An artificial intelligence computing method and apparatus, an electronic device, and a medium

Abstract

The application discloses an artificial intelligence calculation method and device, electronic equipment and medium, and relates to the technical field of artificial intelligence, comprising the following steps: collecting multi-source data of an object to be modeled, and generating state, control and topological relationship input triplets; performing low-rank embedding on a state tensor historical sequence, and executing explicit time extrapolation to generate a basic prediction result; jointly extracting low-rank context variables based on state tensor and control tensor historical data, and broadcasting to all nodes; fusing the basic prediction result, the control tensor and the low-rank context variables, executing graph-constrained axial implicit spatial correction to generate a spatial correction increment; and fusing the basic prediction result and the spatial correction increment based on a gating mechanism to generate a final prediction result. The application can explicitly represent hidden environmental factors, take into account global trends and local constraints, and significantly improve prediction accuracy, system robustness and interpretability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence computing method, apparatus, electronic device, and medium. Background Technology

[0002] In recent years, artificial intelligence computing methods have made rapid progress in scientific computing, industrial forecasting and complex system modeling. With the widespread acquisition of multi-source data and the popularization of high-performance computing platforms, time series forecasting, graph neural networks and Transformer-type models based on deep learning have gradually become core tools for processing high-dimensional, unstructured and multimodal data.

[0003] However, existing technologies still have several limitations. First, traditional time series prediction methods rely heavily on one-way mapping of historical data, making it difficult to fully consider the coupling relationship between multiple sources of information, resulting in insufficient generalization ability in complex systems or unknown environments. Second, existing graph neural networks or Transformer models have high computational costs when processing high-dimensional continuous field data and lack explicit representation of hidden environmental variables, making it difficult to provide interpretability and long-term rolling stability. Third, existing fusion methods usually adopt simple weighting or end-to-end training strategies, lacking fine-tuning for spatial topological relationships and local neighborhood constraints, resulting in error accumulation in prediction results at local nodes or boundary regions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an artificial intelligence computing method that solves the problems of insufficient accuracy in state prediction and inadequate characterization of implicit environmental factors in existing technologies under conditions of multi-source data and complex spatial topology.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an artificial intelligence computing method, which includes collecting multi-source data of an object to be modeled and generating state, control and topological relationship input triples;

[0008] Low-rank embedding is performed on the history sequence of the state tensor, and explicit temporal extrapolation is performed to generate the basic prediction results;

[0009] Low-rank context variables are extracted from historical data of state tensor and control tensor and broadcast to all nodes.

[0010] By integrating the basic prediction results, control tensors, and low-rank context variables, axial implicit spatial correction with graph constraints is performed to generate spatial correction increments.

[0011] The final prediction result is generated by fusing the basic prediction results with the spatial correction increment based on the gating mechanism.

[0012] As a preferred embodiment of the artificial intelligence computing method described in this invention, the steps for collecting multi-source data of the object to be modeled and generating state, control, and topological relation input triples are as follows:

[0013] Acquire observation data of the target object at each time step. The observation data includes state field data characterizing the internal state distribution of the system, control field data characterizing the externally applied forces, and topological data characterizing the connection relationships between spatially discrete nodes.

[0014] The state field data and control field data at each time step are organized into state tensor representations and control tensor representations of a unified dimension according to a preset spatial grid layout. The tensor representations are then flattened along the spatial dimension into node feature matrices to obtain the node state matrix. and node control matrix ;

[0015] The node state matrix sequence, the node control matrix sequence, and the topology data are combined into an input triplet.

[0016] In a preferred embodiment of the artificial intelligence computing method described in this invention, the steps for generating the basic prediction results are as follows:

[0017] Based on the sequence of node state matrices in the input triplet, select a length of... Historical Window Window to history Every moment The node state matrix is ​​linearly embedded using the state embedding matrix to generate a hidden embedding matrix;

[0018] Perform low-rank decomposition on the state embedding matrix, and use the low-rank decomposed state embedding matrix to re-embed the sequences in the history window linearly, outputting the history embedding matrix sequence.

[0019] Using the embedding matrix at the current time as the query term, the query matrix, key matrix, and value matrix are calculated respectively using the learnable weight matrix;

[0020] Calculate the first [number] based on the correlation between the query item and each historical key. The attention weight of each historical moment for predicting the next moment;

[0021] By using attention weights to perform a weighted summation of the historical value matrices at each time step, the main trend hidden representation for the next time step can be obtained. ;

[0022] The hidden representation of the main trend is mapped back to the original state space using a learnable output mapping matrix, generating basic prediction results.

[0023] In a preferred embodiment of the artificial intelligence computing method described in this invention, the steps for extracting low-rank context variables are as follows:

[0024] Window to history The node state matrix sequence and node control matrix sequence are spatiotemporally pooled respectively, and the pooling results are concatenated to generate a joint observation vector;

[0025] Low-rank context variables are extracted based on joint observation vectors through low-rank mapping and nonlinear activation.

[0026] Broadcast the low-rank context variables to all spatial nodes to obtain the node-level context matrix.

[0027] In a preferred embodiment of the artificial intelligence computing method described in this invention, the steps for generating the spatial correction increment are as follows:

[0028] The basic prediction results, the node-level context matrix, and the node control matrix at the next time step are concatenated along the feature dimension to obtain the joint input matrix;

[0029] The joint input matrix is ​​restored to the form of a spatial grid tensor, and the horizontal axis features and vertical axis features are extracted along the two spatial axes of the spatial grid, respectively.

[0030] Based on the horizontal axis features and the vertical axis features, the horizontal and vertical query matrices and key matrices are generated respectively through learnable projection matrices, and the axial relationship kernel matrices of the horizontal and vertical axes are calculated.

[0031] Construct a graph constraint mask matrix based on the topology data. ;

[0032] Based on the axial relationship kernel matrix and the graph constraint mask matrix, the final implicit correction kernel matrix is ​​constructed.

[0033] The joint input matrix is ​​spatially corrected using an implicit correction kernel matrix to generate a spatial correction increment.

[0034] In a preferred embodiment of the artificial intelligence computing method described in this invention, the steps for generating the final prediction result are as follows:

[0035] A gated weight matrix is ​​generated based on the joint input matrix. The gated weight matrix is ​​then used to weight and fuse the basic prediction results with the spatial correction increment to obtain the final node-level prediction results.

[0036] The joint loss function is constructed using the node-level prediction results, the model parameters are trained and updated, the newly collected data is input into the trained model, and the final prediction result is output.

[0037] When multi-step recursive prediction is required, the final prediction result obtained at the current moment is used as one of the new historical inputs, and the prediction steps are repeated to form the future. The prediction results at each time point;

[0038] The final prediction results are restored to spatial tensor form for subsequent visualization analysis.

[0039] In a preferred embodiment of the artificial intelligence computing method described in this invention, the steps of constructing a joint loss function using node-level prediction results and training and updating the model parameters are as follows:

[0040] Based on the deviation between the final prediction result and the actual result, a single-step prediction error term is constructed.

[0041] Based on the multi-step recursive requirements in practical applications, a rolling stability term is defined.

[0042] Introduce context-compact regularization and graph smoothing terms;

[0043] The single-step prediction error term, rolling stability term, context compact regularization term, and graph smoothing term are jointly weighted to obtain the joint loss function, and the weighting weights are determined by cross-validation.

[0044] The model is trained using real historical data as the training set, constrained by a joint loss function, optimized using the Adam optimizer with gradient descent, and its parameters are updated. The iteration stops and the model is output when the loss no longer decreases significantly during continuous iteration.

[0045] Secondly, the present invention provides an artificial intelligence computing device, comprising,

[0046] The data acquisition module is used to acquire the state field tensor, control field tensor, and topological structure data of the object to be modeled, and flatten the state field tensor and control field tensor into node feature matrices to form the state, control, and topological relationship input triples required for subsequent calculations.

[0047] The temporal extrapolation module is used to perform low-rank embedding on the feature matrix of state nodes at each time point within the historical window, and to perform explicit temporal extrapolation on the embedding sequence based on the attention mechanism to generate basic prediction results that reflect the evolution trend of the system.

[0048] The context extraction module is used to perform spatiotemporal pooling and concatenation on the feature sequences of state nodes and control nodes within the historical window, extract global context variables based on low-rank mapping, and broadcast the context variables to all spatial nodes to form a node-level context matrix.

[0049] The spatial correction module is used to splice the joint input matrix, extract axial features along each spatial axis and calculate the axial relationship kernel matrix, construct the implicit correction kernel matrix by combining the spatial adjacency matrix, and perform spatial correction on the joint input matrix to generate spatial correction increments.

[0050] The prediction output module is used to generate a gating coefficient matrix based on the joint input. Based on the gating coefficient matrix, the basic prediction results and the spatial correction increment are weighted and fused element by element to obtain the final prediction node matrix.

[0051] The beneficial effects of this invention are as follows: by generating state, control, and topological relationship input triples, a complete and structured input is provided for subsequent historical embedding and spatial correction; by using low-rank embedding and explicit time extrapolation, the parameter scale is reduced, while the generated basic prediction results characterize the main trend evolution law of the system without applying complex spatial coupling corrections; by extracting and broadcasting low-rank context variables, the subsequent spatial correction module can obtain global working condition awareness information; by fusing basic predictions, control, and context execution graph-constrained axial implicit spatial corrections, the prediction results maintain continuity and consistency in local nodes and boundary regions; and by generating the final prediction results through a gating mechanism, a balanced control of the main trend and local corrections is achieved. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of the artificial intelligence calculation method in Example 1.

[0054] Figure 2 This is a structural diagram of the artificial intelligence computing device in Example 1.

[0055] Figure 3 This is a flowchart of the explicit time extrapolation used in Example 1 to generate the basic prediction results. Detailed Implementation

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0058] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0059] Example 1, referring to Figure 1 , Figure 2 and Figure 3 This is the first embodiment of the present invention, which provides an artificial intelligence computing method, including the following steps:

[0060] S1. Collect multi-source data of the object to be modeled and generate input triples for state, control and topological relationships;

[0061] Specifically, the observation data of the target object at each moment is acquired. The observation data includes state field data characterizing the internal state distribution of the system, control field data characterizing the externally applied action, and topological data characterizing the connection relationship between spatial discrete nodes.

[0062] The state field data includes at least one of the following: velocity field, pressure field, vorticity field, displacement field, stress field, strain field, temperature field, heat flux density field, and electric potential field.

[0063] The control field data includes at least one of the following: volume force field, surface force field, heat source distribution field, boundary velocity profile, boundary pressure gradient, material property spatial distribution field, and applied excitation power density distribution field.

[0064] The topological data includes the coordinate information and adjacency matrix of each spatial discrete node;

[0065] The state field data and control field data at each time step are organized into state tensor representations and control tensor representations of a unified dimension according to a preset spatial grid layout. The tensor representations are then flattened along the spatial dimension into node feature matrices to obtain the node state matrix. and node control matrix ;

[0066] The node state matrix sequence, the node control matrix sequence, and the topology data are combined into an input triplet.

[0067] S2. Perform low-rank embedding on the state tensor history sequence and execute explicit temporal extrapolation to generate basic prediction results;

[0068] Specifically, based on the sequence of node state matrices in the input triplet, a length of [missing information] is selected. Historical Window Window to history Every moment The node state matrix is ​​linearly embedded using the state embedding matrix to generate the hidden embedding matrix, expressed as:

[0069] ;

[0070] in, Indicates time The hidden embedding matrix, Indicates time The node state matrix, Represents the state embedding matrix, This represents a time-location encoding matrix used to distinguish the temporal positions of different historical moments;

[0071] Perform low-rank decomposition on the state embedding matrix, and use the low-rank decomposed state embedding matrix to re-embed the sequences in the history window linearly, outputting the history embedding matrix sequence.

[0072] Using the embedding matrix at the current time as the query term, the query matrix, key matrix, and value matrix are calculated respectively using the learnable weight matrix;

[0073] Calculate the first [number] based on the correlation between the query item and each historical key. The attention weights for predicting the next time step from each historical moment are expressed as follows:

[0074] ;

[0075] in, Indicates the first Attention weights for predicting the next moment from each historical moment. Represents the query matrix. This represents the transpose of the key matrix. Indicates a historical moment index. Indicates the scaling factor;

[0076] By using attention weights to perform a weighted summation of the historical value matrices at each time step, the main trend hidden representation for the next time step can be obtained. ;

[0077] The main trend hidden representation is mapped back to the original state space using a learnable output mapping matrix to generate the basic prediction result, expressed as follows:

[0078] ;

[0079] in, Indicates the basic prediction result, This indicates the output mapping matrix.

[0080] S3. Extract low-rank context variables based on historical data of state tensor and control tensor, and broadcast them to all nodes;

[0081] Specifically, regarding historical windows The node state matrix sequence and node control matrix sequence are spatiotemporally pooled respectively, and the pooling results are concatenated to generate a joint observation vector, expressed as follows:

[0082] ;

[0083] in, Represents the joint observation vector. This represents a vector concatenation operation. This indicates a spacetime pooling operation;

[0084] Based on the joint observation vector, low-rank context variables are extracted through low-rank mapping and nonlinear activation, expressed as follows:

[0085] ;

[0086] in, Represents low-rank context variables, Represents a non-linear activation function. and The low-rank mapping matrix is ​​represented by the state embedding matrix. Low-rank decomposition yields, Indicates the bias term;

[0087] Broadcast the low-rank context variable to all spatial nodes to obtain the node-level context matrix, expressed as:

[0088] ;

[0089] in, Represents the node-level context matrix. This represents a column vector consisting entirely of 1s.

[0090] S4. Integrate the basic prediction results, control tensor and low-rank context variables, perform axial implicit spatial correction with graph constraints, and generate spatial correction increments.

[0091] Specifically, the basic prediction results, the node-level context matrix, and the node control matrix of the next time step are concatenated along the feature dimension to obtain the joint input matrix;

[0092] The joint input matrix is ​​restored to the form of a spatial grid tensor, and the horizontal axis features and vertical axis features are extracted along the two spatial axes of the spatial grid, respectively.

[0093] The expressions for extracting the horizontal axis features and the vertical axis features are as follows:

[0094] ;

[0095] ;

[0096] in, Indicates the horizontal axis number The eigenvectors of the row, Indicates the longitudinal axis number Eigenvectors of rows and columns and Represents a nonlinear projection mapping. and Represents a pointwise linear mapping. and These represent the discrete dimensions of the two-dimensional spatial grid in the horizontal and vertical directions, respectively. This represents the joint input matrix restored to the form of a spatial grid tensor;

[0097] Based on the horizontal and vertical axis features, query matrices and key matrices are generated horizontally and vertically respectively using learnable projection matrices. The axial relationship kernel matrices for the horizontal and vertical axes are then calculated, as expressed in the following expressions:

[0098] ;

[0099] ;

[0100] in, The kernel matrix represents the axial relationships in the horizontal direction, used to describe the spatial correlation between rows in the horizontal direction. The kernel matrix represents the longitudinal axial relationships and is used to describe the spatial correlation between the columns in the longitudinal direction. and These represent the horizontal and vertical query matrices, respectively. and These represent the transposes of the horizontal and vertical key matrices, respectively. Indicates the scaling factor;

[0101] Construct a graph constraint mask matrix based on the topology data. When node With nodes When there is a physical connection, spatial adjacency, or predefined coupling relationship between them, let The value is 1, otherwise The value is 0;

[0102] Based on the axial relationship kernel matrix and the graph constraint mask matrix, the final implicit correction kernel matrix is ​​constructed, and its expression is:

[0103] ;

[0104] in, Represents the implicit correction kernel matrix. and These represent the horizontal and vertical identity matrices, respectively. , and This represents learnable weight parameters used to balance the contributions of horizontal associations, vertical associations, and local graph associations. Represents the Kronecker product;

[0105] The joint input matrix is ​​spatially corrected using an implicit correction kernel matrix to generate a spatial correction increment, expressed as follows:

[0106] ;

[0107] in, Indicates spatial correction increment. Represents a non-linear activation function. Represents the joint input matrix, This represents the learnable correction mapping matrix. This indicates the bias term.

[0108] S5. Based on the gating mechanism, the basic prediction results and spatial correction increments are fused to generate the final prediction results;

[0109] Specifically, a gated weight matrix is ​​generated based on the joint input matrix. This gated weight matrix is ​​then used to weight and fuse the basic prediction results with the spatial correction increment, yielding the final node-level prediction result, expressed as:

[0110] ;

[0111] ;

[0112] in, This represents the gate weight matrix, with a range of . , Represents a learnable gated mapping matrix. Indicates the bias term. This represents the final node-level prediction matrix. This represents element-wise multiplication;

[0113] The joint loss function is constructed using the node-level prediction results, the model parameters are trained and updated, the newly collected data is input into the trained model, and the final prediction result is output.

[0114] When multi-step recursive prediction is required, the final prediction result obtained at the current moment is used as one of the new historical inputs, and the prediction steps are repeated to form the future. The prediction results at each time point;

[0115] The final prediction results are restored to the form of a spatial tensor for subsequent visualization analysis;

[0116] Furthermore, based on the deviation between the final predicted result and the actual result, a single-step prediction error term is constructed, with the following expression:

[0117] ;

[0118] in, This represents the single-step prediction error term. Indicates the length of the training set sample sequence. Indicates a time index. Indicates time The predicted state results Indicates time The actual result, Represents square Norm;

[0119] Based on the multi-step recursive requirements in practical applications, a rolling stability term is defined, with the following expression:

[0120] ;

[0121] in, This represents the rolling stability term, used to constrain the cumulative error when the model continues to make rolling predictions using its own prediction results as input. This represents the overall mapping function of the method of the present invention;

[0122] Introducing context-compact regular expressions, the expression is:

[0123] ;

[0124] in, This indicates a context-tightening regularization term, used to prevent the values ​​of context variables from being excessively amplified. Represents the square of the Frobenius norm;

[0125] A graph smoothing term is introduced to maintain the physical continuity between adjacent nodes in the graph structure. The expression is:

[0126] ;

[0127] in, Indicates the graph smoothing term. This represents the set of edges determined by topological data. and These represent the first and second parts of the final prediction result. The node and the first The predicted vector of each node;

[0128] The single-step prediction error term, rolling stability term, context compact regularization term, and graph smoothing term are jointly weighted to obtain the joint loss function, and the weighting weights are determined by cross-validation.

[0129] The model is trained using real historical data as the training set, constrained by a joint loss function, optimized using the Adam optimizer with gradient descent, and its parameters are updated. The iteration stops and the model is output when the loss no longer decreases significantly during continuous iteration.

[0130] This embodiment also provides an artificial intelligence computing device, including:

[0131] The data acquisition module is used to acquire the state field tensor, control field tensor, and topological structure data of the object to be modeled, and flatten the state field tensor and control field tensor into node feature matrices to form the state, control, and topological relationship input triples required for subsequent calculations.

[0132] The temporal extrapolation module is used to perform low-rank embedding on the feature matrix of state nodes at each time point within the historical window, and to perform explicit temporal extrapolation on the embedding sequence based on the attention mechanism to generate basic prediction results that reflect the evolution trend of the system.

[0133] The context extraction module is used to perform spatiotemporal pooling and concatenation on the feature sequences of state nodes and control nodes within the historical window, extract global context variables based on low-rank mapping, and broadcast the context variables to all spatial nodes to form a node-level context matrix.

[0134] The spatial correction module is used to splice the joint input matrix, extract axial features along each spatial axis and calculate the axial relationship kernel matrix, construct the implicit correction kernel matrix by combining the spatial adjacency matrix, and perform spatial correction on the joint input matrix to generate spatial correction increments.

[0135] The prediction output module is used to generate a gating coefficient matrix based on the joint input. Based on the gating coefficient matrix, the basic prediction results and the spatial correction increment are weighted and fused element by element to obtain the final prediction node matrix.

[0136] This embodiment also provides a computer electronic device applicable to artificial intelligence computing methods, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the artificial intelligence computing method proposed in the above embodiment.

[0137] This computer electronic device can be a terminal, comprising a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0138] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the artificial intelligence computing method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0139] In summary, this invention provides a complete and structured input for subsequent historical embedding and spatial correction by generating state, control, and topological relationship input triples; it reduces the parameter scale through low-rank embedding and explicit time extrapolation, while the generated basic prediction results characterize the main trend evolution of the system without applying complex spatial coupling corrections; by extracting and broadcasting low-rank context variables, the subsequent spatial correction module can obtain global condition-aware information; by fusing basic predictions, control, and context execution graph-constrained axial implicit spatial corrections, the prediction results maintain continuity and consistency in local nodes and boundary regions; and by generating the final prediction results through a gating mechanism, a balanced control of the main trend and local corrections is achieved.

[0140] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An artificial intelligence computing method, characterized in that: include, Collect multi-source data of the object to be modeled and generate input triples for state, control, and topological relationships; Low-rank embedding is performed on the history sequence of the state tensor, and explicit temporal extrapolation is performed to generate the basic prediction results; Low-rank context variables are extracted from historical data of state tensor and control tensor and broadcast to all nodes. By integrating the basic prediction results, control tensors, and low-rank context variables, axial implicit spatial correction with graph constraints is performed to generate spatial correction increments. The final prediction result is generated by fusing the basic prediction results with the spatial correction increment based on the gating mechanism.

2. The artificial intelligence calculation method as described in claim 1, characterized in that: The steps for collecting multi-source data of the object to be modeled and generating state, control, and topological relation input triples are as follows. Acquire observation data of the target object at each time step. The observation data includes state field data characterizing the internal state distribution of the system, control field data characterizing the externally applied forces, and topological data characterizing the connection relationships between spatially discrete nodes. The state field data and control field data at each time step are organized into state tensor representations and control tensor representations of a unified dimension according to a preset spatial grid layout. The tensor representations are then flattened along the spatial dimension into node feature matrices to obtain the node state matrix. and node control matrix ; The node state matrix sequence, the node control matrix sequence, and the topology data are combined into an input triplet.

3. The artificial intelligence calculation method as described in claim 2, characterized in that: The steps for generating the basic prediction results are as follows: Based on the sequence of node state matrices in the input triplet, select a length of... Historical Window Window to history Every moment The node state matrix is ​​linearly embedded using the state embedding matrix to generate a hidden embedding matrix; Perform low-rank decomposition on the state embedding matrix, and use the low-rank decomposed state embedding matrix to re-embed the sequences in the history window linearly, outputting the history embedding matrix sequence. Using the embedding matrix at the current time as the query term, the query matrix, key matrix, and value matrix are calculated respectively using the learnable weight matrix; Calculate the first [number] based on the correlation between the query item and each historical key. The attention weight of each historical moment for predicting the next moment; By using attention weights to perform a weighted summation of the historical value matrices at each time step, the main trend hidden representation for the next time step can be obtained. ; The hidden representation of the main trend is mapped back to the original state space using a learnable output mapping matrix, generating basic prediction results.

4. The artificial intelligence calculation method as described in claim 3, characterized in that: The steps for extracting low-rank context variables are as follows: Window to history The node state matrix sequence and node control matrix sequence are spatiotemporally pooled respectively, and the pooling results are concatenated to generate a joint observation vector; Low-rank context variables are extracted based on joint observation vectors through low-rank mapping and nonlinear activation. Broadcast the low-rank context variables to all spatial nodes to obtain the node-level context matrix.

5. The artificial intelligence calculation method as described in claim 4, characterized in that: The steps for generating the spatial correction increment are as follows: The basic prediction results, the node-level context matrix, and the node control matrix at the next time step are concatenated along the feature dimension to obtain the joint input matrix; The joint input matrix is ​​restored to the form of a spatial grid tensor, and the horizontal axis features and vertical axis features are extracted along the two spatial axes of the spatial grid, respectively. Based on the horizontal axis features and the vertical axis features, the horizontal and vertical query matrices and key matrices are generated respectively through learnable projection matrices, and the axial relationship kernel matrices of the horizontal and vertical axes are calculated. Construct a graph constraint mask matrix based on the topology data. ; Based on the axial relationship kernel matrix and the graph constraint mask matrix, the final implicit correction kernel matrix is ​​constructed. The joint input matrix is ​​spatially corrected using an implicit correction kernel matrix to generate a spatial correction increment.

6. The artificial intelligence calculation method as described in claim 5, characterized in that: The steps for generating the final prediction result are as follows: A gated weight matrix is ​​generated based on the joint input matrix. The gated weight matrix is ​​then used to weight and fuse the basic prediction results with the spatial correction increment to obtain the final node-level prediction results. The joint loss function is constructed using the node-level prediction results, the model parameters are trained and updated, the newly collected data is input into the trained model, and the final prediction result is output. When multi-step recursive prediction is required, the final prediction result obtained at the current moment is used as one of the new historical inputs, and the prediction steps are repeated to form the future. The prediction results at each time point; The final prediction results are restored to spatial tensor form for subsequent visualization analysis.

7. The artificial intelligence calculation method as described in claim 6, characterized in that: The steps for constructing a joint loss function using node-level prediction results and training and updating the model parameters are as follows. Based on the deviation between the final prediction result and the actual result, a single-step prediction error term is constructed. Based on the multi-step recursive requirements in practical applications, a rolling stability term is defined. Introduce context-compact regularization and graph smoothing terms; The single-step prediction error term, rolling stability term, context compact regularization term, and graph smoothing term are jointly weighted to obtain the joint loss function, and the weighting weights are determined by cross-validation. The model is trained using real historical data as the training set, constrained by a joint loss function, optimized using the Adam optimizer with gradient descent, and its parameters are updated. The iteration stops and the model is output when the loss no longer decreases significantly during continuous iteration.

8. An artificial intelligence computing device, characterized in that: include, The data acquisition module is used to acquire the state field tensor, control field tensor, and topological structure data of the object to be modeled, and flatten the state field tensor and control field tensor into node feature matrices to form the state, control, and topological relationship input triples required for subsequent calculations. The temporal extrapolation module is used to perform low-rank embedding on the feature matrix of state nodes at each time point within the historical window, and to perform explicit temporal extrapolation on the embedding sequence based on the attention mechanism to generate basic prediction results that reflect the evolution trend of the system. The context extraction module is used to perform spatiotemporal pooling and concatenation on the feature sequences of state nodes and control nodes within the historical window, extract global context variables based on low-rank mapping, and broadcast the context variables to all spatial nodes to form a node-level context matrix. The spatial correction module is used to splice the joint input matrix, extract axial features along each spatial axis and calculate the axial relationship kernel matrix, construct the implicit correction kernel matrix by combining the spatial adjacency matrix, and perform spatial correction on the joint input matrix to generate spatial correction increments. The prediction output module is used to generate a gating coefficient matrix based on the joint input. Based on the gating coefficient matrix, the basic prediction results and the spatial correction increment are weighted and fused element by element to obtain the final prediction node matrix.

9. An electronic device comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A storage medium having a program stored thereon, characterized in that: When the program is executed, it implements the method as described in any one of claims 1 to 7.

Images