned use, transaction price, and plot ratio; The factory building transaction data includes transaction time, area, number of floors, building year, transaction unit price, and transaction location code; The business registration information of the enterprise includes the enterprise name, industry classification, registered address and establishment date; The policy announcement text data includes the policy number, release date, applicable region, and corresponding text content; The location-related infrastructure data includes spatial distribution information of transportation facilities, industrial parks, schools, and service outlets around the site. The factory building image data includes images of the main facade of the factory building and remote sensing aerial images, with an image resolution of no less than 512×512 pixels; Data cleaning and processing were performed on land transfer data, factory transaction data, enterprise business registration information, policy announcement texts, location and supporting facilities data, and factory image data. Incomplete records were deleted and field formats were standardized. Missing values ​​in multi-source heterogeneous data are filled by using time series interpolation to fill in missing timestamps and by using nearest neighbor interpolation to fill in missing spatial attributes. Timestamp alignment is performed on time fields in multi-source heterogeneous data to ensure that time information in factory transaction data, land transfer data and policy announcement texts is unified into a standardized time series. Unify the spatial coordinates in multi-source heterogeneous data, convert the coordinate systems of data from different sources into a unified coordinate system, and establish the geographical location labeling of factory nodes; The land transfer data, factory transaction data, enterprise business information, policy announcement texts, location and supporting facilities data, and factory image data, which have been cleaned, filled, aligned and have unified coordinates, are formatted and integrated to generate a standardized dataset.

[0008] Optionally, a spatial map structure for the factory site is constructed, and weighted values ​​are applied to the edges based on geographical distance, industry similarity, and policy overlap to generate a spatial adjacency matrix. Specifically, this includes: Based on the spatial coordinate vector corresponding to each factory record in the standardized dataset, each factory plot is treated as a graph node, and a set of nodes is constructed, where each node corresponds to a factory sample with a unique geographical location. Based on the relationship between the factory nodes, determine whether to establish a connection edge in turn, and combine all the connection edges into an edge set. Specifically, if the Euclidean distance between two factory nodes in geographic space is less than the preset geographic proximity threshold, then establish a connection edge between them based on geographic proximity. If the industries to which the enterprises corresponding to the two factory nodes belong have upstream and downstream supply chain collaboration relationships, then establish a connection edge between the two to establish supply chain collaboration relationships. If two factory nodes are located in the same policy coverage area or are regulated by the same policy, then a connecting edge with a policy coverage relationship is established between them; For any two graph nodes that satisfy the connection conditions, calculate the weight value of the edge. The weight value of the edge consists of three parts: geographical proximity, industrial chain similarity, and policy coverage overlap, to obtain the edge weight set. By combining the set of nodes, the set of edges that meet the connection conditions, and the corresponding set of edge weights, a spatial graph structure is constructed, where each node represents a factory plot and each edge represents the connection relationship between nodes under the conditions of geographical proximity, industrial chain collaboration, or policy coverage. Based on the spatial graph structure, the set of nodes, the set of edges, and the corresponding set of edge weights are combined to generate a spatial adjacency matrix. Each element in the spatial adjacency matrix is ​​used to represent the spatial connection relationship and strength between any two factory nodes.

[0009] Optionally, the step of constructing a multimodal feature vector set based on a standardized dataset, performing normalization and alignment processing to form a node feature representation set specifically includes: The structured attribute vector, policy text encoding vector, and image feature vector corresponding to each factory record are extracted from the standardized dataset to obtain a multimodal feature vector set, where: The structured attribute vector consists of fields such as factory area, number of floors, building age, and land use, and is formed into a fixed-length numerical vector after numerical standardization. The policy text encoding vector obtains a contextual embedding representation that reflects the semantics of policy elements by inputting the policy announcement text corresponding to each record into a pre-trained language model; The image feature vector is obtained by inputting the factory image into a visual neural network model to extract image feature information, resulting in a fixed-dimensional image representation; The structured attribute vector, policy text encoding vector, and image feature vector are normalized respectively. The structured attribute vector is normalized using the zero mean unit variance method, while the policy text encoding vector and image feature vector are normalized using the vector norm method. The normalized structured attribute vector, policy text encoding vector, and image feature vector are aligned in a predetermined order, and the three types of modal sub-vectors are combined into a unified node feature vector by vector concatenation. The total dimension of the node feature vector is equal to the sum of the dimensions of each modal sub-vector. The node feature vector serves as a joint representation of a single factory node in terms of structured, textual, and image information. All node feature representations are arranged sequentially according to node number to form a set of node feature representations.

[0010] Optionally, the step of inputting the node feature representation set and the spatial adjacency matrix into the improved spatiotemporal graph neural network model to obtain the joint embedding representation matrix specifically includes: The node feature representation set and the spatial adjacency matrix are simultaneously input into the improved spatiotemporal graph neural network model, which includes two parallel backbone structures: a graph neural network branch and a Transformer branch. In the graph neural network branch, multi-order graph convolution operations are performed on the feature representation of each node based on the spatial adjacency matrix to calculate the spatial feature matrix; In the Transformer branch, based on the historical transaction price records and historical policy text content corresponding to each factory node in the standardized dataset, a multimodal temporal feature sequence is constructed in chronological order. Each time step contains the transaction price features of the node at the corresponding time point and the semantic features of the corresponding policy text after encoding. All time steps are combined to form the temporal input tensor of the node. The temporal input tensor is input into the multi-layer Transformer encoder structure, and position encoding, multi-head self-attention calculation, feedforward neural network transformation and layer normalization operations are performed sequentially to extract the deep feature representation of each node in the time dimension, thereby obtaining the temporal feature representation corresponding to each node. The temporal feature representations of all nodes are arranged according to the node number to form a temporal feature matrix. Each row of the temporal feature matrix corresponds to the temporal feature representation of a node under the historical transaction information and policy evolution sequence. The spatial feature matrix and the temporal feature matrix are simultaneously input into the interactive fusion layer. First, the feature dimensions of the two are uniformly mapped so that the spatial feature representation and the temporal feature representation can be operated in the same dimensional space. Then, based on the dimensional alignment, a feature concatenation operation is performed to combine the spatial feature vector and temporal feature vector corresponding to each node according to the feature dimension direction to form a node-level fused feature representation. After the splicing is completed, the fused feature representation is input into the multilayer perceptron structure. Through linear transformation, nonlinear activation and weight learning, the deep extraction of the fused features is achieved to obtain the joint embedding representation matrix. Each row of the joint embedding representation matrix corresponds to the comprehensive feature of a factory node under the dual contexts of spatial dependence and temporal evolution.

[0011] Optionally, in the graph neural network branch, multi-order graph convolution operations are performed on the feature representation of each node based on the spatial adjacency matrix. The computation of the spatial feature representation specifically includes: The connection relationships between all nodes are processed based on the spatial adjacency matrix. First, an identity matrix is ​​introduced into the original adjacency matrix to add self-loop connections. Then, the degree value of each node is calculated to form a degree matrix. Finally, the adjacency matrix after introducing self-loops is normalized using the symmetric normalization formula to obtain the normalized adjacency matrix. Set the set of node feature representations as the initial feature input, denoted as the initial node feature matrix. Each row of the initial node feature matrix corresponds to the multimodal feature representation of a factory node. Using the initial node feature matrix and normalized adjacency matrix as input, multi-layer graph convolution operations are performed sequentially. In each layer, a new node feature matrix is ​​calculated through neighborhood feature aggregation and weight transformation. Specifically, the node feature matrix of the current layer is multiplied by the normalized adjacency matrix to achieve neighborhood information fusion, and then multiplied by the trainable weight matrix of the current layer. After transformation by a nonlinear activation function, the output feature matrix of the layer is obtained. This output is used as the input of the next layer, and the above process is repeated until a preset number of graph convolution layers are completed, finally obtaining the node spatial feature representation under multi-layer perception. After completing the calculation of all graph convolutional layers, the node feature output of the final layer is obtained. This output is denoted as the spatial feature representation matrix. The spatial feature representation matrix consists of the deep feature representation of each node after multi-order neighborhood aggregation. Each row of the matrix corresponds to the semantic expression of a factory node in the spatial structure.

[0012] Optionally, the step of constructing a reinforcement learning controller based on the joint embedding representation matrix to optimize the improved spatiotemporal graph neural network model specifically includes: The joint embedding representation matrix is ​​input into the reinforcement learning controller, which uses the training process of the improved spatiotemporal graph neural network model as its operating environment and the internal policy network as its agent. In reinforcement learning controllers, the set of states is defined as a set of states that change over time, where each state is composed of three types of indicators: current estimation error, model response time, and model stability. In the reinforcement learning controller, an action set is defined. Each action in the action set corresponds to adjusting key parameters in the training process of the improved spatiotemporal graph neural network, specifically including adjusting the learning rate, adjusting the weights of the loss function, and adjusting the weights of the multimodal fusion. The action set serves as the basis for the policy network to select and execute actions during the training process. In the reinforcement learning controller, a reward function is set, which is weighted and calculated with estimation error, response time and model stability as input parameters to obtain the reward value; The policy network samples a predefined set of actions and determines the optimal action based on the feedback from the current state and the reward function. During training, state acquisition, action selection, reward calculation, and parameter updates are continuously performed to obtain an optimized and improved spatiotemporal graph neural network model.

[0013] Optionally, applying the optimized improved spatiotemporal graph neural network model to the valuation task, and outputting a set of valuation results, specifically includes: The optimized and improved spatiotemporal graph neural network model is used as the estimation model. Regression prediction operation is performed on the input joint embedding representation matrix. The joint embedding representation matrix is ​​formed by splicing and fusing multiple node features. Each row vector corresponds to the comprehensive feature of a factory node under the dual contexts of spatial dependence and temporal evolution. In the output layer of the valuation model, a regression prediction function is set, and a nonlinear mapping is performed on each joint embedding representation to convert the joint embedding representation into the corresponding predicted market value. The regression prediction function takes the joint embedding representation as input and completes the mapping calculation based on the trained parameter set. The final output value is used as the predicted market value of the target factory. The joint embedding representation of each node is processed using the quantile regression method to calculate the node's credibility score; The predicted market value and credibility score of each plant node are combined and output to form a set of valuation results.

[0014] Optionally, the mechanism for periodically or in real-time triggering of online updates to the improved spatiotemporal graph neural network model based on historical error feedback and market changes specifically includes: After the valuation results set is generated, the deviation between the predicted market value and the actual transaction value is continuously monitored, and the valuation error is calculated. The online update mechanism is triggered when the valuation error exceeds a preset threshold, or when external market data and policy announcement text are updated. Once the online update mechanism is triggered, the newly collected standardized dataset is merged with the existing historical dataset to form an incremental training dataset containing both historical and new samples. This incremental training dataset retains the support of historical data for model stability while incorporating the latest data to reflect market changes and policy shifts, providing a complete data foundation for subsequent transfer training. The improved spatiotemporal graph neural network model is transferred to the incremental training dataset. While keeping the original parameters generally stable, the parameter set is updated using a mini-batch iterative method. The loss function on the incremental dataset is calculated and the gradient is obtained. The parameters are then gradually adjusted in combination with the learning rate to obtain the updated parameter set. The updated parameter set is applied to the improved spatiotemporal graph neural network model to form an online updated valuation model.

[0015] The beneficial effects of this invention are: This invention introduces a multi-source heterogeneous data fusion mechanism at the data level to integrate land transfer, factory transaction, business registration, policy announcements, locational supporting facilities, and factory image information into a standardized dataset, achieving comprehensive coverage of factors influencing the market value of factories. Compared with traditional valuation methods that rely on a single data source, this method improves the completeness and consistency of input data, reduces prediction errors caused by missing information or data imbalance, and provides a solid data foundation for subsequent model building.

[0016] This invention proposes an improved spatiotemporal graph neural network in terms of model structure. It combines graph neural network branches and Transformer branches to model spatial and temporal features respectively, and aligns and deeply fuses these features through an interactive fusion layer. This dual-backbone structure can simultaneously capture the spatial relationships between the factory and surrounding facilities, as well as the temporal patterns of historical transactions and policy dynamics, achieving a unified representation of spatial and temporal features. This structural innovation overcomes the limitations of traditional model structures, which are singular and unable to simultaneously consider both spatial and temporal factors. This is the core creative aspect of this invention, enhancing the adaptability and predictive accuracy of valuation results in complex market environments.

[0017] At the optimization and application level, this invention introduces a reinforcement learning controller, which uses estimation error, response time and model stability as reward functions to dynamically adjust the learning rate, loss function weights and multimodal feature fusion ratio, so that the model can maintain adaptive adjustment when market environment and policy conditions change. At the same time, this invention sets up an online update and cross-regional transfer mechanism to ensure the generalization and real-time performance of the model when deployed in multiple regions through incremental training and transfer learning.

[0018] These mechanisms enable the valuation system to be continuously updated, dynamically adapted, and widely promoted, providing reliable technical support for risk control in financial institutions, real estate asset management, and decision-making in manufacturing enterprises. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0020] Figure 1 This is a flowchart of an automatic prediction method for factory market value based on deep learning proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved spatiotemporal graph neural network model in the automatic prediction method for factory market value based on deep learning proposed in this invention; Figure 3 This is a schematic diagram of the dynamic optimization mechanism of the reinforcement learning controller in the automatic prediction method for factory market value based on deep learning proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figures 1-3 A method for automatically predicting the market value of factory buildings based on deep learning includes the following steps: Collect heterogeneous data from multiple sources and perform data cleaning, missing value imputation, timestamp alignment and spatial coordinate unification to form a standardized dataset; Construct a spatial graph structure for factory plots, and assign weights to edges based on geographical distance, industry similarity, and policy overlap to generate a spatial adjacency matrix; A multimodal feature vector set is constructed based on a standardized dataset, and then normalized and aligned to form a set of node feature representations. By inputting the node feature representation set and the spatial adjacency matrix into the improved spatiotemporal graph neural network model, a joint embedding representation matrix is ​​obtained; Based on the joint embedding representation matrix, a reinforcement learning controller is constructed to optimize the improved spatiotemporal graph neural network model; The optimized and improved spatiotemporal graph neural network model is applied to the valuation task, and a set of valuation results is output. Based on historical error feedback and market changes, an improved spatiotemporal graph neural network model is updated online periodically or in real time.

[0023] In this embodiment, the process of collecting multi-source heterogeneous data and performing data cleaning, missing value imputation, timestamp alignment, and spatial coordinate unification to form a standardized dataset specifically includes: The data collected includes multi-source heterogeneous data such as land transfer data, factory transaction data, enterprise registration information, policy announcement text data, location and supporting facility data, and factory image data. The land transfer data includes fields such as plot number, transfer time, planned use, transaction price, and plot ratio; The factory building transaction data includes transaction time, area, number of floors, building year, transaction unit price, and transaction location code; The business registration information of the enterprise includes the enterprise name, industry classification, registered address and establishment date; The policy announcement text data includes the policy number, release date, applicable region, and corresponding text content; The location-related infrastructure data includes spatial distribution information of transportation facilities, industrial parks, schools, and service outlets around the site. The factory building image data includes images of the main facade of the factory building and remote sensing aerial images, with an image resolution of no less than 512×512 pixels; Data cleaning and processing were performed on land transfer data, factory transaction data, enterprise business registration information, policy announcement texts, location and supporting facilities data, and factory image data. Incomplete records were deleted and field formats were standardized. Missing values ​​in multi-source heterogeneous data are filled by using time series interpolation to fill in missing timestamps and by using nearest neighbor interpolation to fill in missing spatial attributes. Timestamp alignment is performed on time fields in multi-source heterogeneous data to ensure that time information in factory transaction data, land transfer data and policy announcement texts is unified into a standardized time series. Unify the spatial coordinates in multi-source heterogeneous data, convert the coordinate systems of data from different sources into a unified coordinate system, and establish the geographical location labeling of factory nodes; The land transfer data, factory transaction data, enterprise business information, policy announcement texts, location and supporting facilities data, and factory image data, which have been cleaned, filled, aligned and have unified coordinates, are formatted and integrated to generate a standardized dataset.

[0024] In this embodiment, the spatial map structure of the factory site is constructed, and weighted values ​​are set for edges based on geographical distance, industry similarity, and policy overlap to generate a spatial adjacency matrix. Specifically, this includes: Based on the spatial coordinate vector corresponding to each factory record in the standardized dataset, each factory plot is treated as a graph node, and a node set is constructed, where each node corresponds to a factory sample with a unique geographical location. The node set is used to represent the factory distribution structure in the entire valuation area. Based on the relationship between the factory nodes, determine whether to establish a connection edge in turn, and combine all the connection edges into an edge set. Specifically, if the Euclidean distance between two factory nodes in geographic space is less than the preset geographic proximity threshold, then establish a connection edge between them based on geographic proximity. If the industries to which the enterprises corresponding to the two factory nodes belong have upstream and downstream supply chain collaboration relationships, then establish a connection edge between the two to establish supply chain collaboration relationships. If two factory nodes are located in the same policy coverage area or are regulated by the same policy, then a connecting edge with a policy coverage relationship is established between them; For any two graph nodes that satisfy the connection condition, calculate the edge weight value. The edge weight value consists of three parts: geographical proximity, industrial chain similarity, and policy coverage overlap, resulting in a set of edge weight values: ; in, Represents a node With nodes The edge weights between them are used to construct the spatial adjacency matrix. The weighting coefficients representing the geographical distance component. Represents the normalized node With nodes Geographical proximity between them The weighting coefficients representing the similarity components of the industrial chain. Represents a node With nodes The higher the value, the stronger the industrial synergy between the corresponding enterprises in the industrial chain. The weighting coefficients representing the policy coverage component. Represents a node With nodes The degree of overlap in the policy areas where they are located; the larger the value, the higher the degree to which the two nodes are affected by the same policy. The set of nodes, the set of edges that meet the connection conditions, and the corresponding set of edge weights are combined to construct a spatial graph structure. Each node represents a factory plot, each edge represents the connection relationship between nodes under the conditions of geographical proximity, industrial chain collaboration, or policy coverage, and the edge weights are used to reflect the strength of the association between nodes. The spatial graph structure serves as the basis for expressing spatial relationships and supports the subsequent generation of the spatial adjacency matrix. Based on the spatial graph structure, the set of nodes, the set of edges, and the corresponding set of edge weights are combined to generate a spatial adjacency matrix. Each element in the spatial adjacency matrix is ​​used to represent the spatial connection relationship and strength between any two factory nodes. If there is an edge connection between the nodes, the edge weight is recorded at the corresponding position in the adjacency matrix. If there is no edge connection between the nodes, the value 0 is recorded at the corresponding position in the adjacency matrix.

[0025] In this embodiment, the step of constructing a multimodal feature vector set based on a standardized dataset, performing normalization and alignment processing, and forming a node feature representation set specifically includes: The structured attribute vector, policy text encoding vector, and image feature vector corresponding to each factory record are extracted from the standardized dataset, forming a multimodal feature vector set, where: The structured attribute vector consists of fields such as factory area, number of floors, building age, and land use, and is formed into a fixed-length numerical vector after numerical standardization. The policy text encoding vector obtains a contextual embedding representation that reflects the semantics of policy elements by inputting the policy announcement text corresponding to each record into a pre-trained language model; The image feature vector is obtained by inputting the factory image into a visual neural network model to extract image feature information, resulting in a fixed-dimensional image representation used to characterize the factory's appearance attributes. The structured attribute vector, policy text encoding vector, and image feature vector are normalized respectively. The structured attribute vector adopts the zero-mean unit variance standardization method to convert each numerical feature into a standard distribution with a mean of 0 and a standard deviation of 1. The policy text encoding vector and image feature vector adopt the vector norm normalization method to scale each vector to a unit length, so as to unify the numerical scale of the features of each modality and make the features of different modalities comparable in the subsequent fusion process. The normalized structured attribute vector, policy text encoding vector, and image feature vector are aligned in a predetermined order, and the three types of modal sub-vectors are combined into a unified node feature vector by vector concatenation. The total dimension of the node feature vector is equal to the sum of the dimensions of each modal sub-vector. The node feature vector serves as a joint representation of a single factory node in terms of structured, textual, and image information. All node feature representations are arranged sequentially according to node number to form a set of node feature representations.

[0026] In this embodiment, the step of inputting the node feature representation set and the spatial adjacency matrix into the improved spatiotemporal graph neural network model to obtain the joint embedding representation matrix specifically includes: The node feature representation set and the spatial adjacency matrix are simultaneously input into the improved spatiotemporal graph neural network model, which includes two parallel backbone structures: a graph neural network branch and a Transformer branch. In the graph neural network branch, multi-order graph convolution operations are performed on the feature representation of each node based on the spatial adjacency matrix to calculate the spatial feature matrix; In the Transformer branch, based on the historical transaction price records and historical policy text content corresponding to each factory node in the standardized dataset, a multimodal temporal feature sequence is constructed in chronological order. Each time step contains the transaction price features of the node at the corresponding time point and the semantic features of the corresponding policy text after encoding. All time steps are combined to form the temporal input tensor of the node. This temporal input tensor is used to characterize the price change trend and policy environment evolution characteristics of the node in the continuous time dimension. The temporal input tensor is input into the multi-layer Transformer encoder structure, and position encoding, multi-head self-attention calculation, feedforward neural network transformation and layer normalization operations are performed sequentially to extract the deep feature representation of each node in the time dimension, thereby obtaining the temporal feature representation corresponding to each node. The temporal feature representations of all nodes are arranged according to the node number to form a temporal feature matrix. Each row of the temporal feature matrix corresponds to the temporal feature representation of a node under the historical transaction information and policy evolution sequence. The spatial feature matrix and the temporal feature matrix are simultaneously input into the interactive fusion layer. First, the feature dimensions of the two are uniformly mapped so that the spatial feature representation and the temporal feature representation can be operated in the same dimensional space. Then, based on the dimensional alignment, a feature concatenation operation is performed to combine the spatial feature vector and temporal feature vector corresponding to each node according to the feature dimension direction to form a node-level fused feature representation. After the splicing is completed, the fused feature representation is input into the multilayer perceptron structure. Through linear transformation, nonlinear activation and weight learning, the deep extraction of the fused features is achieved to obtain the joint embedding representation matrix. Each row of the joint embedding representation matrix corresponds to the comprehensive feature of a factory node under the dual contexts of spatial dependence and temporal evolution.

[0027] In this embodiment, within the graph neural network branch, multi-order graph convolution operations are performed on the feature representation of each node based on the spatial adjacency matrix. The computation of the spatial feature representation specifically includes: The connection relationships between all nodes are processed based on the spatial adjacency matrix. First, an identity matrix is ​​introduced into the original adjacency matrix to add self-loop connections. Then, the degree value of each node is calculated to form a degree matrix. The adjacency matrix after introducing self-loops is normalized using a symmetric normalization formula to obtain a normalized adjacency matrix. The normalized adjacency matrix is ​​used to aggregate the neighborhood information of nodes according to the adjacency strength during graph convolution. Set the set of node feature representations as the initial feature input, denoted as the initial node feature matrix. Each row of the initial node feature matrix corresponds to the multimodal feature representation of a factory node. Using the initial node feature matrix and normalized adjacency matrix as input, multi-layer graph convolution operations are performed sequentially. In each layer, a new node feature matrix is ​​calculated through neighborhood feature aggregation and weight transformation. Specifically, the node feature matrix of the current layer is multiplied by the normalized adjacency matrix to achieve neighborhood information fusion, and then multiplied by the trainable weight matrix of the current layer. After transformation by a nonlinear activation function, the output feature matrix of the layer is obtained. This output is used as the input of the next layer, and the above process is repeated until a preset number of graph convolution layers are completed, finally obtaining the node spatial feature representation under multi-layer perception. After completing the calculation of all graph convolutional layers, the node feature output of the final layer is obtained. This output is denoted as the spatial feature representation matrix. The spatial feature representation matrix consists of the deep feature representation of each node after multi-order neighborhood aggregation. Each row of the matrix corresponds to the semantic expression of a factory node in the spatial structure.

[0028] In this embodiment, the step of constructing a reinforcement learning controller based on the joint embedding representation matrix to optimize the improved spatiotemporal graph neural network model specifically includes: The joint embedding representation matrix is ​​input into the reinforcement learning controller, which uses the training process of the improved spatiotemporal graph neural network model as its operating environment and the internal policy network as its agent. In a reinforcement learning controller, the set of states is defined as a group of states that change over time. Each state is composed of three types of indicators: current valuation error, model response time, and model stability. The valuation error is used to characterize the degree of deviation between the prediction result and the actual transaction price. The model response time is used to characterize the time required to complete a single valuation calculation. The model stability is used to characterize the convergence performance and result fluctuation of the model during continuous training iterations. The above three types of indicators together constitute the state description at any time step, which serves as the input basis for the reinforcement learning controller to select actions and optimize strategies. In the reinforcement learning controller, an action set is defined. Each action in the action set corresponds to adjusting key parameters in the training process of the improved spatiotemporal graph neural network, specifically including adjusting the learning rate, adjusting the weights of the loss function, and adjusting the weights of the multimodal fusion. The action set serves as the basis for the policy network to select and execute actions during the training process. A reward function is set in the reinforcement learning controller to generate feedback signals based on the model's performance during training. The reward function is calculated by weighting the estimation error, response time, and model stability as input parameters. The estimation error measures the deviation between the predicted result and the actual result, the response time measures the computation time of the model in a single estimation task, and the model stability measures the convergence consistency of the model in multiple training iterations. The three factors are multiplied by their respective weight coefficients and combined to form the reward value. The reward value serves as the basis for the policy network to select and update actions, thereby optimizing the training process of the improved spatiotemporal graph neural network. The policy network samples a predefined set of actions, determines the optimal action based on the feedback from the current state and the reward function, and applies the optimal action to the training process of the improved spatiotemporal graph neural network to dynamically adjust key parameters, including the learning rate, loss function weights, and multimodal fusion weights, thereby achieving adaptive updates to the model training strategy. During training, state acquisition, action selection, reward calculation, and parameter updates are continuously performed. At each time step, the reinforcement learning controller acquires state information consisting of the current estimation error, response time, and model stability. Based on the state information, the policy network selects the corresponding optimization action and adjusts the learning rate, loss function weights, and multimodal fusion weights. Then, the reward value is calculated based on the adjusted model performance and used to update the parameters of the policy network. Through multiple iterations, the effectiveness of the policy is gradually improved until the training process converges, and finally, the optimized improved spatiotemporal graph neural network model is obtained.

[0029] In this embodiment, applying the optimized and improved spatiotemporal graph neural network model to the estimation task and outputting a set of estimation results specifically includes: The optimized and improved spatiotemporal graph neural network model is used as the estimation model. Regression prediction operation is performed on the input joint embedding representation matrix. The joint embedding representation matrix is ​​formed by splicing and fusing multiple node features. Each row vector corresponds to the comprehensive feature of a factory node under the dual contexts of spatial dependence and temporal evolution. In the output layer of the valuation model, a regression prediction function is set, and a nonlinear mapping is performed on each joint embedding representation to convert the joint embedding representation into the corresponding predicted market value. The regression prediction function takes the joint embedding representation as input and completes the mapping calculation based on the trained parameter set. The final output value is used as the predicted market value of the target factory. The joint embedding representation of each node is processed using the quantile regression method, and the confidence score of the node is calculated. The confidence score is used to measure the uncertainty level of the prediction result, and the value ranges from zero to one. The closer the value is to one, the higher the confidence of the prediction result. The predicted market value and credibility score of each plant node are combined and output to form a set of valuation results.

[0030] In this embodiment, the mechanism for periodically or in real-time triggering of the online update mechanism for the improved spatiotemporal graph neural network model based on historical error feedback and market changes specifically includes: After the valuation result set is generated, the deviation between the predicted market value and the actual transaction value is continuously monitored, and the valuation error is calculated. The valuation error is then used to construct the online update trigger conditions. The online update mechanism is triggered when the valuation error exceeds a preset threshold, or when external market data and policy announcement text are updated. Once the online update mechanism is triggered, the newly collected standardized dataset is merged with the existing historical dataset to form an incremental training dataset containing both historical and new samples. This incremental training dataset retains the support of historical data for model stability while incorporating the latest data to reflect market changes and policy shifts, providing a complete data foundation for subsequent transfer training. The improved spatiotemporal graph neural network model is transferred to the incremental training dataset. While keeping the original parameters generally stable, the parameter set is updated using a mini-batch iterative method. The loss function on the incremental dataset is calculated and the gradient is obtained. The parameters are then gradually adjusted in combination with the learning rate to obtain the updated parameter set. The updated parameter set is applied to the improved spatiotemporal graph neural network model to form an online updated valuation model, which is used to replace the original model to continue performing the valuation task. When deployed in multiple regions, the online updated valuation model is migrated to the target region, and the standardized dataset within the region is retrained or fine-tuned to generate a region-specific model copy to support the task of predicting the market value of factory buildings across cities and regions. The online updated valuation model has the ability to dynamically adapt to market and policy changes, thereby improving the accuracy and stability of the prediction results.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to the valuation of factory buildings in a large industrial park. This park has seen active trading in the past two years, with factory types including standard factory buildings, industrial park buildings, and warehousing and logistics facilities. Financial institutions have high requirements for valuation accuracy in mortgage loan approvals, while real estate developers and manufacturing companies also need to quickly grasp the value of factory buildings for asset management and expansion decisions. Existing valuation methods rely on human experience and rule-based calculations, with an average valuation deviation rate exceeding 15%, and a single valuation taking between 10 and 20 minutes, making it difficult to meet the needs of high-frequency trading and dynamic markets.

[0032] In this scenario, the invention collects and integrates land transfer data, factory transaction records, business registration information, industrial policy announcements, locational supporting facilities, and factory images. After cleaning, missing value imputation, timestamp alignment, and spatial coordinate unification, a standardized dataset is formed. A spatial graph structure is established based on the location of factory plots and industrial relationships, and weighted edges are generated according to geographical distance, industrial chain synergy, and policy coverage to obtain a spatial adjacency matrix. Subsequently, spatial features are extracted using graph neural network branches, and time-series features of transactions and policies are modeled using Transformer branches. Dual-branch alignment and deep fusion are achieved in the interactive fusion layer to generate a joint embedding representation. During training, a reinforcement learning controller is introduced, using valuation error, response time, and stability as reward functions to dynamically adjust the learning rate and loss weights, ensuring that the model can adapt to market and policy changes in real time. Finally, the optimized model is applied to factory valuation, quickly outputting market value predictions and credibility scores, and achieving adaptive transfer when new data arrives through an online update mechanism.

[0033] To verify the effectiveness, 300 historical transaction cases in the park were selected as a sample, with transaction prices ranging from 12 million to 85 million yuan, covering different areas, floors, and land types. Experimental results show that the average deviation rate between the valuation results of this invention and the actual transaction price is 4.7%, which is better than the 15.2% of the traditional method; the average time for valuation per case is 2.8 minutes, improving processing efficiency by approximately 5 times; the correlation coefficient between the predicted results and the actual transaction price reaches 0.93, while the traditional method is only 0.76. Further verification was conducted within two weeks after the policy adjustment. The valuation deviation rate of this invention was consistently controlled within 5%, while the deviation rate of the static model rose to over 12%, demonstrating the adaptive advantage of this invention in a dynamic market environment.

[0034] Table 1. Comparative Experiment Results of Factory Market Value Valuation ; As shown in Table 1 above, this invention demonstrates improvements over traditional methods in key performance indicators for predicting the market value of factory buildings. Firstly, regarding valuation accuracy, the average valuation deviation rate of this invention is controlled at 4.7%, significantly better than the 15.2% of the traditional method, with an overall error reduction of 10.5 percentage points. This result indicates that this invention, through the joint modeling of multi-source heterogeneous data fusion and an improved spatiotemporal graph neural network, can more comprehensively capture the spatial, temporal, and policy characteristics affecting factory building value, thereby effectively improving the accuracy of the valuation results. Secondly, in terms of valuation efficiency, the traditional method takes an average of 15 minutes per valuation, while the invention takes only 2.8 minutes per valuation, representing an efficiency improvement of approximately 5 times.

[0035] This advantage demonstrates the invention's rapid response capability under multimodal feature parallel processing and GPU acceleration, meeting the business needs of large-scale, high-frequency valuation. Furthermore, regarding the correlation between valuation results and actual transaction prices, the correlation coefficient of this invention reaches 0.93, higher than the 0.76 of the traditional method, indicating that its prediction results are more consistent with actual market trends. Finally, in terms of system stability, this invention scores 96 points, while the traditional method only scores 81 points, a difference of 15 points, demonstrating the robustness and reliability of this invention in long-term operation.

[0036] Table 2 Comparison of model performance before and after policy adjustment ; As can be seen from Table 2 above, the present invention demonstrates improved stability and adaptability in predicting the market value of factory buildings compared to the traditional static model before and after policy adjustments. Firstly, under unchanged policy conditions, the average valuation deviation rate of the traditional static model is 14.8%, while the deviation rate of the present invention's model is only 4.9%, a reduction of 9.9 percentage points in overall error. This demonstrates that the present invention can maintain high-precision valuation capabilities even under normal market conditions.

[0037] Secondly, in the week following the policy adjustment, market volatility intensified, and the bias rate of the traditional static model remained high at 12.1%, while the bias rate of the model of this invention remained at around 5.0%, a reduction of 7.1 percentage points. This demonstrates that the present invention, through its reinforcement learning controller and online update mechanism, can quickly adapt to the new policy environment in a short period of time. Furthermore, in the two and three weeks after the policy implementation stabilized, the bias rates of the traditional static model were 11.3% and 11.0%, respectively, while those of the model of this invention were 4.8% and 4.6%, respectively, with the overall error consistently remaining below 5%, exhibiting long-term stability.

[0038] This result demonstrates that, through dynamic optimization and transfer training, the present invention enables the model to maintain high valuation accuracy and stability even when markets and policies change, solving the problem of declining prediction performance of traditional static models in dynamic environments and providing financial institutions and industrial enterprises with a more reliable valuation tool.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatically predicting the market value of factory buildings based on deep learning, characterized in that, Includes the following steps: Collect heterogeneous data from multiple sources and perform data cleaning, missing value imputation, timestamp alignment and spatial coordinate unification to form a standardized dataset; Construct a spatial graph structure for factory plots, and assign weights to edges based on geographical distance, industry similarity, and policy overlap to generate a spatial adjacency matrix; A multimodal feature vector set is constructed based on a standardized dataset, and then normalized and aligned to form a set of node feature representations. By inputting the node feature representation set and the spatial adjacency matrix into the improved spatiotemporal graph neural network model, a joint embedding representation matrix is ​​obtained; Based on the joint embedding representation matrix, a reinforcement learning controller is constructed to optimize the improved spatiotemporal graph neural network model; The optimized and improved spatiotemporal graph neural network model is applied to the valuation task, and a set of valuation results is output. Based on historical error feedback and market changes, an improved spatiotemporal graph neural network model is updated online periodically or in real time.

2. The method for automatically predicting the market value of factory buildings based on deep learning according to claim 1, characterized in that, The standardized dataset includes structured attribute vectors, policy text encoding vectors, factory image feature vectors, unified timestamp fields, and geospatial coordinate vectors.

3. The method for automatically predicting the market value of factory buildings based on deep learning according to claim 1, characterized in that, Constructing a spatial map structure for factory site plots, and assigning weighted values ​​to edges based on geographical distance, industry similarity, and policy overlap, generates a spatial adjacency matrix, specifically including: Based on the spatial coordinate vector corresponding to each record in the standardized dataset, a set of nodes is constructed with each factory plot as a graph node; Based on the relationship between the factory nodes, determine whether to establish a connecting edge in turn, and combine all connecting edges into an edge set; For any two graph nodes that satisfy the connection conditions, calculate the weight value of the edge. The weight value of the edge consists of three parts: geographical proximity, industrial chain similarity, and policy coverage overlap, to obtain the edge weight set. Combine the set of nodes, the set of edges that satisfy the connection conditions, and the corresponding set of edge weights to construct a spatial graph structure; Based on the spatial graph structure, the set of nodes, the set of edges, and the corresponding set of edge weights are combined to generate a spatial adjacency matrix.

4. The method for automatically predicting the market value of factory buildings based on deep learning according to claim 1, characterized in that, The process of constructing a multimodal feature vector set based on a standardized dataset, performing normalization and alignment processing, and forming a node feature representation set specifically includes: The structured attribute vector, policy text encoding vector, and image feature vector corresponding to each factory record are extracted from the standardized dataset to obtain a multimodal feature vector set; The structured attribute vector, policy text encoding vector, and image feature vector are normalized respectively; The normalized structured attribute vector, policy text encoding vector, and image feature vector are dimensionally aligned and connected to form node feature representations; All node feature representations are arranged sequentially according to node number to form a set of node feature representations.

5. The method for automatically predicting the market value of factory buildings based on deep learning according to claim 1, characterized in that, The step of inputting the node feature representation set and the spatial adjacency matrix into the improved spatiotemporal graph neural network model to obtain the joint embedding representation matrix specifically includes: The node feature representation set and the spatial adjacency matrix are simultaneously input into the improved spatiotemporal graph neural network model, which includes two parallel backbone structures: a graph neural network branch and a Transformer branch. In the graph neural network branch, multi-order graph convolution operations are performed on the feature representation of each node based on the spatial adjacency matrix to calculate the spatial feature matrix; In the Transformer branch, a time-series input tensor is constructed based on the historical transaction price sequence and historical policy text encoding sequence corresponding to each node in the standardized dataset. The temporal input tensor is input into the multi-layer Transformer encoder structure, and position encoding, multi-head self-attention calculation, feedforward neural network transformation and layer normalization operations are performed sequentially to obtain the temporal dependency feature representation of each node. The temporal dependency feature representations of all nodes are arranged according to the node number to form a temporal feature matrix. The spatial feature matrix output by the graph neural network branch and the temporal feature matrix output by the Transformer branch are aligned and jointly fused in the interactive fusion layer, and then combined into a joint embedding representation matrix by splicing.

6. The method for automatically predicting the market value of factory buildings based on deep learning according to claim 5, characterized in that, In the graph neural network branch, multi-order graph convolution operations are performed on the feature representation of each node based on the spatial adjacency matrix. The computation of the spatial feature matrix specifically includes: The connection relationships between all nodes are processed based on the spatial adjacency matrix to obtain the normalized adjacency matrix; Set the set of node feature representations as the initial feature input, denoted as the initial node feature matrix. Each row of the initial node feature matrix corresponds to the multimodal feature representation of a factory node. Using the initial node feature matrix and the normalized adjacency matrix as input, multi-layer graph convolution operations are performed sequentially. In each layer, a new node feature matrix is ​​calculated through neighborhood feature aggregation and weight transformation. After completing the calculation of all graph convolutional layers, the node feature matrix of the final layer is obtained, denoted as the spatial feature matrix.

7. The method for automatically predicting the market value of factory buildings based on deep learning according to claim 1, characterized in that, The optimization of the improved spatiotemporal graph neural network model by constructing a reinforcement learning controller based on the joint embedding representation matrix specifically includes: The joint embedding representation matrix is ​​input into the reinforcement learning controller, which uses the training process of the improved spatiotemporal graph neural network model as the environment and the policy network as the agent. In a reinforcement learning controller, a set of states is defined, where each state consists of three indices: current estimation error, response time, and model stability. In the reinforcement learning controller, a set of actions is defined, where each action includes an adjustment operation on the learning rate, loss function weights, and multimodal fusion weights; In the reinforcement learning controller, a reward function is set, which is weighted and calculated with estimation error, response time and model stability as input parameters to obtain the reward value; The action set is sampled through a policy network, and the optimal action is determined based on the feedback from the current state and the reward function. During training, state acquisition, action selection, reward calculation, and parameter updates are continuously performed to obtain an optimized and improved spatiotemporal graph neural network model.

8. The method for automatically predicting the market value of factory buildings based on deep learning according to claim 1, characterized in that, The process of applying the optimized and improved spatiotemporal graph neural network model to the valuation task, and outputting a set of valuation results, specifically includes: The optimized and improved spatiotemporal graph neural network model is used as the estimation model to perform regression prediction operations on the input joint embedding representation matrix; In the output layer of the valuation model, a regression prediction function is set, and a nonlinear mapping is performed on each joint embedding representation to obtain the predicted market value. The joint embedding representation of each node is processed using the quantile regression method to calculate the node's credibility score; The predicted market value and credibility score of each plant node are combined and output to form a set of valuation results.

9. The method for automatically predicting the market value of factory buildings based on deep learning according to claim 1, characterized in that, The mechanism for periodically or in real-time triggering online updates of the improved spatiotemporal graph neural network model based on historical error feedback and market changes specifically includes: After the valuation results set is generated, the deviation between the predicted market value and the actual transaction value is continuously monitored, and the valuation error is calculated. The online update mechanism is triggered when the valuation error exceeds a preset threshold, or when external market data and policy announcement text are updated. Once the online update mechanism is triggered, the newly collected standardized dataset will be merged with the existing historical dataset to form an incremental training dataset. The improved spatiotemporal graph neural network model was transferred to an incremental training dataset for training. The parameter set of the improved spatiotemporal graph neural network model was updated through mini-batch iterative updates to obtain the updated parameter set. The updated parameter set is applied to the improved spatiotemporal graph neural network model to form an online updated valuation model.

Images