A target object multi-gradient prediction method fusing multi-source data
By constructing a spatiotemporal attention mechanism and reinforcement learning, and integrating features from multiple data sources, the stability and reliability issues of multi-gradient prediction were resolved, enabling accurate prediction of mining resource development and engineering assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 中国矿业报社
- Filing Date
- 2026-04-14
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies struggle to effectively integrate multi-source heterogeneous data, lack the ability to hierarchically represent multi-gradient features, and suffer from insufficient stability and reliability in prediction results, making it difficult to meet the scientific decision-making needs of mining resource development and engineering assessment.
By constructing a spatiotemporal attention mechanism, combining temporal and spatial features, extracting multi-source data features using bidirectional long short-term memory networks and graph convolutional networks, and adjusting model weights using reinforcement learning, multi-gradient prediction is achieved.
It improves the comprehensiveness and accuracy of multi-source data fusion, can adapt to policy adjustments or emergencies in real time, provides confidence intervals for fluctuations in target objects, and enhances the stability and reliability of predictions.
Smart Images
Figure CN122390136A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multi-source data analysis technology, and specifically to a multi-gradient prediction method for target objects that integrates multi-source data. Background Technology
[0002] Accurate forecasting of target objects, such as the sales of mining products, is a core element for scientific decision-making in subsequent resource development and engineering assessments. The results of accurate forecasting directly impact project classification, resource allocation, and implementation efficiency. Currently, target object forecasting increasingly relies on multi-source data, including the mine's geographical location and transportation information, dynamic data on mine production, and textual reports. Different types of data have their advantages; for example, spatial data can characterize the distribution of mine locations, time-series data can reflect the dynamic trends of target objects, and textual data can supplement implicit patterns and constraints, providing a foundation for improving the comprehensiveness of forecasts.
[0003] However, existing prediction methods still face significant technical bottlenecks in practical applications. First, multi-source data exhibit heterogeneity (structural, image, text, etc.), making it difficult for traditional fusion methods to uncover the intrinsic relationships between data, which can easily lead to information redundancy or loss of key features, affecting the reliability of prediction inputs. Second, the attributes of target objects often exhibit multi-gradient features, while existing methods mostly focus on a single scale or single indicator, lacking the ability to hierarchically represent multi-gradient features, making it difficult to adapt to the refined needs of different decision-making scenarios. Third, due to the sparsity of data, observation errors, and the complexity of the target objects themselves, existing methods have weak tolerance for uncertainty, resulting in insufficient stability and credibility of prediction results, which can easily lead to decision bias.
[0004] Therefore, in response to the pain points of difficulty in multi-source data fusion, insufficient multi-gradient feature representation, and limited prediction reliability, there is an urgent need for a method that can effectively integrate multi-source heterogeneous data and achieve accurate multi-gradient prediction of target objects, so as to provide technical support for scientific decision-making in subsequent resource development, engineering evaluation and other projects of the mine. Summary of the Invention
[0005] The purpose of this invention is to provide a multi-gradient prediction method for target objects that integrates multi-source data, in order to solve the technical problem that traditional methods in the prior art are unable to integrate unstructured data such as policy texts and satellite remote sensing data with heterogeneous types. Therefore, traditional methods are unable to meet the requirements due to defects such as spatiotemporal separation and single data.
[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:
[0007] A method for multi-gradient prediction of target objects by fusing multi-source data includes the following steps:
[0008] Step 100: Extract evaluation indicators related to the prediction of the target object, form a change sequence of the target object, and construct a data set of evaluation indicators and change sequence of the target object on the time axis;
[0009] Step 200: Establish a target object prediction model based on the dataset and spatiotemporal attention mechanism;
[0010] Step 300: Capture the evaluation indicators at the current moment and input the evaluation indicators at the current moment into the target object prediction model to predict the target object. The prediction of the target object includes short-term expectations and medium- and long-term predictions.
[0011] Step 400: Input the real-time evaluation indicators and target objects into the data set according to the time axis to update the data set, thereby correcting the target object prediction model. Iterate continuously according to the time axis to update the short-term expectations and medium-to-long-term predictions of the target objects.
[0012] As a preferred embodiment of the present invention, in step 100, the method for constructing a data set of evaluation indicators and target object change sequences on the time axis is as follows:
[0013] Capture the time points when the value of the target object changes, and obtain the evaluation index corresponding to each time the value of the target object changes.
[0014] The evaluation indicators include time series data, spatial data, and text data. The time series data includes the output volume and inventory data of the target object; the spatial data includes the geographical location of the target object's transactions and transportation network information; and the text data includes audiovisual materials containing the target object.
[0015] As a preferred embodiment of the present invention, the method for establishing the target object prediction model in step 200 is as follows:
[0016] Data preprocessing: Data processing is performed on the evaluation indicators and target objects in the dataset, including data cleaning, normalization, and text feature extraction for the evaluation indicators in the dataset;
[0017] Constructing a spatiotemporal graph: Determining the weights of spatiotemporal graph nodes and edges, wherein spatiotemporal graph nodes include the geographical locations where the target object appears, connecting different spatiotemporal graph nodes in pairs to form spatiotemporal graph edges to construct the spatiotemporal graph, and determining the weight of each spatiotemporal graph edge;
[0018] Construct a target object prediction model: Use the time feature calculation module to extract time features from the time series data;
[0019] The spatial feature calculation module is used to process the spatiotemporal map and extract spatial features;
[0020] By combining temporal and spatial attention using a spatiotemporal attention calculation module, the new features processed by the spatiotemporal attention calculation module are input into a fully connected layer, and the final target object prediction value is obtained through linear transformation.
[0021] As a preferred embodiment of the present invention, the specific implementation steps for data cleaning, normalization processing, and text feature extraction of the evaluation indicators in the dataset are as follows:
[0022] Remove missing and outlier values from the evaluation metrics and target objects in the dataset;
[0023] The time series data and spatial data in the evaluation index are normalized to the interval [0, 1] using the Min-Max method;
[0024] For text data, text features are extracted by using a pre-trained language model to convert the text data into numerical vectors.
[0025] As a preferred embodiment of the present invention, the weight of each spatiotemporal graph edge is dynamically calculated based on the historical numerical correlation between different geographical locations containing the target object, and the weight of the spatiotemporal graph edge formed by two spatiotemporal graph nodes is determined by the Pearson correlation coefficient of two numerical time series containing transactions of the target object. The specific calculation method is as follows:
[0026] ;
[0027] In the formula, xi and yi are the values of two geographical locations containing the target object at the i-th time point, respectively; and These are the numerical averages of two geographical locations containing the target object; n is the length of the time series.
[0028] Based on the Pearson correlation coefficient, the weights of the edges in the spatiotemporal graph are calculated using the following formula:
[0029] ;
[0030] The spatiotemporal graph is constructed by using the weights of the edges formed by connecting the two spatiotemporal graph nodes containing the geographic locations of the target objects.
[0031] As a preferred embodiment of the present invention, when extracting time features from the time series data using the time feature calculation module, a bidirectional long short-term memory network is specifically used to extract the time features from the time series data. The specific implementation method is as follows:
[0032] The output quantity of the target object, inventory data, and audiovisual materials containing the target object are input into the bidirectional long short-term memory network. Specifically, the input data is a three-dimensional tensor containing the number of samples, time step, and number of features.
[0033] The bidirectional long short-term memory network outputs extracted features, specifically a three-dimensional tensor containing the number of samples, time step, and number of hidden layer neurons.
[0034] As a preferred embodiment of the present invention, the spatial feature calculation module uses a graph convolutional network to process the spatiotemporal graph and extract spatial features, wherein the input data of the graph convolutional network is the node features and adjacency matrix of the spatiotemporal graph, and the output data of the graph convolutional network is the spatial features extracted from the spatiotemporal graph;
[0035] The node features are specifically a two-dimensional tensor containing the number of nodes and node features. The node features include the annual output corresponding to the geographical location of the target object for each spatiotemporal graph node, transportation network information, and audiovisual materials containing the target object.
[0036] The adjacency matrix is a two-dimensional tensor containing information on the number and weight of nodes. Specifically, the adjacency matrix contains the connection relationships between multiple spatiotemporal graph nodes and the weight of the spatiotemporal graph edge formed by every two spatiotemporal graph nodes.
[0037] The output data of the graph convolutional network is specifically a two-dimensional tensor that includes the number of nodes and the output feature dimension.
[0038] As a preferred embodiment of the present invention, when the spatiotemporal attention calculation module combines temporal attention and spatial attention, the output data of the temporal feature extraction module and the output data of the spatial feature extraction module are concatenated to form a new feature, which is then input into the spatiotemporal attention calculation module, and the feature weighted by the spatiotemporal attention mechanism is output. Specifically, this includes the following steps:
[0039] The importance of the same geographical location containing the target object at different time steps is calculated using the time attention branch. The specific calculation formula is as follows:
[0040] ;
[0041] In the formula, The time attention branch weights are represented by Qt, which is the time query vector, indicating the feature information of the output and inventory of the same geographical location containing the target object at the current time step t, with dimension d. qWq is the temporal attention weight matrix with dimensions dq×dk; Kti is the key vector at the i-th time step, representing the vector obtained after linear transformation of the feature information at time step i, with dimensions dk; dk is the dimension of the key vector.
[0042] T is the transpose symbol, which is the transpose of the Wq matrix;
[0043] Softmax() is a general function that maps a set of real numbers to a probability distribution such that the sum of all elements is 1;
[0044] The spatial attention branch is used to learn the interactions between different geographical locations containing the target object. The specific calculation formula is as follows:
[0045] ;
[0046] In the formula, dv represents the spatial attention branch weights; Vi and Vj are the spatial value vectors of the i-th and j-th geographic locations containing the target object, respectively, representing the feature information of different geographic locations, with a dimension of dv; Wv is the spatial attention weight matrix, used to learn the interaction relationship between different geographic locations, transforming the spatial value vectors of different geographic locations to calculate their correlation, with a dimension of dv×dv; dv is the dimension of the spatial attention vector;
[0047] By combining temporal attention and spatial attention, the final spatiotemporal attention weights are obtained. It is used for weighted aggregation of temporal and spatial features, and the specific calculation formula is as follows:
[0048] .
[0049] As a preferred embodiment of the present invention, the bidirectional long short-term memory network in the temporal feature calculation module, the graph convolutional network in the spatial feature calculation module, the weight matrix of the spatiotemporal attention calculation module, and the fully connected layer are trained, wherein the training parameters are specifically the bias terms in each neural network layer.
[0050] In a preferred embodiment of the present invention, in step 400, an evaluation index related to the target object to be predicted at the current time is obtained, the evaluation index is input into the target object prediction model, the target object prediction result at the current time is obtained, and a 95% confidence interval for the time step is obtained, and the confidence interval is used as the predicted value interval of the target object.
[0051] Compared with the prior art, the present invention has the following advantages:
[0052] (1) The spatiotemporal attention mechanism provided by this invention can not only capture time information such as target objects and output information in multiple time periods, but also capture spatial information such as different production geographical locations and different transaction geographical locations, realize the linkage of time and space information, integrate multiple data types such as historical values, output, inventory, production geographical location, transaction geographical location, transportation network, and policy text, improve the comprehensiveness and accuracy of the model, and also improve the modeling accuracy, making the predicted target objects more meaningful.
[0053] (2) The method provided by the present invention uses reinforcement learning to adjust the weights of the spatiotemporal attention model, which can adapt to the impact of policy adjustments or sudden events in real time, and the prediction results are interval data, providing a confidence interval for the fluctuation of the target object, helping users to better manage risks. Attached Figure Description
[0054] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.
[0055] Figure 1 A flowchart illustrating the short-term target object prediction method provided in an embodiment of the present invention;
[0056] Figure 2 This is a schematic diagram illustrating the construction process of the target object prediction model provided in an embodiment of the present invention. Detailed Implementation
[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0058] Example 1
[0059] like Figure 1 As shown, this invention provides a multi-gradient prediction method for target objects that integrates multi-source data. The multi-source data includes structured data such as the target object's output values and inventory levels, unstructured data such as audiovisual materials containing the target object, and geospatial data such as the target object's geographical location distribution and transportation network. By integrating the above multi-source data, the modeling accuracy of the prediction model is significantly improved. The method includes the following steps:
[0060] Step 100: Extract evaluation indicators related to the prediction of the target object, form a change sequence of the target object, and construct a data set of evaluation indicators and change sequence of the target object on the time axis;
[0061] Step 200: Establish a target object prediction model based on the dataset and spatiotemporal attention mechanism;
[0062] Step 300: Capture the evaluation indicators at the current moment and input the evaluation indicators at the current moment into the target object prediction model to predict the target object. The prediction of the target object includes short-term expectations and medium- and long-term predictions.
[0063] Step 400: Input the real-time evaluation indicators and target objects into the dataset according to the time axis to update the dataset, thereby correcting the target object prediction model. Iterate continuously according to the dataset to update the short-term and medium-to-long-term predictions of the target objects.
[0064] For example, in the constructed evaluation index and target object change sequence data set, the target object change sequence is the target object collected every 15 minutes. Through the above process, only one point can be predicted each time (that is, the next 15 minutes). The prediction result is put into the data set, the data set is updated and then another point is predicted (that is, a point in the next 30 minutes). By continuously looping, the short-term and medium-to-long-term values of the target object can be predicted.
[0065] The spatiotemporal attention mechanism provided in this embodiment can not only capture temporal information such as target objects and output information over multiple time periods, but also capture spatial information containing target objects, thereby achieving the linkage of time and spatial information, improving modeling accuracy, and making the predicted target objects more meaningful. Furthermore, it integrates various data types such as historical values, outputs, inventory, geographical locations, transportation networks, and policy texts of target objects, enhancing the comprehensiveness and accuracy of the model.
[0066] In step 100, the method for constructing a dataset of evaluation indicators and target object change sequences on the time axis is as follows:
[0067] It captures the time points when the value of the target object changes, and obtains the evaluation index corresponding to each time the value of the target object changes. The target object can be a target object from the past five years or longer, and its corresponding evaluation index.
[0068] The evaluation indicators include time series data, spatial data, and text data. Time series data includes the output volume and inventory data of the target object; spatial data includes the geographical location of the target object's transactions and transportation network information; and text data includes audiovisual materials containing the target object, such as news events.
[0069] In step 200, as Figure 2 As shown, the implementation method for establishing the target object prediction model is as follows:
[0070] (1) Data preprocessing: Data processing is performed on the evaluation indicators and target objects in the dataset. Data cleaning, normalization and text feature extraction are performed on the evaluation indicators in the dataset.
[0071] The specific steps for data cleaning, normalization, and text feature extraction of evaluation metrics in the dataset are as follows:
[0072] Remove missing and outliers from the evaluation metrics and target objects in the dataset;
[0073] The Min-Max method was used to normalize the time series and spatial data in the evaluation indicators to the interval [0, 1].
[0074] For text data, text features are extracted by using a pre-trained language model to convert the text data into numerical vectors.
[0075] (2) Constructing the spatiotemporal graph: Determine the weights of the spatiotemporal graph nodes and edges. The spatiotemporal graph nodes include the geographical locations where the target object appears. Connect different spatiotemporal graph nodes in pairs to form spatiotemporal graph edges to construct the spatiotemporal graph. Determine the weight of each spatiotemporal graph edge. It is important to note that:
[0076] The spacetime graph is a hypothetical graph, consisting of a set of interconnected spacetime graph nodes (connected in pairs), which may or may not intersect. For example:
[0077] For example, nodes in the spatiotemporal graph represent the geographical locations of target objects, such as the London Metal Exchange and the Shanghai Futures Exchange, and geographical locations such as the Greenbushes lithium mine and the Atacama Salt Flat.
[0078] The connection between the London Stock Exchange and the Shanghai Futures Exchange indicates a certain correlation between the geographical locations of the two trading targets, and the degree of correlation is reflected by the edges (weights).
[0079] The London Stock Exchange and the Greenbushes lithium mine node are connected, indicating a relationship between the geographical location of the production target and the geographical location of the trading target. An increase or decrease in Greenbushes lithium production directly affects the London Stock Exchange's figures.
[0080] The weight of each spatiotemporal graph edge is dynamically calculated based on the historical numerical correlation between different geographical locations containing the target object. The weight of the spatiotemporal graph edge formed by two spatiotemporal graph nodes is determined using the Pearson correlation coefficient between two time series containing transactions of the target object. The specific calculation method is as follows:
[0081] ;
[0082] In the formula, xi and yi are the values of two geographical locations containing the target object at the i-th time point, respectively; and These are the numerical averages of two geographical locations containing the target object; n is the length of the time series.
[0083] Based on the Pearson correlation coefficient, the weights of the edges in the spatiotemporal graph are calculated using the following formula:
[0084] ;
[0085] The spatiotemporal graph is constructed by using the weights of the edges formed by connecting the two spatiotemporal graph nodes containing the geographic locations of the target objects.
[0086] It should be noted that the different geographical locations containing the target objects specifically include the geographical locations of the production target objects and the geographical locations of the transaction target objects. Calculating the historical numerical correlation between different geographical locations containing the target objects includes the correlation between production target objects, the correlation between transaction target objects, and the correlation between production target objects and transaction target objects.
[0087] The method provided in this embodiment uses reinforcement learning to adjust the weights of the spatiotemporal attention model, which can adapt to the impact of policy adjustments or emergencies in real time.
[0088] (3) Constructing a target object prediction model: The time feature calculation module is used to extract the time features in the time series data, the spatial feature calculation module is used to process the spatiotemporal graph and extract the spatial features, the spatiotemporal attention calculation module is used to combine time attention and spatial attention, the new features processed by the spatiotemporal attention calculation module are input into the fully connected layer, and the final target object prediction value is obtained through linear transformation.
[0089] When extracting time features from time series data using the time feature calculation module, a bidirectional long short-term memory network is specifically used to extract time features from the time series data. The specific implementation method is as follows:
[0090] Input the output of the target object, inventory data, and audiovisual materials containing the target object into the bidirectional long short-term memory network. The input data is specifically a three-dimensional tensor containing the number of samples, time step, and number of features.
[0091] The output features of the bidirectional long short-term memory network are specifically a three-dimensional tensor containing the number of samples, time step, and number of hidden layer neurons.
[0092] The spatial feature calculation module uses a graph convolutional network to process the spatiotemporal graph and extract spatial features. The input data of the graph convolutional network is the node features and adjacency matrix of the spatiotemporal graph, and the output data of the graph convolutional network is the spatial features extracted from the spatiotemporal graph.
[0093] The node features are specifically a two-dimensional tensor containing the number of nodes and node features. The node features include the annual output corresponding to the geographical location containing the target object, transportation network information, and audiovisual materials containing the target object for each spatiotemporal graph node.
[0094] The adjacency matrix is a two-dimensional tensor containing information about the number and weights of nodes. Specifically, the adjacency matrix contains the connection relationships between multiple spatiotemporal graph nodes and the weights of the spatiotemporal graph edges formed by every two spatiotemporal graph nodes.
[0095] The output data of a graph convolutional network is specifically a two-dimensional tensor that includes the number of nodes and the dimension of the output features.
[0096] When using the spatiotemporal attention calculation module to combine temporal attention and spatial attention, the output data of the temporal feature extraction module and the output data of the spatial feature extraction module are concatenated to form a new feature input to the spatiotemporal attention calculation module. The output is the feature weighted by the spatiotemporal attention mechanism. The specific steps include:
[0097] The importance of the same geographical location containing the target object at different time steps is calculated using the time attention branch. The specific calculation formula is as follows:
[0098] ;
[0099] In the formula, The time attention branch weights are represented by Qt, which is the time query vector, indicating the feature information of the output and inventory of the same geographical location containing the target object at the current time step t, with dimension d. q Wq is the temporal attention weight matrix with dimensions dq×dk; Kti is the key vector at the i-th time step, representing the vector obtained after linear transformation of the feature information at time step i, with dimensions dk; dk is the dimension of the key vector.
[0100] T is the transpose symbol, which is the transpose of the Wq matrix;
[0101] Softmax() is a general function that maps a set of real numbers to a probability distribution such that the sum of all elements is 1;
[0102] The spatial attention branch is used to learn the interactions between different geographical locations containing the target object. The specific calculation formula is as follows:
[0103] ;
[0104] In the formula, dv represents the spatial attention branch weights; Vi and Vj are the spatial value vectors of the i-th and j-th geographic locations containing the target object, respectively, representing the feature information of different geographic locations, with a dimension of dv; Wv is the spatial attention weight matrix, used to learn the interaction relationship between different geographic locations, transforming the spatial value vectors of different geographic locations to calculate their correlation, with a dimension of dv×dv; dv is the dimension of the spatial attention vector;
[0105] By combining temporal attention and spatial attention, the final spatiotemporal attention weights are obtained. It is used for weighted aggregation of temporal and spatial features, and the specific calculation formula is as follows:
[0106] .
[0107] The bidirectional long short-term memory network in the temporal feature calculation module, the graph convolutional network in the spatial feature calculation module, the weight matrix and the fully connected layer in the spatiotemporal attention calculation module are trained, and the training parameters are specifically the bias terms in each neural network layer.
[0108] In simple terms, the training process involves calculating the output result based on a given bias term and input features. The mean squared error between the predicted output and the actual result is then calculated. If the mean squared error is within the allowable range (0.000001), the training ends; otherwise, the parameter values are modified, and the above process is repeated until the requirement is met.
[0109] In step 400, the evaluation index related to the target object to be predicted at the current time is obtained, the evaluation index is input into the target object prediction model, the target object prediction result at the current time is obtained, and the 95% confidence interval of the time step is obtained. The confidence interval is used as the predicted value range of the target object, the current prediction result is input into the historical dataset, the historical dataset is updated, and the target object of the next time step is predicted, thereby obtaining the short-term trend of the target object.
[0110] The prediction results of this implementation are interval data, providing a confidence interval for the fluctuation of the target object, helping users to better manage risks.
[0111] The spatiotemporal attention mechanism provided in this implementation can not only capture temporal information such as target objects and output information over multiple time periods, but also spatial information such as different production geographical locations and different transaction geographical locations, realizing the linkage of time and spatial information. It integrates multiple data types such as historical values, output, inventory, geographical location, transportation network, and policy text, improving the comprehensiveness and accuracy of the model, while also improving modeling precision, making the predicted target objects more meaningful.
[0112] The method provided in this embodiment uses reinforcement learning to adjust the weights of the spatiotemporal attention model, which can adapt to the impact of policy adjustments or sudden events in real time. Moreover, the prediction results are interval data, providing a confidence interval for the fluctuation of the target object, helping users to better manage risks.
[0113] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.
Claims
1. A method for multi-gradient prediction of target objects by fusing multi-source data, characterized in that, Includes the following steps: Step 100: Extract evaluation indicators related to the prediction of the target object, form a change sequence of the target object, and construct a data set of evaluation indicators and change sequence of the target object on the time axis; Step 200: Establish a target object prediction model based on the dataset and spatiotemporal attention mechanism; Step 300: Capture the evaluation indicators at the current moment and input the evaluation indicators at the current moment into the target object prediction model to predict the target object. The prediction of the target object includes short-term expectations and medium- and long-term predictions. Step 400: Input the real-time evaluation indicators and target objects into the data set according to the time axis to update the data set, thereby correcting the target object prediction model. Iterate continuously according to the time axis to update the short-term expectations and medium-to-long-term predictions of the target objects.
2. The method for multi-gradient prediction of target objects by fusing multi-source data according to claim 1, characterized in that, In step 100, the method for constructing a dataset of evaluation indicators and target object change sequences on the time axis is as follows: Capture the time points when the value of the target object changes, and obtain the evaluation index corresponding to each time the value of the target object changes. The evaluation indicators include time series data, spatial data, and text data. The time series data includes the output volume and inventory data of the target object; the spatial data includes the geographical location of the target object's transactions and transportation network information; and the text data includes audiovisual materials containing the target object.
3. The method for multi-gradient prediction of target objects by fusing multi-source data according to claim 2, characterized in that, In step 200, the method for establishing the target object prediction model is as follows: Data preprocessing: Data processing is performed on the evaluation indicators and target objects in the dataset, including data cleaning, normalization, and text feature extraction for the evaluation indicators in the dataset; Constructing a spatiotemporal graph: Determining the weights of spatiotemporal graph nodes and edges, wherein spatiotemporal graph nodes include the geographical locations where the target object appears, connecting different spatiotemporal graph nodes in pairs to form spatiotemporal graph edges to construct the spatiotemporal graph, and determining the weight of each spatiotemporal graph edge; Construct a target object prediction model: Use the time feature calculation module to extract time features from the time series data; The spatial feature calculation module is used to process the spatiotemporal map and extract spatial features; By combining temporal and spatial attention using a spatiotemporal attention calculation module, the new features processed by the spatiotemporal attention calculation module are input into a fully connected layer, and the final target object prediction value is obtained through linear transformation.
4. The method for multi-gradient prediction of target objects by fusing multi-source data according to claim 3, characterized in that, The specific steps for data cleaning, normalization, and text feature extraction of the evaluation indicators in the dataset are as follows: Remove missing and outlier values from the evaluation metrics and target objects in the dataset; The time series data and spatial data in the evaluation index are normalized to the interval [0, 1] using the Min-Max method; For text data, text features are extracted by using a pre-trained language model to convert the text data into numerical vectors.
5. The method for multi-gradient prediction of target objects by fusing multi-source data according to claim 3, characterized in that, The weight of each spatiotemporal graph edge is dynamically calculated based on the historical numerical correlation between different geographical locations containing the target object. The weight of the spatiotemporal graph edge formed by two spatiotemporal graph nodes is determined using the Pearson correlation coefficient between two time series containing transactions of the target object. The specific calculation method is as follows: ； In the formula, xi and yi are the values of two geographical locations containing the target object at the i-th time point, respectively; and These are the numerical averages of two geographical locations containing the target object; n is the length of the time series. Based on the Pearson correlation coefficient, the weights of the edges in the spatiotemporal graph are calculated using the following formula: ； The spatiotemporal graph is constructed by using the weights of the edges formed by connecting the two spatiotemporal graph nodes containing the geographical locations of the target objects.
6. The method for multi-gradient prediction of target objects by fusing multi-source data according to claim 3, characterized in that, When extracting time features from the time series data using the time feature calculation module, a bidirectional long short-term memory network is specifically used to extract the time features from the time series data. The specific implementation method is as follows: The output quantity of the target object, inventory data, and audiovisual materials containing the target object are input into the bidirectional long short-term memory network. Specifically, the input data is a three-dimensional tensor containing the number of samples, time step, and number of features. The bidirectional long short-term memory network outputs extracted features, specifically a three-dimensional tensor containing the number of samples, time step, and number of hidden layer neurons.
7. The method for multi-gradient prediction of target objects by fusing multi-source data according to claim 5, characterized in that, The spatial feature calculation module uses a graph convolutional network to process the spatiotemporal graph and extract spatial features. The input data of the graph convolutional network are the node features and adjacency matrix of the spatiotemporal graph, and the output data of the graph convolutional network are the spatial features extracted from the spatiotemporal graph. The node features are specifically a two-dimensional tensor containing the number of nodes and node features. The node features include the annual output corresponding to the geographical location of the target object for each spatiotemporal graph node, transportation network information, and audiovisual materials containing the target object. The adjacency matrix is a two-dimensional tensor containing information on the number and weight of nodes. Specifically, the adjacency matrix contains the connection relationships between multiple spatiotemporal graph nodes and the weight of the spatiotemporal graph edge formed by every two spatiotemporal graph nodes. The output data of the graph convolutional network is specifically a two-dimensional tensor that includes the number of nodes and the output feature dimension.
8. The method for multi-gradient prediction of target objects by fusing multi-source data according to claim 5, characterized in that, When the spatiotemporal attention calculation module combines temporal attention and spatial attention, the output data of the temporal feature extraction module and the output data of the spatial feature extraction module are concatenated to form a new feature, which is then input into the spatiotemporal attention calculation module. The output is the feature weighted by the spatiotemporal attention mechanism. The specific steps include: The importance of the same geographical location containing the target object at different time steps is calculated using the time attention branch. The specific calculation formula is as follows: ； In the formula, The time attention branch weights are represented by Qt, which is the time query vector, indicating the feature information of the output and inventory of the same geographical location containing the target object at the current time step t, with dimension d. q Wq is the temporal attention weight matrix with dimensions dq×dk; Kti is the key vector at the i-th time step, representing the vector obtained after linear transformation of the feature information at time step i, with dimensions dk; dk is the dimension of the key vector. T is the transpose symbol, which is the transpose of the Wq matrix; Softmax() is a general function that maps a set of real numbers to a probability distribution such that the sum of all elements is 1; The spatial attention branch is used to learn the interactions between different geographical locations containing the target object. The specific calculation formula is as follows: ； In the formula, dv represents the spatial attention branch weights; Vi and Vj are the spatial value vectors of the i-th and j-th geographic locations containing the target object, respectively, representing the feature information of different geographic locations, with a dimension of dv; Wv is the spatial attention weight matrix, used to learn the interaction relationship between different geographic locations, transforming the spatial value vectors of different geographic locations to calculate their correlation, with a dimension of dv×dv; dv is the dimension of the spatial attention vector; By combining temporal attention and spatial attention, the final spatiotemporal attention weights are obtained. It is used for weighted aggregation of temporal and spatial features, and the specific calculation formula is as follows: 。 9. The method for multi-gradient prediction of target objects by fusing multi-source data according to claim 5, characterized in that, The bidirectional long short-term memory network in the temporal feature calculation module, the graph convolutional network in the spatial feature calculation module, the weight matrix of the spatiotemporal attention calculation module, and the fully connected layer are trained, wherein the training parameters are specifically the bias terms in each neural network layer.
10. The method for multi-gradient prediction of target objects by fusing multi-source data according to claim 1, characterized in that, In step 400, an evaluation index related to the target object to be predicted at the current time is obtained, the evaluation index is input into the target object prediction model, the target object prediction result at the current time is obtained, and a 95% confidence interval for that time step is obtained, and the confidence interval is used as the predicted value interval of the target object.